Methods of non-uniformity compensation for infrared detector arrays

14119/01 ϋ1ΤΝ-Ν19ΓΝ A7J Him ΠΙΤΠΝ-Α ^97 Π1Ί911ϋΠ mtPtll

IMPROVED METHODS OF NON-UNIFORMITY COMPENSATION FOR INFRARED DETECTOR ARRAYS PCTAJSOO/17622 WO 01/03423 1

IMPROVED METHODS OF NON-UNIFORMITY COMPENSATION FOR INFRARED DETECTOR ARRAYS

TECHNICAL FIELD OF THE INVENTION

The present invention relates generally to infrared detectors, and more particularly to a method of compensating for non uniformities among detector elements 5 of an infrared detector array.

BACKGROUND OF THE INVENTION

Infrared detectors provide thermal images for temperature measurement and heat detection. They are 10 used for various applications, such as for military, industrial, and medical applications. In its simplest form, an infrared detector is a device, such as a photosensitive diode, that generates an electric current when exposed to infrared radiation. This current is 15 dependent on the intensity and wavelength of the radiation and can be used in many different ways to produce an infrared picture.

Infrared detectors may be configured as a single element (detector), a small array of elements, a long 20 linear array, or a full two-dimensional array. When the detector is a full two-dimensional array, the entire image is recorded at once, and the array is referred to as a "staring" array. However, with smaller arrays, the image is scanned over the array. The small array 25 requires a serial scan to sweep the image in two-dimensions, whereas the linear array requires a "pushbroom" scan to sweep the image across the array in one dimension.

The current produced by an infrared detector is 30 amplified and processed to provide a more useful detector WO 01/03423 PCT/USOO/17622 2 output. The processing reduces interference due to external and internal causes, such as electrical noise.

The ideal response of an infrared detector array is that each detector element exhibit the same linear 5 voltage response for given temperature changes in the irradiation of the array. However, one type interference with a good detector signal is electrical noise due to detector non-uniformity among detector elements. The uniformity differences have both spatially and temporally 10 dependent causes. A number of methods have been tried for compensating non uniformity of infrared detector arrays. Generally, all involve some sort of data processing. Some methods use a uniform calibration source, typically using a 15 chopper and controlled temperature. Other methods are scene-based, which means that they use an image comprised of one or more objects or patterns. The scene-based methods may be further categorized into mechanical and non-mechanical methods. 20 The "dithered scan" method of non uniformity compensation is a scene-based mechanical method. The detector array views a scene through suitable optics. During a given time frame, the incident flux is sensed by each detector element. At the end of the time frame, the 25 array data is delivered for processing and the array is displaced ("dithered") a fixed distance, typically a distance equal to the width or height of one detector element, in either, the horizontal or vertical direction. The scene flux is assumed to be stable throughout the 30 dither cycle. Thus, during the next time frame, each detector element is exposed to the flux seen by one of its neighbors during the prior time frame. These WO 01/03423 PCT/USOO/17622 3 detector pairs can be "linked" analytically, such as by averaging their outputs. By a suitable choice of a dither pattern, each detector can be linked with several of its neighbors, to adjust gain and offset differences. 5 A dithered scan approach is described in an article by

William F. O'Neil, "Dithered Scan Detector Compensation", Proc. IRIS Passive Detectors, 1992, Vol. 1, pp. 123 -134.

10 SUMMARY OF THE INVENTION

One aspect of the invention is a "one-step" method. of calculating offset correction values for detector elements of an infrared detector array. The array is dithered with respect to the scene that is to be imaged. 15 The dither motion is such that pairs of neighboring detector elements are exposed to the same scene flux.

The method assumes the scene flux is stable over the dither cycle. With this assumption, neighborhood averaging is used to adjust offset differences, based on 20 their responses to the same scene irradiance. The neighborhood averaging uses the sum of "shifted image differences", where shifted images (matrices of detector output values) from a first field and a second field, respectively, are differenced. This sum is divided by 25 four. A scene term may be added to compensate for the bias.

The "one-step" method is the mathematical equivalent of "two-step" methods. Specifically, for offset correction calculations, a detector element and its 30 neighbors are exposed to the same scene flux. This permits calculation of a local response average, which is subtracted from the detector element1s output to WO 01/03423 PCT/US00/17622 4 determine the output correction for that detector element.

Generally, for a given image, a complete nonuniformity correction process involves calculations for each detector element. A multiplicative gain correction factor is calculated and applied to that detector element's output. The method is iterative, such that new gain correction updates become constant (equal to unity), reducing gain errors to smaller and smaller errors. Then an offset correction is interactively calculated and subtracted, such that the new offset correction updates become constant (equal to zero) . The image can be displayed during the iterations with successively smoothed gain and offset errors, or the display can be delayed until the correction values become stable to some predetermined level.

The method can be mathematically described with matrix notation, which shows that the method is equivalent to a low pass filtering process. For each detector element, the calculations use local neighbor averages, which. include values representing the detector element whose error is being calculated as well as values representing its neighbors. As a result, the method is a true low pass filter, as compared to previous techniques. It better smooths the high frequency components of the gain and offset matrices. WO 01/03423 PCT/US00/17622 5

BRIEF DESCRIPTION OF THE DRAWINGS FIGURES IA and IB are block diagrams of infrared detector systems in accordance with the invention. FIGURE 2 illustrates a portion of a scene to be viewed by a one dimensional or two dimensional detector array. FIGURES 3A and 3B illustrate the basic steps of the method of the invention, as applied to a one-dimensional array. FIGURES 4A and 4B illustrate a one dimensional array with gain errors and its dither motion. FIGURES 5A and 5B illustrate a one dimensional array with offset errors and its dither motion. FIGURES 6A and 6B illustrate the basic steps of the method of the invention, as applied to a two-dimensional array. FIGURE 7 illustrates the dither pattern for a two-dimensional array. FIGURE 8 illustrates a detector element and its neighbors for purposes of gain correction for a two-dimensional array. FIGURE 9 illustrates a detector element and its neighbors for purposes of offset correction for a two-dimensional array. FIGURES 10 and 11 illustrate an alternative dither pattern for offset correction for a two-dimensional array. FIGURES 12 and 13 illustrate two more diagonal dither patterns for two-dimensional arrays. FIGURE 14 illustrates an alternative rectangular dither pattern for offset correction for two-dimensional arrays. WO 01/03423 PCT/US00/17622 6 FIGURE 15 illustrates an alternative rectangular dither pattern for gain and offset correction for two-dimensional arrays. FIGURE 16 illustrates the smoothing of gain errors for a "bland" scene. FIGURES 17A - 17D illustrate how two detector elements might be exposed to different scene paths, as a result of dither bias.

