Method of automatically detecting evaluation zones in images of mechanical parts

The invention relates to a method for automatic detection of areas that can be assessed in images of mechanical parts. It applies in particular to the inspection of mechanical parts by fluoroscopy.

The inspection by radioscopy of a part is generally carried out by means of several views making it possible to inspect different zones of the part. The images obtained for the different views often have several areas. Certain zones are said to be non-appraisable when they are saturated in very light or very dark gray levels, or when the contrast is not sufficient there to allow the detection of faults, or when they do not represent the part; the other areas are said to be assessable and are used to search for possible faults.

The known methods, for example from document EP-A-0 627 693, for determining the areas that can be appraised in a mechanical part image generally consist in marking the different areas of the image and then in determining the exact contour of the areas using a method known as a watershed, LPE for short. The problem with these methods is that they are not fully automatic and require an operator to mark areas of the image. Determining markers is a delicate operation and must be carried out by an operator qualified in image analysis. Manual methods are particularly long and tedious in the frequent cases where they are applied to a control of mechanical parts in series and where the same treatment must be applied to a series of images showing the same scene with objects which can be located at different places and / or have variable shapes. Furthermore, the absence of a systematic method for marking areas can lead to erroneous interpretations of the images and call into question the reliability of the parts control.

A first object of the invention is to provide a method for automatically detecting the different areas in an image of a mechanical part. Another object of the invention is to provide a method for automatically determining a set of markers associated with different areas of the image, these markers having optimized dimensions.

For this, the invention consists in a first phase called manual segmentation phase to establish a description of the images to be segmented using one or more reference images representative of the images to be segmented. During this first phase, the detection of the zones in the reference images is carried out by an operator by means of manual segmentation.

In a second phase, called the phase of determining and optimizing the marking parameters, the invention consists in defining simplification transformations of the reference images making it possible to accentuate a contrast or a difference between two or more zones and to determine and set image thresholding intervals to obtain markers for the areas detected in the first phase. Each marker is a subset of a single image area and has dimensions optimized to speed up the image segmentation process and make it more stable against noise.

Finally, in a third phase, the invention consists in automatically marking the zones of new images by applying to these images the simplification transformations defined in the second phase and by using the parameters fixed during the second phase. The exact contour of the zones is then determined using the method known as the watershed.

According to the invention, the method of automatic detection of the zones which can be assessed in a mechanical part image, consisting in marking the zones of the image and in determining the exact contour of the zones using a method called the watershed line, is characterized. in that to carry out the marking of the zones, it consists:<ul id="ul0001" list-style="dash"><li>in a first phase (1), performing manual segmentation of at least one reference image to detect the areas of this reference image;</li><li>in a second phase (2), to define and apply at least one simplification transformation to the reference image, this transformation possibly being the identity transformation, to determine and optimize thresholding intervals for the zones of the reference image allowing to obtain markers characteristic of each zone, to fix the optimized thresholding intervals;</li><li>in a third phase (3), performing the automatic marking of the zones of a new image by successively applying to the new image the simplification transformation and the thresholding intervals fixed during the second phase.</li></ul>

Other features and advantages of the invention will appear clearly in the following description given by way of nonlimiting example and made with reference to the appended figures which represent:<ul id="ul0002" list-style="dash"><li>FIG. 1a, an image of a mechanical part comprising three zones;</li><li>Figure 1b, an example of a set of markers chosen to detect the different areas in the image of Figure 1a;</li><li>FIG. 1c, a transformed image of the mechanical part after application of a gradient transformation;</li><li>FIG. 1d, the result of the segmentation of the image of FIG. 1a into three zones, obtained after application of the LPE method, according to the invention;</li><li>FIG. 2, a block diagram of the different phases of the process for detecting areas in an image of a mechanical part, according to the invention;</li><li>Figure 3a, a block diagram of the steps of the marking process during a first phase called manual segmentation phase, according to the invention;</li><li>FIG. 3b, a block diagram of the steps of the marking process during a second phase, called the phase of determining and optimizing the marking parameters, according to the invention;</li><li>FIG. 4a, an image comprising 2 zones A and B;</li><li>FIG. 4b, the histograms hA and hB of the two zones A and B of the image of FIG. 3a;</li><li>FIG. 4c, an image of two markers obtained after thresholding of the image shown in FIG. 3a, according to the invention;</li><li>FIG. 5, a block diagram of the steps for determining and optimizing the thresholds for thresholding an image comprising two zones A and B, according to the invention.</li></ul>

The segmentation of an image consists in sharing an image in several zones and makes it possible to recognize objects or regions of homogeneous appearance.

