FR2864668A1 - Leather fault detection procedure uses cartography system with digital image acquisition apparatus to establish likelihood of faults - Google Patents

Abstract

The procedure uses a cartography system (2) with a digital image acquisition apparatus (16) to produce an image of a leather skin (13) being tested, which is processed and has a statistical algorithm applied to it. This makes it possible to calculate, within an analysis window centred on a given pixel of the image and using a mask with at least two regions (R1, R2), the relative likelihood of the presence of a fault in the first region relative to the second, and thus obtain an image of probability indicating the likelihood of faults for each pixel. The cartography system used for the procedure comprises a conveyor belt (7, 8) that carries the skins to an acquisition point (4) with a digital camera (15) and lighting with reflective (17) and transmission (19) sources.

Description

The present invention relates to a process for qualifying a leather

  using a mapping of leather defects by image processing, as well as a device for implementing this method.

  In the remainder of the text, the term leather is used to denote a skin or a leather whatever its stage of manufacture and its mode of presentation: whole, segmented or cut in thickness.

  Leather, the raw material of tannery and tannery, is heterogeneous in its structure and surface appearance.

  The leather has a natural grain due to the implantation of the hairs of the animal, the homogeneity of the grain depending on their size and distribution. Depending on the species, the leather has different characteristics of size, flexibility, fineness of the grain.

  To these natural variations of the leather grain are added deteriorations or defects that affect the quality of the finished leather and therefore strongly influence the end use and the market value of the leather.

  From the stage of the live animal to that of the tannerie-tannerie, defects can be caused during the stages: -levage, due to sanitary condition, parasitism, mechanical lesions or genetic origin; - slaughter, due to the transport and restraint of animals, to the skin of leather; preservation, due to the implementation of the method, and the storage conditions; tanning and tanning, due to mechanical operations on the machine or the effects of chemicals.

  The diversity of the nature of the defects implies a constant adaptation of the process of transformation of the leathers, certain defects modifying the appearance of the finished leather, the others altering the structural properties even of the leather.

  The position of the defect on the leather and the possibility of integrating it into the mounting margins at the manufacturing level of the finished product must be taken into account to evaluate the value of the leather.

  The detection of defects is therefore of decisive interest, both for the tanner, for the quality control, the optimization of its process according to each skin, and the establishment of historical defects, that for the manufacturer of finished leather products, for quantitative and qualitative sorting, maximizing the use of the surface for cutting leather.

  In the remainder of the text, the term wet-blue designates a leather after chrome tanning, giving it a blue color. The leather at this stage is wet.

  Currently, the controls are performed visually and manually, according to criteria established by the tanner and his client manufacturer.

  In the tanner, a first sorting is done after spinning at the wet-blue stage. The check is initially done on only one part of the leather, then a sorting station deepens the analysis of the leather on all its surface, the sorters allotting the leathers according to a dozen choices.

  The sorting remains dependent on the appreciation, the know-how and the regularity of each sorter during the same day. There is no automatic help for sorting wet-blue leathers.

  The error rate in the sorting of semi-finished leathers varies between 2 and 4% depending on the tanners. Some families of faults are poorly detected in the wet-blue and are gradually revealed in the dyeing and finishing operations, thus making any anticipation impossible.

  An automatic and early detection from the wet-blue would allow the tanner to reorientate the leathers as soon as possible.

  At the manufacturer, the detection of defects occurs in three stages: - sorting reception of finished leather; - cutting leather; - the quality control of cut pieces.

  When sorting in reception, the leather coming from the tanner is checked on their size and thickness measurements and on the presence of defects, in comparison with the level of quality required. The leathers are then affected in the transformation process according to their quality.

This step is manual.

  Leather cutting is a step that requires experience and optimization of the placement of the pieces on the leather. The permanent compromise to be made by the cutter begins to be computerized with automatic placement software, coupled with digital cutting tables.

  However, even on the most automated systems, the phase of entry of defects and areas of quality remains manual, which penalizes the profitability and productivity of the whole.

  Attempts to automate the detection of defects in leather have been made, however unsuccessfully, because these approaches did not take into account either an image quality sufficient to highlight the defects, or the highly noisy appearance of obtained images, the lack of reproducible geometric structures on the leathers. In addition, these solutions do not reveal structural defects.

  The object of the present invention is to provide a method and an automatic defect mapping device on a leather to solve these technical problems by processing an image of sufficient quality to highlight the defects and taking into account the noisy appearance of the images obtained and the lack of reproducible structure, and furthermore to detect both appearance defects and structural defects.

  For this purpose, the subject of the present invention is a process for qualifying a leather using a leather defect mapping, implementing a mapping device, comprising means for acquiring a digital image of the leather, and a processing the image obtained comprising at least one phase consisting in applying to the image of the leather, after any preceding steps, a generalized likelihood ratio statistical algorithm or GLRT, making it possible to calculate, in a centered analysis window on a given pixel of the image and comprising a mask with at least two regions, the likelihood ratio of the presence of a defect in a first region with respect to a second region and thus obtaining a probability image indicating for each pixel the value of this likelihood ratio.