DETAILED DESCRIPTION OF THE INVENTION

The following description is of a method of calculating gain correction values and offset correction values for infrared detector arrays. The method is also described in U.S. Patent No. _ (U.S. Patent

Application Serial No. 08/918,642), entitled Non Uniformity Compensation for Infrared Detector Arrays", to J. Young, et al. , assigned to Raytheon Company.

The array is dithered so that two or more neighboring detector elements of the array look at the same location of a scene. A dither motion occurs once per field so that each new field contains pixel values for the same location of the scene but viewed by a different detector element. The scene is assumed stable, so that if there were no gain or offset errors, a pixel value would be the same in both fields even though it was sensed by different detector elements. Then, the fields of pixel data are processed to calculate a gain correction value and an offset correction value for each detector element. The original fields of pixel data are corrected by multiplying each detector element's response by its gain correction and by subtracting its offset correction. WO 01/03423 PCT/US00/17622 7

As explained below, for each detector element, the gain and offset errors are calculated from local averages. The local average for a particular detector element includes a term for that detector element as well as terms for its neighboring detector elements. For offset correction, local averages are calculated for "opposing" sets of neighboring detector elements. Gain and offset errors are "smoothed" in successive iterations of the method, making the method equivalent to low pass filtering the errors. More specifically, the gain and the offset of each detector element can be expressed in matrix form as [Gain] and [Offset], respectively. These matrices are shifted, summed, and averaged, in operations that can be represented as convolutions. The convolutions are with "masks" that result in low pass filtered versions of [Gain] and [Offset]. The process is recursive so that after successive iterations, [Gain] and [Offset] become constant.

The following description includes embodiments of the method for one-dimensional and two-dimensional arrays. In the one-dimensional method, the dither is in one direction. In the two-dimensional method, the dither is in two directions. For both methods, the embodiments described herein compute local averages with values from "near neighbor" detector elements, resulting in a filter domain of a certain size. In other embodiments, local averages could be computed with a larger "neighborhood", with computations modified accordingly, to provide a larger filter domain.

The present invention is directed to improvements of these methods. A first improvement is a simplified "one- step" method that can be used for both one-dimensional WO 01/03423 PCT/US00/17622 8 and two-dimensional arrays. A second improvement is directed to compensation for dither bias as well as non uniformity. 5 System Overview FIGURE IA is a block diagram of a one-dimensional infrared detector system in accordance with the invention. A detector array 11A has a number of individual detector elements DA_N. In the embodiment of 10 FIGURE IA, array 11A scans horizontally. However, as explained below, the invention is easily modified for vertical scanning.

Each detector element DA.N of array 11A provides a response signal, indicating flux irradiance, on a 15 separate channel. Each channel has an amplifier 12 and an analog-to-digital. converter 13 that samples the signal. The N data signals are delivered in parallel to processor 14. At each time, t, the data received by processor 14 are pixel values representing intensity a 20 given sample time.

It is assumed that the detector elements DA.N of array 11A have non-uniform responses. That is, at least two of the detectors, and probably all of them to one extent or another, do not provide the same response to a 25 given temperature in the scene. The invention is designed to compensate these non uniformities regardless of whether they are linear or non-linear.

As indicated by the arrows in FIGURE IA, a dither motion is achieved by displacing array 11A in the 30 vertical direction. In a simple implementation of the method, the displacement is equal to the height of one detector element D. This can be accomplished with WO 01/03423 PCT/USOO/17622 9 mechanical stepper motors or other opto-mechanical means known in the art. After processing, the image data may be manipulated so that the images register in the same place for display. 5 Processor 14 may be any device capable of performing the mathematical operations described below. Therefore, processor 14 might be a general purpose processor programmed to perform the appropriate operations with programming stored in a memory 14a. Or processor 14 10 might be dedicated logic circuitry or of some combination of software and hardware driven devices.

The output of processor 14 is a corrected signal, from which an image can be generated. The embodiments of this description are designed to provide real time 15 correction for displayed images, but the same method could be used for stored image data. FIGURE IB illustrates the same system as FIGURE IA, but with a two-dimensional array 11B. As explained below, for two dimensional arrays, in the example of this 20 description, the dither is in both the vertical and horizontal direction. The invention can also be modified for a diagonal dither pattern. FIGURE 2 illustrates a portion of a scene to be viewed by a one-dimensional array, such as array 11A, or 25 by a two-dimensional array, such as array 11B. The scene is divided into a number of pixels, each representing an area whose irradiance flux is sensed by a detector element of array 11A or 11B. For simplicity of illustration, only the pixels sensed by a 4x4 portion of 30 detector elements are explicitly illustrated. In practice, a number of pixels corresponding to the scan of array 11A or to the size of array 11B comprises an image WO 01/03423 PCT/USOO/17622 10 "field". The true flux incident upon each pixel, indexed by (x,y) coordinates, is <$(x,y) . The pixel values for two adjacent pixels, at (x,y) and at (x+l,y) , are shown as being light grey and dark grey, respectively.

For a two-dimensional array, such as array 11B of FIGURE IB, the desired response of the array to the sce:je flux of FIGURE 2 can be represented as: [Image] = g · [Φ] + o , where [Image] is a two-dimensional matrix of the response value of each detector element D and [Φ] is a two-dimensional matrix of values representing the actual scene flux at the location viewed by each detector element D. Each value in [Φ] corresponds to one pixel location in the scene space. In this ideal response, gain and offset are global values, g and o.

An image corrupted only by fixed pattern noise can be represented as: [Image] = [Gain] · [Φ] + [Offset] , where [Gain] is a two-dimensional matrix that contains the gain values for each detector element D, and [Offset] is a two-dimensional matrix that contains the offset values for each detector element D. Each value in [Gain] is comprised of the global gain, g, plus the detector element's gain error. Each value in [Offset] is comprised of the global offset, o, plus the detector element's offset error.

The dithering technique described below uses dithering and local neighbor averaging td smooth the high frequency components of the [Gain] and [Offset] matrices. The gain correction is achieved by multiplying a gain correction term to equalize the response gain ratios between neighboring detector elements. Offset correction WO 01/03423 PCT/US00/17622 11 is achieved by subtracting an. offset correction term to reduce the difference between individual detector element responses and the local . average responses. After gain and offset correction, the image could be represented as: [Corrected Image] = [Image]·[Gain Correction] - [Offset Correction] = [Smoothed Gain] · [Φ] + [Smoothed Offset]

As explained below, the gain and offset correction process recursively corrects the images to achieve smoother and smoother gains and offsets until the following result is achieved: [Smoothed Gain] = g = constant , and [Smoothed Offset] = o = constant

For a one-dimensional array, such as array 11A of FIGURE IA, the above matrices would be replaced by column vectors.