Figure 1a shows a fluoroscopic image of a solid blade. This image has three zones, the zone in the middle of the image being dawn, the zone around dawn being due to the circular field of an image intensifier, the external zone in black corresponding to the mask of the camera of Shooting. To segment this image and recognize these three zones using an LPE method, it is necessary to define a set of markers characteristic of the zones of the image and a transformed image of the original image in which the outlines of the different zones are evidence.

FIG. 1b shows an example of choice of a set of markers, each of which is characteristic of a single area of ​​the image. The three areas are marked by squares in different gray levels.

FIG. 1c presents an image in which the contours of the different zones appear very clearly in white. This image was obtained by applying a transformation called "morphological gradient" to the original image. This morphological gradient transformation consists in performing successively on the original image, a morphological dilation, a morphological erosion and a subtraction between the dilated and eroded images.

All the markers and the transformed image being defined, the segmentation of the original image can then be carried out using the LPE method.

The LPE method consists in extending the markers by following the relief of the transformed image (by considering the image as a topographic surface) and in determining the exact contours of the zones in the image.

FIG. 1d shows the result of the segmentation of the original image represented in FIG. 1a after application of the LPE method using the markers and the transformed image represented respectively in FIGS. 1b and 1c.

FIG. 2 represents a block diagram of the three phases of the method of detecting the zones in images of mechanical part, according to the invention.

The first two phases 1 and 2 are learning phases. In phase 1, a manual segmentation of one or more reference images is carried out in order to obtain a description of the different areas of these reference images. In phase 2, parameters for marking the areas of the reference images are determined, optimized and fixed, these marking parameters being the most effective simplification transformations and image thresholding intervals allowing optimal quality of separation to be obtained areas of reference images. These phases require the presence of an operator, in particular to perform a segmentation of the reference images and to define the simplification transformations. The search for optimum thresholding intervals is carried out automatically.

The third phase 3, is a fully automatic phase consisting in using the zone marking parameters defined and fixed during the second phase to mark the zones of new images. This third phase allows automatic control of mechanical parts in series without requiring the presence of an operator to mark the areas of the images. The exact contour of the zones is obtained by then using the method known as the LPE watershed.

Figures 3a and 3b show two block diagrams corresponding respectively to the first two phases of the marking process according to the invention. FIG. 3a relates to a first phase, known as the manual segmentation phase during which reference images stored in a learning base are segmented into different zones by a manual segmentation method. This first phase includes a first step 10 which consists in choosing a reference image in the learning base, a step 11 during which an operator indicates the number of zones in the chosen image, a step 12 during which the operator marks manually the zones on the image by means of a pointer, a step 13 during which the exact contours of the zones are determined using the LPE watershed method. The manual segmentation of the reference image is then terminated and a test is performed in a step 14 to determine whether all the images available in the learning base have been segmented. If the test is negative, another reference image is extracted from the learning base and steps 10 to 14 are repeated with this new reference image. If the test is positive the first phase is completed.

FIG. 3b relates to the second phase of the method of marking the areas in an image, called the phase of determining and optimizing the marking parameters, during which the areas of the reference images determined during the first phase are analyzed so as to extract markers with optimized dimensions and zero intersection between them.

As the position of the zones in the reference images is known by the manual segmentation carried out during the first phase, it is possible to carry out measurements and statistics in each of these zones. In particular, it is possible to calculate histograms which provide information on the distribution of the gray levels in each zone, or to calculate particle sizes which make it possible to know the shape of each zone.

The invention consists in using these measurements and these statistics to determine and optimize markers. For this, in a step 20, the images segmented during the first phase are used to develop a base of simplification transformations of these images.

The simplification transformations of the image are chosen by the operator so as to accentuate a contrast or a difference between two or more zones of the image.

