Methods and systems for the rapid detection of concealed objects

FIELD OF THE INVENTION

The present invention relates generally to X-ray based methods and systems for detection of concealed threats, and threat resolution, and more specifically to improved methods and systems, using dual stage scanning to process luggage for faster inspection with reduced false alarm rate.

BACKGROUND OF THE INVENTION

Conventional X-ray systems produce radiographic projection images, which are then interpreted by an operator. These radiographs are often difficult to interpret because objects are superimposed. A trained operator must study and interpret each image to render an opinion on whether or not a target of interest, a threat, is present. With a large number of such radiographs to be interpreted, and with the implied requirement to keep the number of false alarms low, operator fatigue and distraction can compromise detection performance.

Advanced technologies, such as dual-energy projection imaging and Computed Tomography (CT), are being used for contraband detection, beyond conventional X-ray systems. In dual-energy imaging it is attempted to measure the effective atomic numbers of materials in containers such as luggage. However, the dual-energy method does not readily allow for the calculation of the actual atomic number of the concealed 'threat' itself, but rather yields only an average atomic number that represents the mix of the various items falling within the X-ray beam path, as the contents of an actual luggage is composed of different items and rarely conveniently separated. Thus dual-energy analysis is often confounded. Even if the atomic number of an item could be measured, the precision of this measurement would be compromised by X-ray photon noise to the extent that many innocuous items would show the "same" atomic number as many threat substances, and therefore the atomic number in principle cannot serve as a sufficiently specific classifier for threat versus no threat.

In X-ray CT cross-sectional images of slices of an object are reconstructed by processing multiple attenuation measurements taken at various angles around an object. CT images do not suffer much from the super-positioning problem present in standard radiographs. However, conventional CT systems take considerable time to perform multiple scans, to capture data, and to reconstruct the images. The throughput of CT systems is generally low. Coupled with the size and expense of CT systems this limitation has hindered CT use in applications such as baggage inspection where baggage throughput is an important concern. In addition, CT alarms on critical mass and density of a threat, but such properties are not unique to explosives. CT based systems suffer from high false alarm rate. Any such alarm is then to be cleared or confirmed by an operator, again interpreting images, or hand searching.

Apart from X-ray imaging systems, detection systems based on X-ray diffraction, or coherent scatter are also known. Their primary purpose is not to acquire images but to obtain information about the molecular structure of the substances an object is composed of. The so-called diffraction or coherent scatter signature is based on BRAGG reflection, that is the interference pattern of X-ray light, which develops when X-rays are reflected by the molecular structure or electron density distribution of a substance.

Various inspection region geometries have been developed and disclosed. Kratky, in Austrian Patent No. 2003753 publishes a refined arrangement of circular concentric apertures combined with an X-ray source and a point detector, to gain the small angle diffraction signature of an object placed between the apertures. More recently Harding in U.S. Patent No. 5,265,144, uses a similar geometry but replaces the point shaped detector aperture with an annular detector configurations. Both patents are incorporated herein by reference.

The resulting diffraction spectra can be analyzed to determine the molecular structure of the diffracting object, or at least to recognize similarity with any one of a number of spectra, which have previously been obtained from dangerous substances.

One approach to detecting explosives in luggage was disclosed in British patent No. 2,299,251 in which a device uses Bragg reflection from crystal structures to identify crystalline and poly-crystalline substances. Substances can be identified because the energy spectrum distribution of the polychromatic radiation reflected at selected angles is characteristic of the crystal structure of the substance reflecting the radiation.

United States Patent Nos. 4,754,469, 4,956,856, 5,008,911, 5,265,144, 5,600,700 and 6,054,712 describe methods and devices for examining substances, from biological tissues to explosives in luggage, by recording the spectra of coherent radiation scattered at various angles relative to an incident beam direction. U.S. Pat. No. 5,265,144 describes a device using concentric detecting rings for recording the radiation scattered at particular angles. Each of the prior art systems and methods, however, suffer from low processing rates because the scatter interaction cross sections are relatively small and the exposure times required to obtain useful diffraction spectra are long, in the range of seconds and minutes. For security inspections, equipment performance has to combine high detection sensitivity and high threat specificity with high throughput, at the order of hundreds of bags per hour.

U.S. Patent No. 5,182,764 discloses an apparatus for detecting concealed objects, such as explosives, drugs, or other contraband, using CT scanning. To reduce the amount of CT scanning required, a pre-scanning approach is disclosed. Based upon the pre-scan data, selected locations for CT scanning are identified and CT scanning is undertaken at the selected locations. The inventors claim the pre-scan step reduces the scanning time required for each scanned item, therefore increasing throughput. However, the use of CT scanning is still inefficient, not threat specific, and does not allow for rapid scanning of objects.

U.S. Patent No. 5,642,393 discloses a multi-view X-ray inspection probe that employs X-ray radiation transmitted through or scattered from an examined item to identify a suspicious region inside the item. An interface is used to receive X-ray data providing spatial information about the suspicious region and to provide this information to a selected material sensitive probe. The material sensitive probe, such as a coherent scatter probe, then acquires material specific information about the previously identified suspicious region and provides it to a computer. The disclosed system does not, however, address critical problems that arise in the course of applying a scatter probe to a selected suspicious region, including the accurate identification of a suspicious region, correction of detected data, and the nature of processing algorithms used.

Accordingly, there is need for an improved automatic threat detection and resolution system that captures data through an X-ray system and utilizes this data to identify threat items in a rapid, yet accurate, manner. Preferably, the system is highly threat specific in order to reliably and automatically discern threats from innocuous materials and items while still being able to process in excess of 100 bags per hour. Preferably the system utilizes relatively inexpensive industrial components, and does not need special support facilities. Additionally, the system should provide for greater accuracy in utilizing pre-scan data to identify an inspection region and in processing scan data.

<u>SUMMARY OF THE INVENTION</u>

One object of the present invention is to provide for an improved scanning process having a first stage to pre-select the locations of potential threats and a second stage to accurately identify the nature of the threat. The improved scanning process increases throughput by limiting the detailed inspection to a small fraction of the total bag volume, and it decreases the frequency of false alarms by applying threat specific analysis.

Another object of the invention is to provide for improved processing techniques performed in association with various scanning systems. The improved processing techniques enable the substantially automated detection of threats and decrease the dependence on operator skill and performance.

Another object of the invention is to provide for a method and system to screen for relatively small amounts of threat material.

Another object of the invention is to provide for an improved method and system for screening for explosives in the form of thin sheets.

Another object of the invention is to provide a screening solution at low cost by utilizing standard industrial components, including relatively low cost and rugged industrial X-ray systems and detector systems.

Accordingly, one embodiment of the present invention provides an apparatus for identifying an object concealed within a container. These objects may be considered threats, such as an illegal drug, an explosive material, or a weapon. The apparatus comprises a first stage inspection system having at least two X-ray projection systems projecting in different directions to generate a first set of data, and a data processing system in data communication with the first stage inspection system. The processors process the first set of data to generate at least two images or image data files. The images are then subjected to a set of image interpretation algorithms, also processed by the processors, to identify target regions in the two images. Since the images are projected in different direction it is possible to back-project identified target regions and to locate those targets in system coordinates.

The first stage inspection system locates potential threat items, regions, and/or areas, based on X-ray images, manual or automatic detection algorithms, and triangulation. A second stage inspection system, discussed below, then focuses on the identified items, regions, and/or areas to produce characteristic signatures which are then used to determine whether a threat is, in fact, present.

In one embodiment, a target region is identified from the two images generated in the first stage inspection system by having an operator select a particular region displayed in the images. In a preferred embodiment, the operator directs a cursor, using an interface, such as a mouse, to position crosshairs on each of the two images. The two crosshairs determine a certain location in system coordinates. The selection for the cross hair location may occur based upon an X-ray image characteristic, such as the X-ray shadow of an object, seen in both images. Optionally, the target selection process may be performed electronically. Target regions are identified from the two images by having a processor execute an algorithm to select regions in the images, which correspond to objects or mass accumulations. With the locations of each of the two images determined, the coordinates corresponding to the physical locations of the target region can be determined and used to direct the system, and, in particular, the conveyor.