Non Uniformity Compensation for One-Dimensional

Arrays FIGURES 3A and 3B illustrates the basic steps of the non uniformity compensation method, as applied to a one-dimensional array. The embodiment of FIGURES 3A and 3B is for a horizontally scanning array, such as array 11A of FIGURE IA, most of whose detector elements have an "upper" and a "lower" neighboring detector element. The exceptions are the "end" detector elements, for which calculations are modified. The embodiment of FIGURES 3A and 3B could be adapted for a vertically scanning array, whose detector elements would have a "left" and a "right" neighboring detector element. In the case of either a WO 01/03423 PCT/US00/17622 12 horizontally or vertically scanning array, the method involves two "near neighbors" of each detector element that is not an end detector element, regardless of whether these near neighbors are "left and right" or "upper and lower" neighbors.

In the embodiment of FIGURES 3A and 3B, a horizontally scanning array, such as array 11A receives data column-by-column. Both gain and offset corrections are also performed column-by-column, consistent with real time data acquisition and processing. Gain correction calculations are performed with data from two adjacent columns of two consecutive fields. Offset correction calculations are performed with data from the same column from two consecutive fields. For a vertically scanning array, the processing would be equivalently performed on horizontal "rows" of data from left and right neighbors.

The calculation of gain and offset correction values is not necessarily synchronous with display of the image. In the example of FIGURES 3A and 3B, the iterations for gain correction values are repeated until they are constant, then the iterations for offset correction values are repeated until they are constant, before the image is corrected and displayed. Thus, pixel data is received and stored and the method of FIGURES 3A and 3B is performed, before the data is corrected for display. Alternatively, uncorrected data and partially corrected data may be displayed as the iterations of FIGURES 3A and 3B are being performed.\ Also, the data may be output for further processing or storage in memory. The decision of when to use the correction values to generate a display is a matter of time or memory constraints, or of the overall system configuration. WO 01/03423 PCT/USOO/17622 13

Steps 301 - 306 are the gain correction part of the method. A gain correction value for each detector element is calculated before the offset correction values. This is because the offset errors do not affect the gain calculation.

Step 301 is receiving the data for a first field. A field is a set of data, acquired during a complete horizontal scan of array 11. Thus, one field contains the pixel values representing a complete scene. FIGURE 4A illustrates Step 301 in simplified form, using a horizontally scanning array 41 and scene 42. For simplicity of explanation, array 41 has only three detector elements, DA.C, and scene 42 has only 4x4 locations to be represented as pixel values. Each location of scene 42 has a "true" flux value, ranging from 1 to 16. The gains of the three detector elements, Da, Db, and Dc, are 1, 2, and 3, respectively. As array 41 scans the first field (i.e., Field 1), it provides pixel values, column-by-column. With the detector element gains, the received pixel values (detector outputs) of Field 1 are:

Field 1 2 4 6 8 6 12 18 24 12 24 36 48

As explained below, the detector outputs may also have offset errors, which, for simplicity of explanation, are not included in the above values.

Step 302 is dithering array 41. The distance of the dither is the height of one detector element. For a horizontally scanned array, such as array 11, the dither is in the vertical direction. For a vertically scanned array, the dither would be in the horizontal direction. WO 01/03423 PCT/US00/17622 14

In either case, in this example, the dither is a distance substantially the same as the dimension of one detector element in the cross direction of the scan. FIGURE 4B illustrates Step 302. The scene is assumed to be substantially stable during the dither, so the scene is illustrated as having the same flux values before and after the dither. As a result of the dither, pairs of neighboring detector elements are exposed to the same scene path during the dither cycle. For example, both Da and DB receive the scene flux from the second row of scene 42.

In a modified version of the method, Step 302 could be performed with fractional dithering, to enhance image resolution. For example, the dither could be % the pixel height. The new pixel values would then be extrapolated to estimate pixel values that would occur with a full dither. These extrapolated pixel values would then be used for calculating correction values. In another modified version, the dither could be for a distance equal to the pixel dimension plus a fraction. It is also possible to dither in distances of two or more pixel dimensions with calculations modified accordingly.

Step 303a is receiving a second field, here Field 2, with array 41 in its new position as illustrated in FIGURE 4B. With the detector element gains, the received pixel values for Field 2 are: 1 2 3 4 4 8 12 16 9 18 27 36

Field 1,

Field 2

Like the values of include offset errors these values do not WO 01/03423 PCT/USOO/17622 15

Step 303b is retrieving two adjacent columns of data from each field. As an alternative to Steps 3 03a and 303b, the process could be performed "on the fly" with pairs of columns of data being processed from Field 2 as the field is being scanned. Also, the columns whose data is being used for calculations could be spaced rather than adj acent.

Step 304 is applying the gain correction values from the previous iteration, if any, to the data for the two columns of the new field (Field 2) and the corresponding data in the previous field (Field 1) . By "corresponding" is meant the pixel values in the same two columns. In the first iteration, there are no previously calculated correction values, but in subsequent iterations, each value will be multiplied by an associated gain correction value. By "associated" is meant that the pixel value is corrected with a correction value of the detector element that sensed that pixel value. After Step 3 04, the data for the current two columns of both fields is now gain-corrected with current gain correction values, although it may still contain offset errors.

Step 305 is calculating a new gain correction for each detector element. In general terms, this calculation involves a computation of a "gain ratio" for each detector element. Then, for each detector element, the gain ratios of that detector element and of its two neighbors are averaged.

In mathematical terms, Step 305 is performed by computing, for each detector element, a gain correction

value from its UN (upper neighbor) gain ratio and its LN (lower neighbor) gain ratio. The general form of the gain correction (GC) computation is: WO 01/03423 PCT/USOO/17622 16 NEW_ GC = (1 + UN gain ratio + LN gain ratio) 3

This calculation is an average of contributing gain ratios. The detector element whose gain correction value is being calculated contributes a term, as do its two near neighbors. The unity term is equivalent to a gain ratio of l/l, so that the detector element's own gain is * used to determine the average. The quotient, 3, is the number of contributing detector elements (the detector element under consideration and its upper and lower neighbors).

Using the data of FIGURES 4A and 4B as a specific example, to begin Step 305, differencing the received pixel values in the first two columns of Field 1 generates the following data: 2

Col 2 - Col 1 = Δ1 = 6 12

Differencing the received pixel values in the first two columns of Field 2 generates the following data: 1

Col 2 - Col 1 = Δ2 = 4 9

Each difference value is "associated" with a detector element, in the sense that the same detector element detected the two pixel values, from which the difference value was calculated.

The difference values from Fields 1 and 2 are used to obtain gain ratios. These ratios compare difference values of neighboring detector elements. For each detector element, an upper neighbor (UN) gain ratio and a WO 01/03423 PCT/USOO/17622 17 lower neighbor (LN) gain ratio are computed. The following table illustrates how these gain ratios are calculated: Δ1 Δ2 Δ1/Δ2 Δ2/Δ1 5 1 - 2 4 2/4 4/2 6 9 6/9 9/6 12 - Each Δ1/Δ2 value is a UN gain ratio -- the difference 10 value associated with a detector element's UN divided by the difference value associated with that detector element. Each Δ2/Δ1 value is a LN gain ratio -- the difference value associated with a detector element's lower neighbor divided by the difference value associated 15 with that detector element.