The difference between two or more zones can for example be a difference in shape, position, texture, etc., which is reflected in the transformed image by a difference in gray levels and which allows the zones concerned to be marked by carrying out thresholds.

There are many image transformation operations such as, for example, identity transformation, erosion, dilation, Laplacian, top hat, contrast correction, high pass filter, pass filter bottom, etc. These transformation operations can also be combined with one another. Among all the possible operations, the operator chooses those which advantageously highlight certain characteristics of the image and which thus make it possible to bring out the information sought.

In general it is necessary to choose several transformations of simplification of the image, because when a transformation makes it possible to increase differences or a contrast in the image, it also introduces noise and parasites.

The noise appearing randomly from one transformation to another, it is then possible to eliminate it by retaining only the information which appears systematically in all the transformed images.

In the case where a single transformation makes it possible to distinguish all the zones of the image, the operator can limit the base of the simplification transformations to this single transformation.

When the basis for the simplification transformations is developed, the manual operations of the process are complete, the determination and optimization of the markers being carried out fully automatically by following the steps described below.

In a step 21, a simplification transformation is extracted from the base of the transformations and applied in a step 22, to one of the reference images.

In a step 23, the content of the zones of the transformed image is analyzed for example by calculating the histograms of these zones, and the intersections between two zones are sought.

In a step 24, the results of the analyzes of the content of the zones are used to determine for each zone a thresholding interval defined by two respectively minimum and maximum thresholding limits making it possible to separate the zones. The threshold limits are defined so as to optimize the quality of the separation between the zones. The detailed steps relating to the determination and optimization of the threshold limits are described in relation to FIG. 4. In a step 25, a test is carried out to determine whether all the simplification transformations have been applied to the chosen reference image. If the test is negative, steps 21 to 24 are again implemented with another simplification transformation. If the test is positive, in a step 26, for each zone, the intersection of the thresholded images obtained for each simplification transformation is calculated, and in a step 27 the results of the intersections relating to the different zones define the optimum markers for these zones for the chosen reference image. All the marking operations defined in steps 20 to 27 are carried out for each of the reference images belonging to the learning base and the final markers retained are obtained by calculating for each marker the intersection of the corresponding corresponding thresholding intervals determined for each of the reference images.

FIGS. 4a, 4b, 4c illustrate an example of determination of the limits of the thresholding intervals of two zones, according to the invention.

FIG. 4a represents an image comprising two zones A and B; FIG. 4b represents the histograms hA and hB of the zones A and B; FIG. 4c represents an image of the two markers obtained after thresholding of the image represented in FIG. 4a.

The two histograms corresponding to the two areas A and B of the image show that there is a non-zero intersection between the two areas. The common part of the two histograms is located between gray levels denoted g and d, g and d being between 0 and 255 and such that d is greater than g.

In this common part, there are image points called pixels, of the zone A which are brighter than certain points of the zone B whereas overall the zone A is darker than the zone B.

These two histograms also show that the image points which have gray levels between 0 and g belong only to area A and that the image points which have gray levels between d and 255 belong only to area B. In this example, the thresholding intervals making it possible to define markers which are characteristic of a single zone and which are as large as possible are therefore the intervals [O, g] for zone A and [d, 255] for zone B.

To define the markers of each zone, the invention consists in searching for the threshold limits g and d which make it possible to take into account a maximum number of gray levels in each zone and which make it possible to obtain a zero intersection between the levels of gray of the two zones.

FIG. 5 represents a synoptic diagram of the various stages of determination and optimization of the thresholds of thresholding of an image comprising two zones A and B. In this figure, only the stages concerning the marking of zone A are represented, the method being identical for the marking of zone B.

The search for thresholding limits is carried out by successive approaches from the histograms of the two areas A and B and after an initialization step 40, by calculating, in a step 41, for each pixel of the area A and of the area B , the probability that this pixel has a gray level between two threshold limits x and y, y being greater than x. x representing zone A or zone B, this probability PX (x, y) is equal to the sum of the histograms of the pixels n belonging to zone X and having a gray level comprised between the threshold limits x and y, reported the sum of the histograms of all the pixels n whatever their gray level between 0 and 255.<maths id="math0001" num=""><img file="EP0769760B1_D0001.tif" /></maths>

To obtain a large marker characteristic only of zone A, the invention consists in searching for the values ​​of the threshold limits x, y for which the probability PA (x, y) is maximum and PB (x, y) is minimum.