Once the target region is determined, either through automatic or operator processing means, a plurality of control commands is produced and used to position, by a multiple-axis motion control system, the second stage inspection system such that an inspection region, at least partially, coincides with the determined target coordinates. In one embodiment, the inspection region is positioned relative to the target region using a plurality of adjustable apertures that can be physically moved. The apertures can be ring-shaped with an adjustable diameter. Optionally, the means for positioning the inspection volume relative to the target region comprises a motion-controlled conveyor operable to move in elevation as well as back and forth relative to the second stage inspection system. Optionally, the inspection region can be moved across the conveyor by mounting the second stage inspection system on a C-arm which is motion controlled to move back and forth across the conveyor, or alternatively by employing a parallel set of fixed linear bearings and synchronized linear motion to effect the same relative movement.

The second stage inspection system generates an inspection volume in space and produces a second set of data having an X-ray signature characteristic of the material in that inspection volume. The X-ray signature characteristic is a diffraction pattern, also called scatter spectrum, and an intensity level associated with that spectrum, and, in addition, a set of dual energy transmission measurements in close proximity to the ray path of the diffraction measurement.

The second stage inspection system comprises a source of X-ray radiation. In one embodiment, it comprises an energy dispersive detector. In another embodiment, it comprises an array of transmission detectors. In a preferred embodiment, it comprises both an energy dispersive detector and an array of transmission detectors. The energy dispersive detector is used to produce a signature of the material in the inspection region and the array of transmission detectors is used to produce data defining at least one of mass, degree of attenuation, area, and average atomic number, of the material in a beampath. Optionally, the array of transmission detectors is in a ring formation. In a preferred embodiment, the array of transmission detectors comprises high energy and low energy detectors. Data generated from the transmission detectors is used to determine a reference spectrum by identifying a spectrum associated with data generated from both the high energy detectors and the low energy detectors. The reference spectrum can be used to correct a diffraction spectrum or to correct for beam hardening.

It is preferred that the present invention includes a method of automatically determining the existence of a threat based upon the second set of data. While a variety of algorithms can be employed for this task, a preferred approach is to use a neural network to process at least one of the first set of data or the second set of data to determine the existence of a threat. In one embodiment, the neural network operates as a back-propagation network having a plurality of nodes and wherein the nodes are organized in a series of successive layers, each layer comprising at least one node that receives inputs from nodes in a prior layer and transmits outputs to nodes in a subsequent layer. Optionally, the nodes in a first layer are weighted in accordance with their distance from at least one node in a second layer. Optionally, the neural network is trained to determine the existence of the threat using a plurality of libraries, such as network accessible libraries, threat libraries, non-threat libraries, and/or other libraries.

These and other embodiments of the present invention will be disclosed in greater detail with reference to the drawings and detailed description.

BRIEF DESCRIPTION OF THE DRAWINGS

These and other features and advantages of the present invention will be appreciated, as they become better understood by reference to the following Detailed Description when considered in connection with the accompanying drawings, wherein: <ul id="ul0001" list-style="none" compact="compact"><li>Figure 1 is a schematic view of one embodiment of the dual stage X-ray scanning system;</li><li>Figure 2 is a schematic view of one embodiment of an X-ray scanning system for the first stage scanning system;</li><li>Figure 3 is a schematic view of one embodiment of the first stage of the X-ray scanning system for identifying a target region;</li><li>Figure 3a depicts exemplary images for identifying the location of an item within a container;</li><li>Figure 4 is a schematic diagram of a cross-section of one embodiment of a preferred beam delivery system for use in a second stage scanning system;</li><li>Figure 5 is a schematic diagram of one embodiment of the beam delivery system of the second stage scanning system;</li><li>Figure 6 is a schematic diagram of an exemplary look up source for transmission spectra;</li><li>Figure 7 is a schematic representation of a beam delivery system having multiple energy dispersive detectors;</li><li>Figure 8 is a graphical representation of an artificial neural network;</li><li>Figure 9 is a flow diagram describing a plurality of steps for practicing one embodiment of the present invention; and</li><li>Figure 10 is a flowchart depicting a process of training the neural network.</li></ul>

DETAILED DESCRIPTION OF THE INVENTION

The methods and systems described herein are directed towards finding, locating, and confirming threat items and substances. Such threats may comprise explosives such as C4, RDX, Semtex, Seismoplast, PE4, TNT, dynamite, PETN, ANFO among others, as well as other contraband such as drugs. Although the embodiments have been described in the context of a baggage inspection system, it should be evident to persons of ordinary skill in the art that items other than luggage such as other packages, mail, and cargo-containers, or even processed food stuffs, can also be analyzed and screened or graded and that the descriptions are exemplary and are not restrictive of the invention. Further, while the invention is described as a dual-stage system and method, the processing techniques discussed herein can be applied to each of the individual scanning stages.

Referring to Figure 1, a dual stage scanning system 100 comprises a housing 130, which encompasses a conveyor system 115 for moving containers, baggage, luggage, or similar object 110 through a plurality of scanning stages 150, 155. A sensor system 165 is connected at the entrance to determine when an object being scanned 110 enters the scan field and communicates with a controller [not shown] to activate or deactivate an X-ray radiation source, 170, 172, as needed. A lead lined tunnel 180 surrounds the conveyor to reduce radiation leakage outside the equipment. At least one radiation source is not expressly depicted in Figure 1 and would be visible if the system were viewed from the opposite side.

Referring to Figure 2, the first stage 150, comprises two X-ray cameras held together by a support structure 220, such as a frame or yoke, for stability. Each camera consists of an X-ray source 170, 171, a X-ray focusing means, such as a collimating slit comprised of a radio-opaque material, for example lead (not shown), and an array of detectors, 200, 201. In one embodiment, it is preferred that the detectors are configured into a L-shape in order to save space. One of ordinary skill in the art would appreciate that other folded configurations may be acceptable, provided that the detectors are appropriately positioned relative to the inspection region and X-ray source.

Behind each slit collimator, a thin sheet of X-rays 210 is formed. Within the sheet, a fan of pencil beams can be defined, shown as dashed lines in Figure 2, by connecting lines between the stationary focus, not shown, and channels in the detector array. Between focus and detector is a tunnel 180 through which the luggage is transported or moved using any means known in the art, including, for example, a conveyor 115, the surface of which is depicted in Figure 2. Wherever in the system radiation has to be transmitted from X-ray sources 170, 171 and through the region defined by tunnel 180, the conveyor belt support structure as well as the tunnel has windows constructed from materials essentially translucent to X-rays. The collimating slits and detector arrays are oriented so that the radiation-fans intersect the main conveyor surface within a few degrees of perpendicular relative to the conveyor surface. The two X-ray sources and their fans point in different directions.

In one preferred embodiment, the detector arrays are mounted on printed circuit boards with a vector positioned normal to their surfaces directed to the X-ray focus. An exemplary printed circuit board has a capacity of 64 channels, and the boards are physically arranged in Venetian blind configuration. The detector arrays consist of linear arrays of silicon photodiodes that are covered with scintillation material, which produces light when exposed to X-rays. The light is detected by the photodiodes that produce corresponding photo current signals. The detectors measure to what degree the X-ray signal has attenuated due to passing through a defined inspection volume. Specifically, the detected data are converted to digital format, corrected for detector gain and offset, and then stored. The required processor means may comprise computing hardware, firmware and/or software known to persons of ordinary skill in the art. When a container under inspection is moving through the tunnel and passing through the X-ray projections, both detector arrays are being sampled repetitively between 50 and 500 times per second. Displaying the line projections on a monitor renders the projection X-ray image.

While a conventional line scan system could be used as the first stage scanning system, it is preferred to use the system as described herein. More specifically, the present invention provides for the placement of at least two X-ray sources such that the directions of the X-ray projections emanating from the sources are mirrored relative to the central vertical plane. Therefore, from the perspective of a view along the path of conveyance through the first stage scanning system, at least one X-ray generator is mounted at a five o'clock position and at least one X-ray generator is mounted at the 7 o'clock position.