As an example of calculating these ratios, in Field 2, Db sensed pixel values with a difference value of 4.

In Field 1, DB's upper neighbor, DA, sensed pixel values in the same locations but with a difference value of 2. 20 Thus, the UN gain ratio for DB is 2/4.

As indicated, because DA has no upper neighbor, no UN gain ratio is available for DA. Also, because Dc has no lower neighbor, Dc has no LN gain ratio.

As an example, for detector element DB, the gain 25 correction is calculated as: GC = (1 + 2/4 + 9/6) B 3 = 1 , where 2/4 is DB's UN gain ratio and 9/6 is DB's LN gain ratio. WO 01/03423 PCT/US00/17622 18

If a detector element is an "end" detector element, it has either no UN or no LN. In this case, when calculating the gain correction value, there are only two contributing detector elements. For example, for the 5 "end" detector element, DA, which has no upper neighbor, the gain correction is calculated as: = α_ί±/_2) A 2 = 1.5 , where 4/2 is DA's LN gain ratio. Likewise, for the other end detector element, Dc, which has no lower neighbor, the gain correction is calculated as: GC. c 2 = .833 10 , where 6/9 is Dc's UN gain ratio.

The following table sets out the gain correction values obtained as a result of Step 305, based on the first two columns of Fields 1 and 2. Each detector element, has an associated gain correction value. In the 15 next iteration of Step 3 04, these values will be used to correct the next column of data. GCa =1.5 GCb = 1 GCc = .833

In step 306, the gain correction values are updated.

This is accomplished by multiplying the old gain correction values GC (i.e. those updated from the

previous iteration) and the gain correction values NEW_GC computed in Step 305. 20 WO 01/03423 PCT/USOO/17622 19

Steps 3 03b through 3 06 are repeated for successive pairs of columns of data across the detector array. As stated above, the column pairs can be adjacent to each other or spaced. The next iteration can re-use one column or use two new columns. The iterations are repeated until the calculated gain correction values become substantially constant.

Steps 307 - 316 are the offset correction part of the non-uniformity compensation method. The offset correction is performed in two parts.

It is assumed that data for at least two fields is stored in memory as a result of the gain correction calculations. Continuing the example above, data for Field 1 and Field 2 is available. In Step 307, two corresponding columns are retrieved for offset correction calculations. FIGURES 5A and 5B illustrate the same scene as illustrated in FIGURES 4A and 4B, but with the detector elements having offset errors rather than gain errors.

As in FIGURES 4A and 4B, the scene is assumed to be stable, so the flux values are illustrated as being the same from field-to-field. Each detector element is assumed to have an offset error, with detector elements Da, Db, and Dc having offset values of 1, 2, and 3, respectively. Thus, for the first column of Field 1, the received pixel values are: 3 5 7 , where each value is the sum of the "true" pixel value and the offset error of the detector element that sensed WO 01/03423 PCT/US00/17622 20 it. For the first column of Field 2, the received pixel values are: 2 4 6

In Step 308, the updated gain correction values from Step 306 and the updated offset correction values calculated in the previous iteration, if any, are applied to the pixel values for the new column in both the new field and the previous field. In the first iteration of Step 308, there are no previously computed offset corrections. In the next iteration of Step 308, previously computed offset correction values will be available, and will be used to correct the new column of data in both Field 1 and Field 2.

In Step 309, offset correction values for each detector element are calculated, using the difference of the output of that detector element and of its upper neighbor. These offset correction values are referred to herein as "upper neighbor offset correction values".

This calculation first determines, for each detector element, the difference between the output of a detector element and the output of the detector element' s upper neighbor, at the same location of the scene. Using the received values set out above, and assuming the gain correction from Step 308, the UN difference values are: 4-3 = 1 6-5 = 1 .

As indicated, detector element DA has no upper neighbor and thus has no UN difference value. Next, for each detector element, its difference value is divided by 2 WO 01/03423 PCT/USOO/17622 21 (the number of contributing detector elements).

Expressed in mathematical form, each upper neighbor (UN) offset correction (OC) value is calculated as: UN OC = UN difference 2

The above offset correction value could be equivalently expressed as a detector element's local .average output, subtracted from that detector element's output. For example, for DB: OCB = IB - % (IB + IA) = (IB - Χ) , where I represents a detector element's output.

For the first iteration, the UN offset correction values are: OCA = OCB =% = 0.5 OCc = = 0.5

In Step 310, the UN offset correction values calculated in Step 309 are applied to the output values for the current column in both the new field and the previous field. Thus, for Field 1 (the previous field), the corrected outputs are: 3-0 = 3 5 - 0.5 = 4.5 7 - 0.5 = 6.5 , and for Field 2 (the new field), the corrected outputs are: 2-0 = 2 4 - 0.5 = 3.5 0.5 5.5 WO 01/03423 PCT/USOO/17622 22

In Step 311, an offset correction value for each detector element is calculated from the corrected data of Step 310, using the difference of the output of that detector element and its lower neighbor. These offset correction values are referred to herein as "lower neighbor offset correction values". The upper and lower neighbors are referred to herein as "opposing" neighbors.

The lower neighbor offset correction calculation determines, for each detector element, the difference between the corrected output of a detector element and the corrected output of that detector element1s lower neighbor, at the same scene location. The resulting difference values are: 3-3.5= -0.5 4.5 - 5.5 = -1

As indicated, detector element Dc has no lower neighbor and thus has no LN difference value. Next, for each detector element, its difference value is divided by 2 (the number of contributing detector elements).

Expressed in mathematical form, each lower neighbor (LN) offset correction (OC) value is calculated as: LN OC = LN difference 2

Like the UN offset correction value calculation, the above LN offset correction value calculation can be equivalently expressed as the local average of a detector element and its lower neighbor, subtracted from that detector element's output. For example, for DB: OCB = IB - M (IB + Ic) = M (IB - Ic) WO 01/03423 PCT/US00/17622 23 , where I represents a detector element's output.

Thus, the LN offset, correction values are: OCA = - (0.5)/2 = -0.25 OCB =-½ =-0.5 OCc = -

In Step 312, these offset correction values are applied to the current column of data in both the previous field and the current field.

In Step 313, the offset correction values are updated by adding to the old offset correction (i.e., that of the previous iteration) the UN offset correction value calculated in Step 309 and the LN offset correction value calculated in Step 311.

Steps 307 - 313 are repeated, column by column, until the offset correction values are substantially constant. The successive columns can be adjacent or spaced.

In Step 314, the gain and offset correction values are now used to correct the data in Field n and Field n+1 for display.