The values ​​of x, y are determined by successive approaches by considering in a step 42, the difference QA (x, y) between the two probabilities PA (x, y) and PB (x, y), this difference being a measure the quality of the separation of zones A and B corresponding to the thresholding interval [x, y] considered.

In a step 43, a test is carried out to determine whether there are thresholding intervals which have not been considered. If the test is positive the values ​​of (x, y) are incremented in a step 44, and steps 41 to 43 are again implemented. If the test is negative in a step 45, the maximum value QA of the difference QA (x, y) between the two probabilities is sought.

In a step 46, the optimal values ​​of the thresholding limits of the zone A are defined. These optimal values ​​are those which make it possible to obtain the maximum value QA of the quality of the marking of zone A.

The expression for the maximum value QA is as follows:<maths id="math0002" num=""><math display="block"><mrow><mtext>QA = sup [PA (x, y) - k PB (x, y)]</mtext></mrow></math><img file="EP0769760B1_D0002.tif" /></maths> k being a weighting coefficient which makes it possible to increase the importance of the probability of zone B compared to zone A and to increase the security of the separation of zones A and B.

In the example shown in FIGS. 4a, 4b, 4c, x and y are equal to 0 and g respectively for zone A and to d and 255 for zone B.

The value of g which makes it possible to obtain a large marker characteristic only of zone A is determined by successive approaches by considering the difference between the two probabilities PA (0, g) and PB (O, g) associated with the quality of the marker ZA of zone A corresponding to the thresholding interval [0, g] considered, and by searching for the maximum QA of this difference as a function of the value of g.

The expression for the maximum value QA of the quality of the marker ZA is as follows:<maths id="math0003" num=""><math display="block"><mrow><mtext>QA = sup [PA (0, g) - k PB (0, g)]</mtext></mrow></math><img file="EP0769760B1_D0003.tif" /></maths>

In the example of FIGS. 4a, 4b, 4c, the value of k has been chosen equal to 50.

In the same way, the value of the bound d, d being greater than g, is obtained by searching for the maximum, as a function of d, of the quality QB of the marker ZB associated with the area B, the expression of QB being the next :<maths id="math0004" num=""><math display="block"><mrow><mtext>QB = sup [PB (d, 255) - k PA (d, 255)]</mtext></mrow></math><img file="EP0769760B1_D0004.tif" /></maths>

The example described with reference to FIGS. 4a, 4b, 4c and to FIG. 5 relates to the marking of two zones. The marking method according to the invention is generalized to a number of zones greater than two by treating the zones two by two in a similar manner.

Markers are defined for all the possible combinations of pairs of zones by determining, for each pair of zones, the thresholds for thresholding the zones by optimizing a quality parameter of the markers of these zones.

In general, for a given image, the threshold limits of the different zones are determined on different transformed images obtained after using several simplification transformations of the image considered. The final marker Zi of an area i is then determined by intersection of all the thresholded images corresponding to the area i.

In the case where the same transformation makes it possible to distinguish several zones of the image, several thresholds can be carried out in the transformed image obtained, each thresholding relating to a pair of determined zones. In this case, for this transformed image considered, the threshold limits of a given zone i are obtained by considering all the threshold intervals in which the zone i intervenes and by calculating the maximum and the minimum of the corresponding threshold limits.

When all the markers of the different areas of the reference images belonging to the learning base have been defined and optimized, the values ​​of the threshold limits are frozen and the most effective simplification transformations, for which the quality parameter of the markers is the most important, are selected. The learning phases are then completed and the third phase of the method then consists in using the selected simplification transformations and the values ​​of the threshold limits set during the second phase to mark zones completely automatically in new images of mechanical parts. not belonging to the learning base. The automatic marking of the zones in a new image is carried out by successively applying to the new image the various selected simplification transformations and the various thresholds the limits of which were fixed during the second phase. The markers for each zone are then obtained by performing the intersection of the thresholded images corresponding to the same zone.

After the marking operation, the new images are segmented using the LPE watershed method.