One of ordinary skill in the art would appreciate that the first stage scanning system is not limited to the specific embodiments described above and that other variations are included within the scope of this invention. In one alternative embodiment, detector arrays are expanded from a single array to multiple parallel arrays of detectors. In a second alternative embodiment, X-ray projections are taken using two-dimensional pixellated detector planes, without requiring the use of a conveyance means. It should be appreciated that, while the present invention will be further described using a description of the invention based on using the line scan configuration of single stationary foci and single line detector arrays in conjunction with a means of conveyance, the present invention includes other systems and methods that generate X-ray projection images and that such systems and methods can be used in the novel dual stage scanning system disclosed herein.

An alternative embodiment uses dual energy imaging. Dual energy imaging can be utilized to display an image where materials of a metallic constituency are suppressed (not displayed) or materials of an organic constituency are suppressed. Having the ability to selectively display certain materials within images helps reduce image clutter. For example, when inspecting containers for masses or explosives, which have little or no metallic component, the "organic materials only" display is preferred. The dual energy approach can be further refined to automatically discriminate between similar materials of higher and lower relative atomic numbers, such as between a plastic comprised of more lower atomic number atoms like hydrogen and carbon and a plastic comprised of more higher atomic number elements like oxygen and nitrogen; or between aluminum (atomic number 13) and steel (atomic number 26).

In one embodiment, dual energy data is generated by using an X-ray tube with extended spectral emission, which is standard, in conjunction with arrays of stacked detectors, where the first detector is positioned to detect more of the lower energy, or so-called softer X-ray photons, and the second detector is positioned to detect the balance of the energy, namely the higher energy, or so-called harder, photons. The second detector is typically positioned behind the first detector. The low energy and high energy measurements are combined in a suitable way using a series of calibration measurements derived from dual energy measurements taken of identified organic and metallic materials of known thicknesses and result in the display of images, including organic only or metal only images. One of ordinary skill in the art would appreciate that various dual energy line scan systems are commercially available.

It is preferred to use projection imaging as the first stage scanning step in this invention. Features shown in the projection images can be used by an operator to make a final decision on whether items identified in a container represent a threat of some type. Additionally, by taking projections from at least two different angles, it is possible to triangulate the location of a potential threat relative to the physical coordinates of the system and use those coordinates to perform a more specific and focused second stage scan. The triangulation process localizes certain items that generate features of interest in the images and identifies their location in the form of system coordinates.

To perform the triangulation process, the images that form the basis of the triangulation process and that are used to identify a target region are first identified. In one embodiment, the images are analyzed by an operator who visually and approximately determines a plurality of X-ray image characteristics, such as degree of attenuation and projected area, associated with mass, atomic number (identified using image color coding), and shape. Operators also use contextual information, such as an X-ray opaque organic mass in a transistor radio or a suspiciously thick suitcase wall. The analytical process is known to those of ordinary skill in the art and includes the interpretation of X-ray image characteristics.

In another embodiment, images are identified by determining the target regions automatically. For example, where the screening target is a mass of plastic explosive, known algorithms, working on dual energy X-ray projection image data, can be combined to automatically find such target. Examples for such algorithm components include, but are not limited to, edge detection, watershed, and connected component labeling.

Referring to Figure 3, a container 110 is moved on a conveyor 115 through a tunnel 180 in x-direction, perpendicular to the plane of the Figure. A first X-ray generator 170, C1, with an X-ray emitting focus projects a fan of X-rays 300 through a slit collimator onto an array of detectors mounted on printed circuit boards 200. One of ordinary skill in the art would appreciate that only a small sampling of detectors are shown in Figure 3 and that a typical system would have a far greater number of detectors, preferably 700 to 800, more preferably 740. As shown, the orientation of the fan plane is perpendicular to the conveyor surface. While a container is being moved along the conveyor surface, the detectors are read out repeatedly, and their signals are converted into digital format by detector electronics that are also mounted on the detector boards 200. The data are being processed and sorted further and stored in a computer [not shown] for display on a monitor [not shown]. Each horizontal line on the monitor corresponds to one particular detector in the array. Therefore, in a system using 740 detectors, the full image is composed of 740 lines.

A second X-ray camera, C2, consisting of X-ray generator 171, slit collimator (not shown) and detector array 201 is mounted in a different orientation, and offset in conveyor direction, by typically 100mm. The detectors aligned with this camera are sampled essentially simultaneously with the detectors of the first camera and produce a second image displayed on a monitor.

Operationally, an item 340 located within the container 110 is recognized in the course of the first stage scan using a detection algorithm or by operator analysis, depending upon the system mode chosen. With the item 340 identified, the approximate centerline X-ray projections 330, 331 that pass through the object can be determined. Each of the centerlines 330, 331 is associated with a certain detector channel, 310 and 311 respectively in each view.

Referring to Figure 3a, once the detector channels have been determined, the location of the associated item 340 can be found in the y-z coordinate system. Two images 380, 381 corresponding to the two views are shown. With knowledge of the detectors associated with the centerlines 331, 330 and the range of detectors, 308 to 314, defined, the y and z coordinates of the item 340 can be derived. The x-coordinate is defined by the direction of conveyor motion and is known because the conveyor motion control system, timing of X-ray exposure, and the fixed offset of the two scan planes are known. The x-coordinate can, for example, be referenced to the beginning, or leading edge of the container, which can be detected by a light curtain or similar position-detecting device. In particular, the two images are referenced to each other precisely in the x-coordinate direction.

The purpose of this triangulation or localization of identified items in a container is to generate control commands that can be used to position and focus the inspection region or inspection volume of the second stage scanning system on the identified item. Therefore, the first inspection stage quickly locates potential threats and determines their coordinates, as referenced to the system, while the second stage focuses on better determining the nature of the identified potential threat. It should be appreciated that, because the first stage characterization of a threat is loosely based on features in X-ray images, it will locate, find, and label, as a potential threat, items which are innocuous, in addition to real threats. Therefore, the performance of a detection system based only on the first stage, as described, suffers from a high false alarm rate.

One of ordinary skill in the art would also appreciate that other elements of the first stage scanning system are not depicted in Figure 1 but would be included in an implementation of the system. For example, a shielding curtain is positioned at both the entrance and exit of the system 100 to protect against radiation leakage to the surrounding environment. The system 100 is controlled by a data interface system and computer system that is capable of rapid, high data rate processing, is in data communication with storage media for the storage of scan data and retrieval of reference libraries, and outputs to a monitor having a graphics card capable of presenting images.

It should also be appreciated that a second stage scan may not be required. In one embodiment, radiographic images from the first stage scan are displayed on a computer monitor for visual inspection with target regions or potential threats identified. An operator may dismiss some of the identified regions or threats based on context, observation, or other analytical tools. If no threats are identified, the container is cleared to exit the inspection system without subjecting it to the second stage of scanning. However, if the operator is unable to resolve an area as being a non-threat, the area is identified as a target region.

The second stage inspection or scanning system closely inspects the identified target locations by deriving more specific information, or a signature, and confirming the first stage threat alarm only if the obtained signature matches the signature of a threat substance or threat item. An alarm confirmed by the second stage system are then taken seriously by operators and indicate the need for further inspection, including, but not limited to, operator image interpretation, additional scanning, and/or hand searching the container.

In a preferred embodiment, the second stage scanning system uses diffracted or scattered radiation to determine the properties of a material, obtain a signature, and, accordingly, identify a threat. Diffracted or scattered radiation comprises photons that have experienced an interaction with the object under investigation. In the special case of small angle scattering, the majority of interactions are elastic or energy-conserving; specifically, the diffracted photon has the same energy as it had before the interaction, just its direction of propagation has changed. If the energy distribution of the scattered photons is being analyzed by an energy-dispersive detector system, which is commercially available, certain properties of the material causing the scatter are being encoded in the signature. Photons scattered under small angles are scattered selectively due to interference effects. Since the process does not change the energy of the photons the signal also contains the distribution of the primary radiation in a simply multiplicative way. The incoming primary radiation, as well as the scattered radiation, encounter further spectral modifications due to other types of interactions, such as Compton scatter and photoelectric absorption, which are not energy preserving. If one wants to view the characteristics of the scattering material, other distracting spectral effects have to be removed.