Step 315 is correcting for "dead" detector elements, which are those that do not detect a signal or can otherwise be known to be defective. A pixel value sensed by another detector element in an adjacent field can be substituted for the pixel value that should have been detected by the dead detector element. For example, if DA were dead, the pixel value detected by its lower neighbor Db in Field 2 can be substituted for the same pixel in Field 1. WO 01/03423 PCT/USOO/17622 24

The data for the current column has now been both gain corrected and offset corrected and is now ready for display. In Step 316, Fields n and n+1 are displayed.

Non Uniformity Compensation for Two-Dimensional

Arrays

The dithered scan method of non uniformity compensation can be used with two-dimensional arrays, as well as one-dimensional arrays. In the case of a two-dimensional array, most of its detector elements have adjoining (i.e. "near") neighbors on all four sides. The exceptions are the detector elements on the first and last columns and the first and last rows. These "end" detector elements have only three neighbors, and in the case of corner detector elements, they have only two neighbors. Also, each detector element has other detector elements that touch at its corners and that can also be considered neighbors. As explained below, for a two-dimensional array, gain correction calculations use local averages of gain ratios of each detector element and its four adjoining neighbors, whereas offset correction calculations use local averages of outputs of each detector element and of its four adjoining neighbors and of two of its corner neighbors. FIGURES 6A and 6B illustrate the basic steps of the method, as applied to a two-dimensional array. As indicated below, both gain correction and offset correction require data from a number of time-contiguous fields. Gain correction uses data from four fields and offset correction uses data from six fields. However, the dither pattern is the same so that gain correction and offset correction may use the same input data. WO 01/03423 PCT/USOO/17622 25

Steps 601 - 604 are the gain correction part of the method. Step 601 is receiving four dithered fields of pixel values. The dither pattern is box-shaped, so that each detector element views four scene locations during a dither cycle. At the end of the cycle, the array is in its original position. Each dither is assumed to be the height or width of one detector element, but as discussed above in connection with FIGURE 3, fractional dithering could be used to enhance resolution. Also, the dither could be for a distance of more than one pixel dimension or of fractional parts thereof. The dither could also be diagonal, with the selection of neighbors modified accordingly. FIGURE 7 illustrates the dithering motion of Step 601. Four positions of a single detector element, DA, are shown, one position for each of four fields. The four fields are identified as Fields1.4. At each dither position, a new pixel value is acquired representing the flux intensity as sensed by DA. During the four field times, each detector element of the array undergoes a similar dither pattern, and acquires a pixel value from a different location. Thus, after one dither cycle, four fields of pixel data have been acquired. As with the one-dimensional array, the dither causes neighboring detector elements to be exposed to the same scene path.

Step 602 is correcting the pixel values for the four fields with gain correction values calculated in the previous iteration. In the first iteration, there are no previously calculated values, and the initial gain correction value is 1.

Step 603 is calculating a new gain correction for each detector element. In general, this calculation WO 01/03423 PCT/US00/17622 26 first involves a computation of a gain ratio for each detector element. Then, for each detector element, the gain ratios of' that detector element and of its four neighbors are averaged.

FIGURE 8 illustrates DA with its four neighbors, Da, Dc, Dd/ and DE, during each of the four fields of FIGURE 7. As stated above, it is assumed that the flux is stable during the four-field dither cycle. In Fieldj, DA detects the flux value, Φ(χ,γ). In Field2, DA has moved one space to the right and detects the flux value Φ(χ+1, y) . In Fields, DB is now in position to detect the flux value Φ(χ+1χ). Finally, in Field4, DB detects the flux value Φ(x,y).

The following table sets out the responses of FIGURE 8, where G is the gain and 0 is the offset for a detector element:

Field 1 2 3 4

Pa

Ga

Pb

Φ(χ,γ) + OA Φ(χ+1,γ) + OA

Gn

Φίχ+l.y) + OB Φ(χ,γ) + OB

For each detector element, the difference between two adjacent flux values detected by that detector element is calculated. For DA, the difference value is: ΔΑ = (Ga · Φ(χ,γ) + OA) - (Ga · Φ(χ+1,γ) + OA) , and for DB, the difference value is: Δβ = (Gb · Φ(χ+1,γ) + OB) - (Gb ·. Φ(χ,γ) + OB)

The gain ratio between detector elements DA and DB is WO 01/03423 PCT/USOO/17622 27 10

gain ratioA B (GB · Φ (x+l,y) + OB) - (Gg · Φ (x,y) + OB) (GA · Φ (x,y) + OA) - (GA · Φ (x+l,y) + OA)

Similar gain ratios can be calculated for any pair of neighboring detector elements.

The gain correction for DA is derived from the gain ratios of DA and of its four neighbors. In mathematical terms, the gain correction for DA is described by: GGa = II·- 2? gain ratio

X-B ff. D.E

A, X

JL *5

A G=

A G. = _ χ + _S + _ϊ + __ + I Σ

-5 X-A,B,C,D,E

A

Gain,

Gain,

Similar calculations can be made to obtain a gain correction value for each detector element.

Step 604 is updating the gain correction values by multiplying the newly calculated gain correction values and the old gain correction values. If each detector element's output were multiplied by its gain correction, the result would be a corrected output value. For example, after Step 604, DA's pixel value is: 15 WO 01/03423 PCT/USOO/17622 28

Corrected IA = GCA ^Gain^ · Φ + Offset^ = 4 ( Σ GainJ ' φ + gca ' 5 Y-A,B,C,D,£ )

= Residual Gain, Φ + New Offset, A A

Using matrix notation, the gain correction calculations can be described in terms of the entire array. The scene flux collected by the array during a four-field dither cycle can be depicted as [Φ] , [Φ]-’, 5 [Φ]_Ι, and [Φ] 1, where each arrow indicates the scene location relative to the initial location. The responses of the detector elements during the dither cycle are represented by four image matrices: [IFieldl] , [YieicX > [Ipieida! » and [I Field4! · Thus : [Yield J = [Gain] • [Φ] + [Offset] [Yield 2! = [Gain] ’ [Φ]- + [Offset] [IField 3! = [Gain] • [Φ]Χ + [Offset] C^Field 4! = [Gain] • [Φ] 1 + [Offset] The difference values for Fields 1 and 2 are calculated as: [IField l! - [Yield 2! = [Gain] · ([Φ] - [Φ]-> , and the shifted image difference values for Fields 3 and 4 are: ([I Field 4 )- [I Field 3 ])1 = ([Gain] · ( [Φ] 1 - [Φ]_Ι))1 20 , which is equivalent to: [IrieidiH - [Yields!1 = [Gain] 1 · ( [Φ] - [Φ]—)

This operation is the same as that used to obtain δα and Δβ, as described above for detector elements DA and WO 01/03423 PCT/USOO/17622 29

From the above difference and shifted difference equations, the gain ratios between neighboring detector elements can be calculated as: JLipield 4-Li-— -Ilpieid 3LL = -[Gain] 1 [^Fieid il “ [I-Fieid 2] [Gain] , which is the same calculation described above for obtaining gain ratioAiB.