The detected signature of a threat is therefore a combination of X-ray properties. One important property is a BRAGG diffraction spectrum, observed at small diffraction angles between 2 and 8 degrees, with a preferred value around 3 degrees.

Figure 4 shows schematically a cross section of a preferred beam delivery system used to obtain BRAGG spectra at small angles. Other beam delivery systems can also be used in the present invention, including those disclosed by Kratky, et al. in Austrian Patent No. 2003753 and Harding in U.S. Patent No. 5,265,144. The preferred system depicted in Figure 4 further includes a transmission detector.

A beam delivery system separates the photon radiation emitted by the focus 400 of the X-ray source 404 into a plurality of beams. A beam 401 is formed by passing through apertures 410 and is directly detected by detectors 402, which are within the beam's direct line-of sight. These beams are referred to as transmission beams. Scatter interactions are detected by blocking direct line-of-sight detection through the use of ring apertures 410, 411 and exposing the associated detector 420 only to scattered radiation 492. Therefore, scatter radiation, generated when certain beams interact with an inspection region or volume 445, can be detected in the same apparatus as transmission radiation.

The choice of ring aperture diameters, distance to focus, and distance to detector determines the effective scatter angle 430 of the photons falling on the detector. In one embodiment, the scatter angle 430 is approximately the same for substantially all photons detected by the detector of the scattered radiation. It is preferred to configure the beam delivery system to establish an effective scatter angle of between two and 8 degrees. It is more preferable to have a scatter angle at or about 3 degrees. Using a beam delivery system having a circular symmetry has the advantage of obtaining a scatter contribution from a larger volume of the material being inspected, thereby increasing the inherently weak scatter signal. Additionally, the scatter spectrum can be cost efficiently detected using only a single detector channel 420 with an entrance aperture in the shape of a hole 421.

The scatter signal is generated by positioning the target region 445, identified in the first stage scan, between the beam forming apertures, irradiating that region 445 using the conical beam 442, and making sure scatter radiation from the target region 445can be detected by the scatter detector. The target region 445, often contained within a container 450, is in the shape of a tube or ring 445 and is referred to as the inspection volume or inspection region. The length, diameter, and wall thickness of the inspection volume depends on the particular shape of the elements of the beam delivery system, including focus size, ring aperture diameter and width, detector opening and overall distance. In a preferred embodiment for the inspection of large luggage, the inspection volume is at or about 60 cubic centimeters.

In a preferred embodiment, as shown in Figure 5, the components of the beam delivery system are mounted to the open ends of a rigid support structure 500 formed in the shape of a C (referred to herein as a C-arm) and aligned with a tolerance of at or about 0.1 millimeters. A first arm of the C-arm comprises a X-ray tube with X-ray focus 172, a beam limiting aperture hole mounted to the tube head 401, and a ring-shaped aperture 410. A second arm holds comprises a transmission detector array 402, a second ring aperture 411, and an energy dispersive detector 420, equipped with an aperture hole.

The energy dispersive detector 420 is positioned to receive scattered radiation from a target object placed on the conveyor running between the arms of the C-arm support structure where a first arm is above the conveyor and a second arm is below the conveyor. The transmission detector is positioned to receive radiation attenuated by the same target object. It is preferable for the C-arm to be mobile and capable of moving in the x-direction along the length of the conveyor. Therefore, the C-arm with tube and detectors can be re-positioned along the length of the conveyor.

In a preferred embodiment, the scatter detector 420 is comprised of cadmium telluride or cadmium zinc telluride and is operated at room temperature, or approximate to room temperature, An exemplary embodiment is available from the e-V Products Company, Saxonburg, Pennsylvania. This type of detector has a spectral resolution performance that is well matched to the limited angular requirements of this application, and therefore the limited spectral resolution of the beam delivery system.

In one mode of operation, the potential threat locations inside a container are found automatically by the first stage, and, based upon the physical coordinates obtained through triangulation, the second stage scanning system is automatically positioned to generate an inspection region that substantially overlaps with the identified target region. Where multiple threat locations are identified, the second stage scanning system is sequentially repositioned to focus on each subsequent target region. To scan each target region, the second stage X-ray source is activated and the scatter detector and transmission detector are sampled simultaneously. In a preferred embodiment, a transmission spectrum associated with the detected transmission data is characterized using a look up reference, figure, table, or chart, and the scatter spectrum is normalized using that identified transmission spectrum.

In another mode of operation, an operator actively identifies images that he or she believes corresponds to a potential threat. X-ray images from the first inspection stage are displayed to the operator, and the operator points to a suspicious object as it appears in both views. To support this functionality, operators use a computer system, comprising a mouse and monitor, to position cross hairs over the areas of interest on each of the images. Using coordinate data generated through triangulation, the second stage scanning system automatically positions itself such that an inspection region overlaps with the target region, activates the X-ray source and simultaneously samples the scatter detector and transmission detector. In a preferred embodiment, a transmission spectrum associated with the detected transmission data is characterized using a look up reference, figure, table, or chart, and the scatter spectrum is normalized using that identified transmission spectrum.

As discussed above, a transmission detector is integrally formed with the beam delivery system, as shown in Figures 4 and 5. A preferred transmission detector comprises a 16 channel array of dual energy detectors. The detector array further comprises pairs of detectors, including a low energy channel that receives and measures a first amount of radiation first (low energy) and a high energy channel that receives and measures a substantial portion of the balance of radiation (high energy). Dual energy detection has been described in connection with the linear scan arrays of the first inspection stage and is known to persons of ordinary skill in the art.

The low energy and high energy detectors measure a plurality of low energy and high energy values that can be used to characterize the material being scanned. In a preferred embodiment, low energy and high energy data are used to reference a look up reference, figure, table, or chart (referred to as a look up source) which contains transmission spectra arranged in accordance with corresponding high and low energy values. The look up source is constructed with high energy values on one axis (i.e. the x-axis), and low energy values on a second axis (i.e. the y-axis). Referring to Figure 6, an exemplary look up source 600 is shown. The source 600 is a graph with high energy values on the x-axis 605 and low energy values on the y-axis 610. Points 615 corresponding to measured spectra 620 are positioned on the graph according to certain linear combinations of the measured high and low dual energy detector signals on the x and y axis.

The transmission spectra used to normalize scatter data is therefore identified by obtaining high energy and low energy data values, identifying the point on the graph corresponding to the detected high and low energy values, and looking up the spectrum associated with that point. Where the detected high and low energy values yield a point on a graph that corresponds to an intermediate point 630 proximate to pre-established points 635, 615, a corresponding transmission spectra 645 can be calculated by performing a two-dimensional interpolation of the spectra 640, 620 associated with the pre-established points 635, 615.

To create the look up source, an exemplary approach places various materials of known composition and thickness, exposes them to X-ray sources, measures the resulting high and low energy data values, and uses the scatter detector to measure the corresponding transmission spectrum. More specifically, the beam path of the beam delivery system is modified to allow a direct beam from the focus through the pinhole to fall on the energy dispersive scatter detector. To further reduce the photon flux into a range that can be tolerated for energy-dispersive measurement, the current of the X-ray source is preferably reduced by a large factor, e.g. 100. Under these parameters, the scatter detector can be used to measure the transmission spectrum. Materials of known composition and thickness are placed in the beam path. The materials are exposed to X-ray radiation. Dual energy measurements are made using the dual energy detectors and a transmission spectrum is obtained using the scatter detector. Through this approach, for each material composition and thickness, a transmission spectrum is obtained and correlated with discrete pairs of dual energy transmission detector readings. This information is then arranged on a chart with the high energy value of the dual energy detector measurement on the x-axis, and the low energy value on the y-axis.

It should be appreciated that, in the disclosed embodiment, the spectra are the looked-up objects of the look up source. Instead of the spectra, however, the look up source can alternatively consist of spectral attenuation functions related to the attenuation of the materials placed in the beam when the look up source is being generated. The spectrum can then be obtained by multiplying one fixed spectrum, for example the spectrum measured without the material placed into the beam, with the spectral attenuation function retrieved from the look up source. Alternatively, the look-up source can contain numbers that are the parameters of analytical expressions, e.g. polynomials, which are formed to describe the attenuation functions in a parametric way.

The presently described approach is preferred because it enables the construction of a transmission detector array from lower cost materials, as opposed to constructing the array using more expensive energy dispersive detectors and support electronics. Moreover, it also addresses the difficult problem of using energy dispersive detectors to measure transmission spectra at the high flux rates that are experienced at the location of the transmission detector in the given configuration and at the same time at which the scatter data are recorded. The required strong attenuation of the transmission beams is a difficult problem that is avoided using the present invention. The look up table is an important element because the preferred dual energy detectors used in the transmission detector cannot deliver spectra directly.

As discussed, transmission spectra are being used to correct the scatter spectra that are being recorded by the energy dispersive detector. Normalizing scatter spectra with transmission spectra corrects for the confounding effects introduced by the specific spectral distribution of the primary radiation, as emitted from the X-ray source, as well as by spectrum-distorting effects known as beam hardening. To correct the scatter spectra, the detected scatter spectra are divided by the looked-up transmission spectra.

A normalized scatter spectrum exhibits a plurality of features. A first feature is that the location of the peaks and valleys of the spectrum are determined by the molecular structure of the materials located in the probe region. A second unrelated feature is that the average spectral signal of the normalized scatter signal, which can be of varying intensity, is linearly related to the gravimetric density of the material in the probe region. This can be used for threat discrimination since most explosives, particularly military explosives, have a density range above that of most other plastic or food items in suitcases.

In one embodiment, the normalized scatter signal is used to identify a threat item by comparing the obtained normalized scatter spectrum and/or spectral signal with a library of scatter signals from known threat items. This comparison can occur automatically by using a processor to compare a library of threat items, stored in a memory, with the obtained scatter signals. Such a library is developed by measuring the normalized scatter signatures of known threat items. In addition to using the transmission detector to generate data used to identify reference spectra, the transmission detector can function in a plurality of other ways. In one embodiment, the transmission detector acts as a position sensor. The transmission beam is interrupted or attenuated momentarily when an object on the conveyor crosses it. Tracking the moment of interruption can provide information on the physical position of the container on the conveyor and be used to appropriately position the beam delivery system or container.

In a second embodiment, the transmission detector array functions as an imaging detector to provide precise attenuation data for certain areas in containers, like container wall areas, where contraband can be hidden. When the circular beam is centered on an edge of a container, the edge of the container can be imaged in good detail, and can help analyze the edges for concealed threats.

In a third embodiment, transmission detector measurements can be used to determine whether the inspection region is, in fact, the same target region previously identified in the first stage scan. If the transmission data correlates with X-ray characteristics different than those obtained in the first stage scan, the relative positioning of the second stage scanning system and the object under inspection may be modified until the transmission data correlates with the same material characteristics that was identified in the first stage scan.

In a fourth embodiment, transmission detector data are also being used to simplify the algorithm-training procedure of the system, as described below, in particular the collection of threat material properties with irregularly shaped threat samples, like sticks of dynamite.

It should be noted that it would appear because the scatter radiation path and transmission path differ downstream from the scatter volume, there would be inconsistencies in the data when scatter and transmission data are combined. This inconsistency is one example of a number of partial volume effects, solutions for which are addressed herein. However, the inconsistencies are not significant and can be tolerated without encountering significant performance degradation of the system as a whole. As shown, Figure 4 is not an isometric schematic and, in reality, the scatter angle is preferably about 3 degrees, and the real path differences are comparatively smaller.

As previously discussed, the second stage scanning system positions an inspection region to physically coincide with the target region identified in the first stage scan. The positioning means may be achieved using any method known in the art. In one embodiment, a plurality of control commands is produced in response to the determination of the location of the target region. The control commands are generated by at least one processor in data communication with a plurality of processors capable of executing the aforementioned triangulation techniques and/or determining the intersection of projection lines to identify the location of the target region in three dimensional system coordinates.

The control commands comprise data signals that drive a three-axis control system. The vertical position of the second-stage inspection volume can be adjusted to the target volume or region of the first stage scan by moving the conveyor system up or down. In another embodiment, the control commands comprise data signals that drive the adjustment of the beam delivery system in the second stage scanning system. The beam delivery system adjustment can include any type of adjustment to the collimation or beam focus, including the physical movement of a plurality of apertures horizontally, vertically, or diagonally, the physical modification of the diameter of the ring aperture by, for example, increasing or decreasing the aperture size. In another embodiment, the position of the support structure, or C-arm, can be modified along the conveyor direction to appropriately position the beam delivery system.

The second stage scan may be compromised when the volume of the target region is smaller than the inspection region of the second stage. In such cases, extraneous material, other than the material identified as being a potential threat, such as air, metal, or container edges, may be included. The resulting scatter radiation is therefore a function of multiple material types and may not be readily identifiable as being the signature of a single substance.

In one embodiment, the present invention comprises a threat recognition process that incorporates a training methodology which relies on libraries in which threat signatures are obtained by combining the threat with other common materials, such as clothing, plastic, air, and metals.. Specifically, the data used in training and developing the detection process are chosen to include data, which are corrupted by errors based on partial volume data from statistically varying containers and threat and non-threat material combinations. When the inspection volume is partially filled with a threat substance and partially filled with a second innocuous substance, a combination signal will be detected by the second scanning stage. The automatic threat recognition methodology recognizes the threat from the combination signal based upon the aforementioned training. An exemplary automatic threat recognition methodology, based on neural networks, is described below.

In a second embodiment, the detected scatter data is corrected for the effects of extraneous materials by preprocessing the data. The motion control system tracks where the inspection volume or region is located in relative to a specific reference point, such as the approximate outlines of the container, and relative to the conveyor system. Because of the ability to measure and track these reference points, the amount and portion of the inspection volume occupied by the conveyor structure can be determined. The conveyor structure includes the belt material as well as the structural member that is underneath the conveyor, which is referred to as the slider bed.

To correct the scatter spectrum for the presence of the conveyor in the inspection volume, the scatter spectrum of the conveyor materials is measured and stored in a reference database. When the scatter spectrum of the inspection region is detected and it is determined that the conveyor occupied a portion of the inspection region, the scatter spectrum is corrected by multiplying the conveyor material scatter spectrum by a weighting factor to account for the size of the inspection volume occupied and that amount is subtracted from the measurement.

Similarly, when part of the inspection volume is filled with air, as in cases when suitcase walls are targeted by the inspection volume, it is known that the contribution of the air-filled portion of the inspection volume to the scatter signal is approximately zero, and therefore, substantially all of the scatter signal can be attributed to the material in the remainder of the inspection volume. By accounting for the air volume contribution, the characterization of the material in the remaining inspection volume is rendered more precise. Optical detectors, such as a plurality of light-curtains, can be positioned across and within the scanning system to generate control signals that convey information about the height and edges of the container relative to the conveyor system and relative to the inspection region. It therefore can be calculated which portion of the inspection region is filled with air.

In another embodiment, transmission values for the scatter beam are measured by an array detector. An exemplary array comprises 16 channels and yields transmission data for 16 subdivisions within the inspection volume. The transmission values can be used to characterize the material distribution in the inspection volume. Based on these transmission values, approximate mass values can be determined for masses contained in each of the 16 subdivisions. For example, where the transmission detector value returns a value indicating the subdivision has material with zero thickness, it can be assumed that the subdivision is occupied by air.

In a preferred embodiment, the inspection volume is subdivided. By reducing the size of the inspection region, one can ensure that fewer differing materials occupy the same region and can therefore avoid the complex composite signals that get generated when multiple materials fill a single inspection region. In one embodiment, system resolution is increased by providing multiple energy dispersive detectors, such as 2, 3, 4, 5, 6 or more, in place of a single energy dispersive detector as shown in Figure 4.