The gain ratios for other pairs of detector elements are : LX.Field -LX-Field 4Σ2. “ [Gain]- [^Field 2] ~ i^Field 3] [Gain] Li.Field 2U—— —LX-Field ill = TGain]I [^Field 3] £ Field 4] [Gain] -LZ-Field 3-Lt—Σ_ —LX-Field 2J-S = [Gainl- [IField 4] ~ [^Field ll = [Gain]

The gain correction may be defined as the averaged sum of the gain ratios:

GC , [Gain]1 [Gain] -1 + -I— + - + [Gain] [Gain] [Gain]1 + [Gain]-[Gain] [Gain] . 1 [Gain] / 0 10 1 c 111 * [Gain] 0 1 / , The convolution operation on the [Gain] matrix (in braces) represents the shifted, summed, and averaged [Gain] matrix. The operation of convolution with the diamond-shaped mask results in a spatially smoothed (two-dimensionally filtered) version of the [Gain] matrix. WO 01/03423 PCT/USOO/17622 30

For end detector elements, including those at the corners, the calculations are modified. In a manner analogous to calculating gain corrections for the end detector elements of a one-dimensional array, only the values of available neighbors are part of the local average.

Steps 601 - 604 are repeated, using successive sets of four fields of data, until the gain correction values become substantially constant.

Referring again to FIGURES 6A and 6B, after the gain correction values are calculated, Steps 605 - 612 are performed for offset correction values. The same dither pattern is used for offset correction as for gain correction. However, for each detector element, two additional neighbors are used, as compared to those used for gain correction. FIGURE 9 illustrates the neighbor detector elements used to estimate the offset correction for DA. As indicated, DA has three lower left neighbors (LLNs), which are DB, Dc, and Dd, and three upper right neighbors (URNs) , which are DE, DF, and DG. Although the following description is in terms of these sets of neighboring detector elements, it should be understood that the method could be modified so as to use upper left neighbors and lower right neighbors. In both cases, the two sets of neighboring detector elements are in opposing corners, and are referred to herein as "opposing" sets of neighbors. Also, the method could be modified so that the two sets of neighboring detector elements are comprised of some number of detector elements other than three. WO 01/03423 PCT/US00/17622 31

Referring to FIGURES 6A and 6B and to FIGURE 9, offset correction is performed in two parts. First, in Step 605, the offset correction for each detector element is calculated with pixel values from the detector element and its LLNs. This data is in Field,, through Fieldn+3. If these offset correction values were applied, the result would be a residual offset error for each detector element. These residual offset errors are further corrected with pixel values from the detector element and its URNs. This data is in Fieldn+2 through Fieldn+5.

Referring again to FIGURE 2 and to FIGURE 9, during each four-field dither cycle, a detector element views the same scene location as three of its neighbors. The following table sets out the scene flux incident on DA and its neighbors during FielcL through Fieldn+5.

Field Scene F lux Locations Da Db Dc D„ De df Do n Φ (x,y) Φ(x-l,y) Φ(χ-1, y+1) Φ(x,y+l) n+1 <t>(x+l,y) Φ(x,y) Φ(χ^+1) Φ(x+1, y+i) n+2 Φ(χ+1, y-l) Φ(x,y-l) φ (x,y) Φ(x+l,y) Φ(x+2, y-l) Φ(x+2, y-2) Φ(χ+1, y-2) n+3 Φ(χ^-1) Φ (x-1, y-l) Φίχ-i.y) φ(χ,γ) Φ(x+l, y-l) Φ(x+l, y-2) ®(x,y-2) n+4 Φ(χ,γ) Φ(x+l, y! Φ (x+l, y-i) Φ(x,y-l) n+5 φ(χ+1^) Φ(x+2, y) Φ(χ+2, y-D Φ(x+l, y-D

During Field„ through Fieldj, DA, DB, Dc, and Dd, successively viewed the same scene flux, <&(x,y) . If the assumption is made that the scene is stable during these WO 01/03423 PCT/USOO/17622 32 .10 15 fields, and if each detector element had the same offset, the four detector elements would have the same responses. The gain is assumed to have already been corrected to a constant, g, for all detector elements. However, because of the offset errors, there are differences between the detector element responses. These differences are the basis of calculating the offset correction.

In Step 606, the data to be used for offset correction calculations is corrected with the previously calculated gain correction values and with any previously calculated offset correction values.

In Step 607, the offset corrections from the LLNs are calculated. These offset corrections are referred to herein as "LLN offset corrections". For DA, the calculation is as follows: OCA - JB(n) - -j (ίΛ(π) + iB(n + 1)+ Ic(n + 2)+ Ip(n + 3)) "4 <Πΐ " 4 7(n + 11 + + 2> * + 3>) , where

TA(n) = gr · 4>(x,y) + OA

IB(n + 1) = g · Φ(χ,γ) + OB

Ic(n + 2) = g · Φ(χ,γ) + Oc

ID(n + 3} = g · Φ(χ, y) + OD

By substitution,

<img img-format="tif" img-content="drawing" file="IL147409AD00021.tif" id="idf0001" />

J (¾ + o, ' oc + 0„)

X=A,B,C,P WO 01/03423 PCT/USOO/17622 33

The offset correction is updated in Step 60 8 by adding it to any previously calculated offset correction. As an example of the corrected data at this point, if DA's response were corrected by subtracting its offset correction, DA's corrected response would be:

Corrected IA = g Φ(χ,γ) + OA - OCA = g · Φ (x,y) + — V C&gt; 4 X=A,B,C,D-

At this point, DA has a residual offset, which is the spatial average of offsets from DA and its LLNs, DB, Dc, Dd. Similar calculations are made for each detector element.

Using matrix notation, the offset correction can be described in terms of the entire array. At the end of the dither cycle, four image fields are received and are shifted as: [IField l] = g · [Φ] + [Offset] flField = g · [Φ] + [Offset]- [IField 3] ** = g · [Φ] + [Offset]-1 [^Field 4] = g · [Φ] + [Offset]1 LLN offset correction is defined as GGLLN PFieldll [^Fieldl] 4 ’ ( ^Fieldl + ^Fieldz]^ + ^Fielda^ + ^Fieldi]^ ) = [offset] - · ([offset] + [offset]- + [offset]-! + [offset]1 ) [offset] = [Offset] - | WO 01/03423 PCT/USOO/17622 34

When subtracting the offset correction term from the image, the residual offset becomes:

Residual 1 1 1 Offset 4 1 1 * [offset]

Mathematically, the smoothing of offset errors is the same as convolving the offset matrix with a rectangular mask. Recursively applying offset corrections to the images is like a low pass filtering process to the offset noise, which will eventually smooth out the high frequency components of the [Offset] matrix.