Referring to Figure 7, a schematic representation of the beam delivery system of Figure 4 780 is shown relative to a beam delivery system having multiple energy dispersive detectors 785. A first system 780 comprises single detector 700s, circular aperture 701s, inspection volume 702s, circular aperture 703s, and X-ray focus 704s. The dark areas represent the presence of radiation blocking material, e.g. 1/4inch lead alloy, and the white areas represent areas that are transparent to X-rays above 30 keV. A second system 785 comprises an X-ray focus 704q, circular aperture 703q, divided inspection volume 702q, detector side beam shaping aperture 701q, and quadruple detector 700q. The aperture 701q is center-symmetric and consists of four slits, each conforming to part of a circle. The centers of the circular slits are chosen to be of the same pattern as the detectors of the quadruple detector 700q. For example, if the detector cluster consists of four channels centered on the four corners of a 2 by 2 mm square, the centers of the partial and circular apertures lay on a circle with diameter equal to the square root of 2 times 2mm. The resulting inspection region for each individual detection region is about one quarter of the full inspection volume. A subdivided inspection region provides a higher spatial resolution of the second stage inspection. Clusters of energy dispersive detectors with their supporting electronics are commercially available from companies such as eV Products, Saxonburg, Pennsylvania.

If more than one scatter detector is being employed, a collimating system of vanes can be placed in front of the detector cluster orthogonal to the surface of the detector and in line with the plane of separation between each detector. Using a separator 705, diffracted radiation is more effectively limited to reach the appropriate channel in the cluster and, consequently, detected signals are more readily associated with materials from specific areas within the inspection region. The separator 705 extends from the surface of the detector cluster toward the surface of the adjacent aperture. The number of separator vanes is dependent on the number of detectors. A typical vane material and thickness is lead alloy of 0.5 mm thickness.

Referring to Figure 9, a flowchart summarizing the operational process of one embodiment of the present invention is provided. A container enters into the first stage scan 905 where it is exposed to a plurality of projected beams 915. From that exposure, X-ray characteristics are determined 920 and target regions containing potential threats are identified 925, 935. If no potential threats are identified, the container is not subjected to a second scanning stage 940. The three dimensional coordinates of the target region is determined 945 and, accordingly, the inspection region generated by the second stage scanning system is coordinated to coincide with the target region 950. The inspection region is subjected to X-ray radiation in order to obtain transmission and spectral data 955. The spectral data is then analyzed 960 to determine the existence of a threat. The data collected in the second stage scan comprises both localized dual energy transmission data and localized BRAGG diffraction spectra, which are subject to statistical variances, originating from photon signal fluctuations, partial volume limitations, or variations of the type of luggage and their contents, among other causes. As such, it is preferred to have a processing methodology that accounts for the fact that the raw data is not sufficiently sensitive to detect threats with sufficiently low false alarm rate.

In a preferred embodiment, the automatic threat resolution is performed by a probabilistic technique in which a plurality of input data points, obtained from the raw spectral scan data, contribute to the probability that the corresponding spectrum belongs to a particular class of threat or non-threat items. Such a probabilistic technique relies on the plurality of input data points as a whole rather than on individual data points. Although probabilistic classification techniques can include explicit, identifiable rules created by a programmer, the preferred techniques utilize a classification procedure that incorporates the results of training. For example, the classification algorithm can be used to process a training set consisting of patterns for structures of known classification. The results of this processing are used to adjust the algorithm, so that the classification accuracy improves as the algorithm learns by processing the training sets.

One type of trainable classifier that can be employed is an artificial neural network. Artificial neural networks attempt to model human biological neural networks to perform pattern recognition and data classification tasks. Neural networks are fine grain parallel processing architectures composed of non-linear processing units, known as neurons or nodes, which attempt to replicate the synaptic-dendritic interconnections found in the human brain.

Different types of neural networks exist. One type of network is a multi-player, feed-forward network. A feed forward network passes a signal by links from input nodes to output nodes, in one direction only. In most implementations, the nodes are organized into multiple layers: the input layer, output layer, and several "hidden layers" in between. The adjacent layers are normally fully interconnected.

Figure 8 depicts a schematic representation of a preferred type of artificial neural network 800 known as a hidden-layer feed-forward network consisting of an input layer 810 of neurons or nodes, at least one hidden layer 820, and an output layer 830. The neuron layers are linked via a set of synaptic interconnections. Each neuron in the input layer is typically connected to each neuron in the hidden layer, and each neuron in the hidden layer is typically connected to each neuron in the output layer, via a synaptic connection; these may be physical, electronic connections, or they may be embodied in software, as may be the neurons themselves, which software operates on conventional digital computers.

The neurons or nodes typically accept several inputs and create a weighted sum (a vector dot product). This sum is then tested against an activation rule (typically a threshold) and then processed through an output function. The output function could be a non-linear function such as a hard-limiter; a sigmoid function; a sine-function or any other suitable function known to a person of ordinary skill in the art. The threshold determines how high the input to that neuron must be in order to generate a positive output of that neuron. A neuron may be considered to be turned on, for instance, whenever its value is above a predetermined value such as, for instance, 0.9 and turned off with a value of less than another value such as 0.1, and has an undefined "maybe" state between those values. The connectivity pattern defines which node receives the output value of a previous node as their input. The connection between two neurons is realized in mathematical terms by multiplying the output of the lower level neuron by the strength of that connection (weight). At each instant of propagation, the values for the inputs define an activity state. The initial activity state is defined upon presentation of the inputs to the network.

The output response of any hidden layer neuron (o_j) and any output layer neuron is a function of the network input to that neuron defined by the difference of that neuron's threshold (θ) and the input to it. The value of the input into each hidden or output layer neuron is weighted with the weight currently stored for the connection strengths between each of the input and hidden layer neurons, and the hidden and output layer neurons, respectively. Summing over all connections into a particular neuron and subtracting this sum from the threshold value may be performed according to the following sigmoid- type Fermi function:<maths id="math0001" num=""><math display="block"><mrow><msub><mrow><mtext>o</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><msub><mrow><mtext> = [1+ exp ( θ</mtext></mrow><mrow><mtext>j</mtext></mrow></msub><msub><mrow><mtext> - ∑</mtext></mrow><mrow><mtext>i</mtext></mrow></msub><msub><mrow><mtext> . w</mtext></mrow><mrow><mtext>ji</mtext></mrow></msub><msub><mrow><mtext> * o</mtext></mrow><mrow><mtext>i</mtext></mrow></msub><msup><mrow><mtext>)] </mtext></mrow><mrow><mtext>-1</mtext></mrow></msup></mrow></math><img file="EP1526392A2_D0001.tif" /></maths> where i and j represent neurons of two different layers with j representing the higher layer; θ_j represents the bias value for j layer neuron; and w_ji represents the strength of the connection between neuron i and neuron j. Alternatively, sine-type functions, or any other suitable function known in the art, may be used to obtain the desired type of response function for the output of a neuron. The weights are chosen so as to minimize the error between the produced result and the correct result. A learning rule defines how to choose the weight values. Several commonly used learning rules are back- propagation, competitive learning, adaptive resonance, and self-organization.

In a preferred embodiment, the artificial neural network uses back-propagation learning. The back-propagation learning algorithm, derived from the chain rule for partial derivatives, provides a gradient descent learning method in the space of weights and can be further understood by reference to D. E. Rumelhart, et al., Parallel Distributed Processing, ch. 8, pp. 322-28 (MIT Press, 1986) and Haykin, Simon (1999), "Neural Networks", Prentice Hall, both of which are incorporated herein by reference.

Back-propagation learning involves a set of pairs of input and output vectors. The network uses an input vector to generate its own, or actual, output vector. The actual output vector is compared with a desired output, or target, vector that may be defined usually in the course of training. The weights are changed to obtain a match between the target vector and the actual output vector. The conventional delta rule may be used for this calculation where the weight for a particular synapse or connection between units is adjusted proportionally to the product of an error signal, delta, available to the unit receiving input via the connection and the output of the unit sending a signal via the connection. If a unit is an output unit, the error signal is proportional to the difference between the actual and target value of the unit. If it is a hidden layer, it is determined recursively in terms of the error signals of the units to which it directly connects and the weights of those connections.