In Steps 609 - 612, DA is grouped with its URNs, DE, DF, and Dg, to calculate its URN offset correction. Referring to the table above, these detector elements receive the scene flux value 4&gt;(x+l,y-l) in Fields through Fieldn+5. These are the four fields retrieved in Step 609. The output values in these fields are corrected in Step 610 with current gain and offset correction values. In Step 611, an analogous calculation as described above for LLNs can be used for URNs,. where the offset correction for DA is calculated by averaging the offset error of DA with those of DE, DF, and DG. The offset correction values are again updated in Step 612.

Steps 605 - 612 are repeated for every four successive fields, alternating the LLN and URN corrections. The alternating corrections ensure that end detector elements are corrected at least once by at least one neighbor.

For offset errors introduced by dither bias, a weighting factor can be applied to the correction value during the offset correction loops. The weighting factor, a &lt; 1, is applied such that only fractions of the WO 01/03423 PCT/US00/17622 35 calculated offset errors are corrected. The combination of a small weighting factor and the alternating neighbor calculations will cancel some correction errors introduced by dither bias and/or scene motion. A variety of dither patterns, single steps or multi-step loops, with either vertical, horizontal, or diagonal steps, can be chosen for offset correction. The common feature of the various dither patterns is that a number of neighboring detector elements are exposed to the same scene flux during a correction cycle. For the rectangular dither pattern described above, corrections can also be calculated with different combinations of neighboring detector elements. For example, fields 1 and 2 (with a single horizontal dither step) could be used to correct offset errors using left neighbors, followed by right neighbors. Then, fields 2 and 3 (with a single vertical dither step) could be used to correct offset errors using down neighbors, followed by up neighbors, etc. FIGURES 10 and 11 illustrate an alternative dither pattern for offset correction. The pattern is diagonal, and in the example of FIGURE 10 is upper right diagonal/return/upper left diagonal/return. FIGURE 11 illustrates DA and its four diagonal neighbors, DB, Dc, Dd, and De. During a four-field dither cycle, DA views three different scene pixels and its four diagonal neighbors also view one or two of the three scene pixels in the same cycle. Local differences are calculated from values of that detector element and its four neighbors to obtain offset corrections. WO 01/03423 PCT/USOO/17622 36 FIGURES 12 and 13 illustrates two other diagonal dither patterns. These patterns may be used for both gain and offset correction. FIGURE 14 illustrates an alternative rectangular, dither pattern for offset correction. FIGURE 15 illustrates an alternative dither pattern for gain and offset correction.

Recursive Operation

The process of FIGURE 3 (for one dimensional arrays) and FIGURE 6 (for two dimensional arrays) may be performed over many iterations. Gain corrections are recursively applied to each new image, with each iteration reducing the gain differences among detector elements. Likewise, the recursive application of offset corrections to each new image reduces the offset differences among detector elements.

In practice, gain non uniformities are not expected to be as big as offset non uniformities, and a rather smoothed gain matrix should be obtained in a few iterations. Offset correction is anticipated to be applied continuously. Gain errors can be corrected in the beginning of an image sequence, or when there is major change in scene radiance. The residual gain errors will also be smoothed out in the offset correction loops. WO 01/03423 PCT/US00/17622 37

I

Bi-Directional and One-Step Algorithms for Offset

Correction

The above-described concepts incorporate an algorithm

I referred to herein as a "bi-directional" algorithm for 5 offset correction. As applied to a one-dimensional array, where the scan direction is horizontal, the algorithm could be described as: calculate correction using upper neighbor averaging, apply the correction to the detector elements, calculate the correction using 10 lower neighbor averaging, apply the correction to the detector elements. The algorithm is "bi-directional" in the sense that two neighbors are used, with a correction being calculated first in one direction then in the other. The algorithm is also "two-step" in the sense 15 that the correction from one neighbor is applied before the correction from the other neighbor is calculated.

The bi-directional two-step algorithm of the preceding paragraph can also be described for a onedimensional array that is scanned vertically. The dither 20 would then involve left and right neighbors. In general, rows and columns are equivalent, and the dither is in the cross-direction of the scan.

The same algorithm can be used for two-dimensional arrays. As an example, for the rectangular dither 25 pattern of FIGURE 7, a first pass (or iteration) would make corrections with left and right neighbors when the detector array is dithered right (or when Field 1 and Field 2 are dithered in the horizontal direction). A second pass (or iteration) would make corrections with 30 upper and lower neighbors when the detector array is dithered up (or when Field 2 and Field 3 are dithered in the vertical direction). This approach requires less WO 01/03423 PCT/USOO/17622 38 computation (two fields rather than four are averaged) and less memory, as compared to approaches that use four neighbors.

Two-step algorithms can be shortened into one-step algorithms. Referring again to the upper and lower neighbor corrections described above for the onedimensional array, the offset correction value from the lower neighbor is: OCM = (I2 - I2‘) (1) , where I2 and I2 are two fields at times tl and t2, respectively, and 1 represents a shift down. For generality, each set of I values is a "matrix", where the matrix is one-dimensional for a one-dimensional (scanning) array. Applying the OC^ to the two fields results in: 1/ = I, - OC^ (2) 1/ = I2 - OCM (3) , where 1/ and 1/ are the corrected fields. The offset correction from the upper neighbor,, using the corrected fields is: OCW = (1/ - I/’) (4)

The combined correction from both neighbors is: ocbi -directional — GGxjj + OCgjQ (5) . By substitution, a one-step equation can be derived from the combined two-step equation: WO 01/03423 PCT/USOO/17622 39

Substitute (1) int o (2) and (3) (ι,-ιΐχχ, + χ &gt;2 = b-yOi-d)' Il + 1 d

Substitute (6) and (7) int o (4)

<img img-format="tif" img-content="drawing" file="IL147409AD00022.tif" id="idf0002" />

1.,, + 1.,2)_(i.I,+I.,2 •'ΧΎΥ'2)

<img img-format="tif" img-content="drawing" file="IL147409AD00023.tif" id="idf0003" />

The bi - directional offset correctionisobtained by combining (1) and (8) AOC = AOCd + AOCy xOi-’O+H’2' -QOi-ii+h-ii) — •I, + - I2 ---if 2 2 2,

<img img-format="tif" img-content="drawing" file="IL147409AD00024.tif" id="idf0004" />

Thus, each one-step value is the mathematical equivalent of combining the two corrections (UN and LN) described above in connection with FIGURE 3 and Steps 309 - 312. In words, each offset correction value is the sum of two "shifted image" difference, where an "image" is a matrix of detector output values. The matrix is onedimensional for a one-dimensional (scanning) array and two-dimensional for a two-dimensional array. Each shifted image difference is the difference from two images in the same field, one shifted with respect to the other. The 1/4 factor is the result of combining the % factor associated with each of the two field differences. 10 WO 01/03423 PCT/US00/17622 40

For a one-step algorithm with left and right neighbor shifting rather than upper and low shifting, each offset correction value would be calculated as: [AOC] = 1-([I], -[I]f +[I]2 -rat)

Scene Term for Offset Correction with Dither Bias FIGURES 17A - 17D illustrate how two detector elements might be exposed to slightly different scene paths, due to a dither bias. One path is shown with dotted lines, the other with solid lines. Detector element DA follows the path Φχ to Φ2, whereas detector element DB follows the path Φ1' to Φ2' . Thus, "dither bias" occurs when there is an error between the intended and the actual displacement of the dither array.