Thus, the training of a neural network is the process of setting the connection weights so that the network produces a desired output in response to any input that is normal for the situation. A supervised training refers to the kind of training that requires a training set, i.e. a set of input-output patterns. The back-propagation algorithm is an efficient technique to train a feed-forward network. It operates to send an error back through the network during the training process, thereby adjusting all the link weights in correspondence with their contribution to the error. The weights of the network therefore gradually drift to a better set of values. The initial weights are chosen randomly within reasonable limits and adjustments are left to the training process.

Referring back to Figure 8, the artificial neural network 800 is trained on a suitably large set of threat and non-threat X-ray raw scan data, to generate an output 840, in accordance with the error back-propagation learning method described above. As described earlier, the required set of threat and non-threat raw scan data for training can be obtained either from the scanning system of the first stage or the scanning system of the second stage or both depending upon whether artificial neural networks are used to process scan data from the first stage or the second stage or from both the stages. Thus, the 'scan data' 805 to be used to train the neural net 800 may comprise of raw attenuation data, raw transmission photon counts, raw diffraction photon spectra or any other data known to a person of ordinary skill in the art.

The purpose of the neural network processing step is to have a processing means capable of recognizing a threat signature. A threat signature is defined as a spectrum, i.e. an array of numbers corresponding, on a one-to-one basis, to the discretized values of a physical quantity, such as the energy of X-rays, and includes unrelated, but relevant, other values, such as transmission detector array data, bag height, and other environmental factors. Although the spectrum may consist of any amount of data points, the present invention preferably operates on a spectrum data set of between 200 and 800 points and, more preferably, of approximately 500 points. Additionally, while the network may consist of any number of layers, it is preferred that it consists of four layers, including one input layer, two hidden layers, and one output layer. Further, while the network can have multiple output nodes with various indicators, it is preferred that, for the present invention, the network comprise a single node the output of which may be interpreted as either "yes, a threat has been recognized" or "no, a threat has not been recognized".

In a preferred embodiment, the raw diffraction spectrum data from the second stage is used to generate the required set of threat and non-threat scan data for training. These spectral counts represent raw, that is non-normalized, scan data 805 that is subsequently used to train the neural network 800. Alternatively, this scan data 805 may be further processed to generate a plurality of normalized data.

The scanning process is repeated to obtain scan data of a sufficiently large number of containers containing threat and non-threat items packaged in a variety of permutations and combinations to model real-world scenarios. This raw scan data, referred to hereinafter as training data, comprise an input-set to be used for training the neural network. Since the training data is obtained by scanning containers containing known materials, each output training data maybe further tagged to identify whether the respective training data represents a defined/known threat or non-threat item. This output training data maybe further stored in a suitable library or database such as a file server on a digital computer system along with the tagged identification information. Furthermore, the library or database of training data may be enhanced to incorporate and reflect all previously known threat materials and their corresponding raw 'scan data'.

In a preferred embodiment, two libraries are generated. A first library has signatures of threats. A larger second library has signatures of innocuous, or non-threat, items. The training process utilizes the threat and non-threat signatures to introduce into the system threat-like and non-threat-like signatures.

A threat-like signature is a linear combination of a sample from the threat library with a plurality of samples, such as two, from the non-threat library. The coefficients of the mix are randomly simulated. A simulated white noise is also added to the generated mixture, with its amplitude also randomly generated, within an interval from zero to a given fraction of the signal. A non-threat-like signature is a mixture of a plurality of non-threat signatures, such as two or three. The coefficients of the mix are randomly simulated. A simulated white noise is also added to the generated mixture, with its amplitude also randomly generated, within an interval from zero to a given fraction of the signal. Using signature mixes, incorporated with noise, the system is trained to recognize a threat or a non-threat by outputting an appropriate recognition answer from the last output node within the context of a reasonable level of noise and interference from overlapping items.

Figure 10 depicts a plurality of steps, in flow diagram format, of one embodiment of the back-propagation training process of the invention. One of ordinary skill in the art would appreciate that the processing is conducted using a computer having a plurality of processors for executing the analytical processes described herein, embodied in at least one software program, a plurality of storage devices for storing the requisite data, library information, and other information necessary to conduct these analyses, and an output device, such as monitor, among other commonly known computing devices and peripherals.

At the beginning of the training process 1003, the synaptic weights and thresholds of the neural net are initialized 1005 with, for example, random numbers. After initialization 1005, the input layer of the neural network is introduced 1010 to a first set of training data and the net is run to receive 1015 an actual output. The neural net makes use of the randomly assigned weights and thresholds to generate an output on the basis of a suitable resolving function such as a sigmoid-type Fermi equation (described earlier), a sine function or any other function known to a person of ordinary skill in the art. The output could be in the form of differentiable signals such as numerals between, say, 0 and 1, in the form of positive or negative states implied by an output numeral of greater than or less than 0 respectively, or any other suitable indication as evident to a person of ordinary skill in the art.

The first set of training data is introduced into the system and, based on the random weights and thresholds, produces an output 'x', i.e. a numeral greater than 0. If the training data represents a threat, this output indication is set as a benchmark to identify a 'threat' while a numeral less than 0 maybe set to identify a 'non-threat' item. Once a suitable benchmark is set, the training process is repeated with the next set of training data and corresponding actual outputs are received. The actual output is compared 1020 with the desired output, defined by an operator with knowledge as to whether input data is or is not representative of a threat, for the corresponding set of training data that was fed to the neural net in step 1010. If the actual output received 1015 is commensurate with the desired or targeted output or, if the difference between the target and actual output falls below a predefined acceptable level, a check 1025 is made to see if the neural net has been trained on the entire set of training data. If not then the next set of training data is introduced to the neural net and the foregoing steps are repeated. The training process continues until the neural net has been trained on the entire set of training data.

If the comparison 1020 suggests that the actual output is not in agreement with the desired or targeted output, the ensuing additional steps are performed. The difference between the actual and desired outputs is used to generate 1030 an error pattern in accordance with a suitable back-propagation rule such as the 'delta rule' or any other error estimation rule known to a person of ordinary skill in the art. The error pattern is used to adjust 1035 the synaptic weights on the output layer such that the error pattern would be reduced the next time if the same set of training data were presented as the inputs. Then the weights of the hidden layers, preceding the output layer, are modified 1040 by comparing what outputs they actually produce with the results of neurons/nodes in the output layer to form an error pattern for the hidden layer.

The error can thus be propagated as far back over as many hidden layers as constituting the neural network. Finally, the weights for the input layer are similarly educated 1045, and the next set of training data is introduced to the neural network to iterate through the learning cycle again. The neural network is therefore trained by presenting each set of training data in turn at the inputs and propagating forwards and backwards, followed by the next input data, and repeating this cycle a sufficient number of times such that the neural network keeps getting closer and closer to the required weight values each time. Thus, the network, through the iterative back-propagation process, establishes a set of weights and thresholds for neural connections so that a desired output pattern is produced for the presented input information. The learned information of a neural network is contained in the values of the set of weights and thresholds.

In exemplary embodiments, the neural network is structured such that, through iterative forward and backward propagation, every node in a layer can be made to contribute to every node in a subsequent layer, only certain nodes in a layer maybe used to contribute to certain nodes in a subsequent layer, or every node in a layer contributes to every node in a subsequent layer but the impact of certain first layer nodes on subsequent layers are weighted relative to other first layer nodes. In a preferred embodiment, the nodes closest to subsequent layer nodes are weighted relative to other nodes in that same layer.

More preferably, links between the input layer and the first hidden layer are not chosen randomly, but selected to have a special distribution. Each hidden layer node is responsible for a region of the spectrum. Links to each hidden layer node from this region have higher weights. Therefore, the farther an input node is from this region, and the less responsible it is, the weaker the link with that input node. Together, the hidden nodes encompass the entire input layer spectrum. By distributing a pre-assigned degree of influence over links, a form of convolution is provided. This embodiment is particularly preferred where preprocessing is not reasonably effectuated because the input data size is too large.

Because new threats may develop over time, it is desirable to have a simple procedure that updates the network to recognize such additional threats. In a preferred embodiment, multiple networks are formed and trained to identify distinct threats. Therefore, new threat recognition is done by implementing a neural network as a set of multiple networks, each trained to identify a specific threat.

Each network group is formed and trained to address and recognize one threat. Thus, there is a one-to-one correspondence between the threats (T₁, T₂, T₃...T_n) and the groups (G₁, G₂, G₃...G_n). Network group G_n is trained to recognize threat T_n. Where G_n recognizes any other threat, T_n-2, T_n-1, T_n+1, it is not considered relevant. G_n is trained using threat signatures of T_n and a corresponding library of non-threat signatures. A group consists of a plurality, such as two, three, or four, completely separate networks, each similarly trained to recognize the same threat. Threat recognition is achieved where the average of the recognition results of each network indicates a threat. Where an additional threat is identified, T_new, a new group of networks, G_new, can be created, without having to retrain or modify all existing groups. This permits the more efficient, incremental addition of new threat recognition networks.

One of ordinary skill in the art would appreciate that the output of these network groups can be handled in various ways. Specifically, a system can output a threat alarm if the recognition result, averaged over all network groups, indicates a threat. A system can output a threat alarm if only one group of networks indicates a threat or if a subset of network groups indicates a threat. In a preferred embodiment, threat recognition is effectuated by monitoring the output of substantially all groups of networks. If the recognition result, averaged over all groups, indicates the existence of a threat, the output results of individual network groups are analyzed. If at least two groups indicate the presence of a threat, then the groups are reviewed to determine which threat has been recognized. One of ordinary skill in the art would appreciate that various derivations of the above-described process can be conducted without departing from the scope of this invention. For example, the threshold analysis can be performed even if only a portion of all network groups is monitored and the secondary analysis can be performed if fewer than two groups indicate a threat.

It is further preferred to regulate the balance between the sensitivity of detection with the selectivity of detection by incorporating an additional input node that is used to inject a sensitivity level into the training process. Systems with higher sensitivity will detect more threats at the expense of having greater false alarms. Systems with higher selectivity will have fewer false alarms, with the disadvantage of possibly not detecting all threats. In the course of operation, it may be necessary to change the sensitivity/selectivity balance, depending on various circumstances, and therefore it is desirable to have a means for doing so.

In one embodiment, a plurality of different networks is stored in a storage device in data communication with processors responsible for executing neural network analytical processes. Each set of networks has a different level of sensitivity, i.e. least sensitive, less sensitive, normal sensitivity, more sensitive, most sensitive. Depending on the requisite level of security (versus requisite level of throughput), the appropriate network set is loaded into the system. Alternatively, a network set having a standard level of sensitivity may be used in operation and a parallel network set, having varying levels of sensitivity, may be concurrently loaded and ready for use, when necessary. Having a parallel network avoids downtime associated with loading new network sets into local memory, or RAM.

Networks having varying degrees of sensitivity are developed by incorporating a sensitivity variable in the training process. With each recognition task, a desired level of sensitivity is communicated to an additional input layer node, thereby inherently incorporating it into the training process. Every training event could further be associated with a randomly selected sensitivity level, selected from within a reasonable range. Training is therefore conducted with the selected level of sensitivity.

The embodiment of back-propagation learning process, as described with reference to the flow diagram of Figure 10, assumes that the training data is first collected by operating the first and/or second scanning systems offline, namely by scanning a large number of baggage containing known threat materials camouflaged amongst non-threat items in a variety of combinations to represent a variety of concealment scenarios. The training data so obtained is then used to educate the neural network that can then be used to operate in real-world situations. Thus, in this embodiment, when the first and/or second stage systems are online, that is operational at real-world sites such as at airports for luggage inspection, the associated neural network is already partially educated (on the training data obtained through test baggage) to discern threat from non-threat items.

However, in another embodiment of the back-propagation learning process, the neural network need not be taught through the scanning of test baggage prior to running the first and/or second scanning systems online. Instead the first and/or second scanning systems can be operated online and the scan data can then be fed into the neural network in real-time. On the basis of this real-time scan data, the neural net is made to classify threat and non-threat items. At the same time, the scanned image of the baggage is also presented to an operator in the form of a visual display such as on a conventional computer screen, as is known in the art. The operator compares the output of the neural net with his own observation of the scanned baggage. In case the output of the neural net is found to be erroneous, the operator prompts the neural net with the correct or desired output, enabling the neural network to adjust its weights and thresholds accordingly. Thus, in the second embodiment when the system is first used, it will have relatively little knowledge about threat and non-threat materials to be identified and recognized. However, with sufficient positive reinforcement of a relationship between the acquired scan data and operator prompted identification, the neural network will learn how to identify threat objects. This self-learning process enables the neural network to learn continuously.

The on-line, self-learning training process does have certain advantages. A company or organization that uses the present invention may not want to share or disclose data to third parties due to privacy or security reasons. Therefore, it may be essential to enable self-learning. Furthermore, the flow of data may change seasonally depending upon how containers change. Specifically, seasonal variations do occur in airline passenger travel where passenger bags may get larger in the winter to accommodate more clothing or the number of total bags may increase due to larger numbers of people traveling in the summer. To address such seasonal variation, it is more practical to allow on-line autonomous adaptation.

Finally, there may be a variety of system users in different geographical regions that experience different types of threat and non-threat items. In such cases, a standard library may not be as helpful as self-taught systems that automatically learn in accordance with its own unique context. More specifically, over time, a system in the field will be trained on containers that have threat and non-threat items unique to their geographic context. Due to operator training and interaction, a particular system would therefore develop a trained processing system tailored to their geographical context. It is preferred, however, that, irrespective of the geographical location, systems get periodically trained using threats that are new or infrequently seen to ensure that the system does not forget the identity of such threats. This update could be performed by the statistically controlled reinjection of threats from existing threat databases.

Although this online self-learning process has been described separately as an embodiment, this continuous self-learning process can be used in conjunction or combination with the offline teaching process of the first embodiment, using test baggage. In a preferred embodiment, the neural network is first trained on scan data obtained by running the X-ray system offline on test baggage and then through operator prompts in real-time operations as well, so that the ability of the neural net, in identifying objects, continuously improves through self-learning. Nevertheless, the system may undergo retraining offline using data from multiple site locations, thereby taking full advantage of the sum total of learning being generated by the operation of multiple systems.

Operationally, acquired scan data is fed into the neural network for identification. If the object is identified with a high degree of confidence, the identification and scan image is conveyed to an operator, along with an indication of what the object may be, to enable the operator to take an action, including conducting a hand search, questioning the container owner, permitting the container to pass, or calling in additional personnel. In one embodiment, the operator provides feedback to the system based upon the identity of the threat/non-threat and action taken. For example, if the system identifies the existence of a threat, the operator can check the container to determine if a threat exists and then inform the system whether it was or was not correct. If correct, the neural network increases its confidence factor for that object's scan data and stores the scan data in a suitable database as an exemplar for retraining. If incorrect, the neural network implements the error back-propagation process to suitably adjust its weights and thresholds and stores the scan data in a suitable database as an exemplar for retraining.

This on-line adaptation process using incoming data requires certain precautions, however. It is preferred that the system utilizes groups of networks, which are accompanied by libraries of threat patterns and innocuous patterns, and that the system is not authorized to modify these libraries. It is further preferred to include an additional library, a buffer-library, that is available for modification based upon incoming data. Specifically, the buffer library comprises incoming new data, and is preferably periodically cleansed of older data. Consequently, the networks are being re-trained using the buffer-library and the two stable libraries, with a proper adaptation time scale. Several previous versions of network groups are stored as a back up and a comparison of newly adapted system with its older versions can be conducted and produced in the form of a report. As described, on-site training can be set up as an automatic feature, but operator input may be required in the rare case of a real alarm.

It is the intention of the inventors to embody within the scope of this patent all embodiments that reasonably and properly come within the scope of this specification. For example, while the dual-stage scanning system has been described with reference to a first stage scanning system comprising a dual-view line scanner and a second stage scanning system, comprising a transmission and scatter scan, other modifications and changes can be made by those of ordinary skill in the art on the basis of the number of beam projections and the singularity or duality of incident beam energies used. Additionally, while the details of processing the scan data using trained classifiers has been used with the second stage scanning unit, one of ordinary skill in the art would appreciate that it could be used with the first stage unit to better help identify threats or target regions