Dither bias will not cause a major breakdown for the gain ratio calculations as long as the common scene path shared by the neighboring detector elements is parallel to the direction of maximum local scene gradient. This is the case illustrated in FIGURE 17A. When there is not enough scene gradient along the dither path, as shown in FIGURES 17C and 17D, a good gain ratio estimate will be difficult to obtain, with or without dither bias. Thresholds can be set based on the direction and degree of the local gradient, as well as the dither path and dither bias.

Perfect dithering is not always possible and usually a fixed bias can be observed and/or measured. Random dither error does not cause the dither correction to break down because randomly generated noise tends to be WO 01/03423 PCT/USOO/17622 41 smoothed out over time by the repeated smoothing in the offset correction process.

Two alternative methods of compensating for dither bias have been described in U.S. Patent Application 5 Serial No. 08/918,373, entitled "Dither Correction for

Infrared Detector Arrays". Consistent with the above-described algorithms, offset corrections are calculated using local neighbor averaging. For bias compensation, one method uses an alternating dither pattern. Another 10 uses a weighting factor. A third method is discussed in this section, and uses a "scene term" that is added to the offset correction value. The following description is in terms of the one-step bi-directional algorithm described in the previous 15 section. For purposes of this section, each newly calculated offset correction value is referred to as a "delta" offset correction value, notated as aOC.

Using the matrix notation described above for offset correction, [I]x and [I]2 represent two consecutive images 20 received by the detector array when the array dithers to . the right from tx to t2. When dither bias is present, the two images are described as follows: [I]i = g - [&lt;X&gt; J -t- [Offset J [I]2 = g - [Φ ]2 + [Offset] = β·([Φ]~* + δ -[VxO]-^) - [Offset] , where the gradient image is defined as: 25 [ν,Φ]=[ΦΓ-[Φ] , and where δ represents the amount of dither bias. If .

5=0, there is no bias (one full pixel dither step). A positive value of δ corresponds to a dither step greater WO 01/03423 PCT/US00/17622 42 than one pixel; a negative value corresponds to a dither step less than one pixel.

As explained above, dither bias can be the result of unintended optical or mechanical error. Intentional 5 dithering of less than or greater than a pixel can be treated the same as unintended dither bias. Many-scanning systems use half-pixel dithering, such that δ = -0.5.

In the case of one-step bi-directional dithering, 10 incremental (delta) offset correction values are calculated as: [Aocj=i-([i],-mr +m2-rar) . By substitution, the offset corrected [I]2 becomes [I]2 - [ΔΟΟ] = g · [Φ ]2 -1 (g · [Φ] - g · [Φ]" + g [Φ]2 - g · [Φ]? ) +[Offset] -1 · {[Offset] - [Offset]-* + [Offset] - [Offset]*”} = g-[0&gt;]2-~g-5-([Vx&lt;Dp -[νχφ]) + |·[1 2 1|* [Offset] = g-[0&gt;]2 + --[12 l]*[Offset]-i-g-0-[V*&lt;X&gt;] 4 4 15 . In the above equation, the last term is an error term, which is introduced after the offset correction is applied. The error term, given as the second derivative of the scene, corresponds to sharp edges in the scene.

For a stationary scene with a fixed dither bias, δ * 0, 20 the repeated delta offset correction with the same error WO 01/03423 PCT/USOO/17622 43 10 term added at each correction update, will gradually erase the scene as a result repeated degrading of sharp edges .

To calculate the first and second derivatives of the scene, the delta image is calculated to give a rough estimate of the scene gradient as follows: ΓΙ]2 — [I]i 1

For δ &lt; O.

For δ s 1 .-+- δ : - (ΐ -+- δ ) - [v x&lt;r&gt; 3 Γΐ]2 - ci]i 1 1 -+- δ [132 — [I3i = S - &lt;1 -+- δ) - 3 A "scene term" is defined as follows for δ &lt; 0: [aocu - —· {([fr- [/]&gt;)- ([fr- [/)07 =~·&amp;amp;·(!+#)·Μ-Γ0+4[ν*ΦΓ} 4 *

For δ &gt; 0, the "scene term" is: ax=;Y-{([/]2- mi)- ([/p- mir} =?ϊ0'{?'ί1+&lt;ί)'[νχφ]'^(1+ί)'[νχΦΓ} = -pg-RV&gt;] WO 01/03423 PCT/USOO/17622 44

As indicated, the "scene term" is based on field difference values. Expressed in words, two field differences are obtained: one from taking the difference between the first image and the second image, and one from taking the difference between the shifted first image and the shifted second image. The field difference and the shifted field difference are then differenced.

The result, divided by 1 + δ, represents the second derivative of the scene. The second derivative values are multiplied times -1/4 and times δ (the dither bias value) .

As discussed above, the algorithm-generated error is the second derivative of the scene. The scene term is based on a calculation of the second derivative so that it can be removed from the correction term.

To obtain a bias-compensated offset correction delta value, the scene term is added to the offset correction delta value: [aOC] compensated = [^OC] + [aOC] scene

When applying the compensated offset correction delta values to the image, the scene error resulting from dither bias disappears, as shown in the following equation:

Ph - [DOC]sc = (pfe - [AOC]) - [AOC^e = (§·[Φ]20[1 2 l]»[0ffset]-|-g6 [v^]) = g [0&gt;]2 + y.[l 2 lJHOffset] 4 WO 01/03423 PCT/US00/17622 45

As an example, using the bi-directional one-step algorithm with half-pixel dithering, δ = -0.5. The bias ratio is -1, and the scene term is: [iOC]SOTt = ί {([I]2 - [I],) - ([I], - [I],)-} 5 The bias-compensated offset correction value is obtained by adding in the scene term to the above-described offset correction value, such that: [aOC] compensated = i^OC] + (-1) OOC] scene 10

Other Embodiments

Although the invention has been described with reference to specific embodiments, this description is not meant to be construed in a limiting sense. Various 15 modifications of the disclosed embodiments, as well as alternative embodiments, will be apparent to persons skilled in the art. It is, therefore, contemplated that the appended claims will cover all modifications that fall within the true scope of the invention. 20