  Advantageously, the method comprises a phase of detection of linear type defects, and a phase of detection of circular type defects.

  According to one embodiment, the phase of detection of linear defects uses a generalized likelihood ratio statistical algorithm or GLRT in its linear form, that is to say using a mask form for the analysis window comprising a band-shaped central region passing through the center of the analysis window and oriented at an angle with respect to the abscissa axis, and two lateral regions and located on either side of the central region, the two regions laterally and completing the analysis window, by performing several likelihood ratio calculations according to several values of the angle and preserving for each pixel: the largest value of the likelihood ratio in a probability image and the value of the angle giving the largest value of the likelihood ratio in a plane of orientation.

  Advantageously, the circular defect detection phase uses a generalized likelihood ratio statistical algorithm or GLRT in its circular form, that is to say by using a mask form for the analysis window comprising a concentric central region with the analysis window, of rectangular shape and a region completing the analysis window, by carrying out several calculations of the likelihood ratio by deforming the central region to increase its size, in the direction of the abscissa and the ordinate, keeping for each pixel : - the largest value of the likelihood ratio in a probability image, and - the abscissa and ordinate dimensions of the central region having given the largest value of the likelihood ratio stored respectively in two sizing images on the abscissa and on the ordinate.

  According to one embodiment, the method comprises a phase for locating circular defects, comprising the steps of: - looking for local maximums of the likelihood ratio in the clusters or sets of image points presenting a likelihood ratio higher than a threshold value; define a neighborhood around this local maximum whose dimensions on the abscissa and on the ordinate are determined using values coming from a detection phase and search for a second local maximum of the likelihood ratio in the residual part of the cluster, if part of the cluster is outside the neighborhood thus defined, - a neighborhood is also defined around this second cluster 35 whose dimensions in abscissa and ordinate are determined using values from a detection phase.

  Advantageously, the values coming from a detection phase making it possible to determine the dimensions on the abscissa and on the ordinate of a neighborhood around this local maximum are the values corresponding to the position of the maximum in the sizing images.

  According to one embodiment, the method comprises a phase of localization of linear defects, comprising the steps of: seeking local maxima of the likelihood ratio in the clusters or sets of image points having a likelihood ratio greater than one threshold value, - defining a neighborhood around this local maximum whose dimensions correspond to those of the region of the phase, this neighborhood being oriented by using the value corresponding to the position of the maximum in the orientation plane and, - a second local maximum of the likelihood ratio in the residual part of the cluster, if a part of the cluster is outside the neighborhood thus defined, a neighborhood also being defined around this second cluster whose dimensions correspond to those of the region of the phase, this neighborhood being oriented using the value corresponding to the position of the maximum in the plane of d 'orientation.

  Advantageously, the phase of localization of the linear defects also comprises a step consisting of fusing nearby neighborhoods, and isolating the neighborhoods with a surface area greater than a given threshold as an altered texture region.

  According to one embodiment, the method comprises a characterization phase comprising a step of performing a determination of the precise shape of the contour of the resolved defects using a polygonal statistical active contour algorithm or CASP.

  Advantageously, the characterization phase comprises a step consisting in calculating geometrical and photometric parameters characteristic of each defect.

  According to one embodiment, the method comprises an identification phase consisting of classifying the defects into categories.

  Advantageously, the method comprises a selection phase of defects to be taken into account by a user.

  According to one embodiment, the method comprises a leather classification phase, allowing a quality level to be assigned, depending on the position, the type and the surface of the defects on the leather.

  Advantageously, the characteristic parameters used in the calculation of the likelihood ratio are the mean and the standard deviation.

  According to one embodiment, the model chosen to represent the densities of probabilities of the intensity of the pixels is the Gaussian model.

  According to one embodiment, the method further comprises a phase of performing a geometric analysis of the leather to identify the quality zones and lines bearing elasticity of the leather.

  Advantageously, the method further comprises a phase consisting in determining a segmentation of the leather in regions of homogeneous hues by using Division and Fusion algorithm using a Tetraarbre.

  According to one embodiment, the method comprises several passes in the circular defect detection phase, corresponding to different analysis window sizes.

  The invention also relates to a mapping device for implementing the method described above, comprising means for conveying leathers, for routing them to an acquisition station comprising a digital camera and lighting means, characterized in that the lighting means comprise a transmission light source through the portion of the conveying means constituting the support of the leather and a source of reflection lighting.

  Advantageously, the angle of the direction of the reflection light source with respect to the plane of the leather is between 15 and 35.

  According to one possibility, the device comprises calculation means allowing a software image processing.

  According to another possibility, the device comprises dedicated electronic calculation means for performing image processing.

  The invention will be better understood from the description which follows, with reference to the appended schematic drawing showing an embodiment of the device according to the invention and illustrating the method according to the invention.

  Figure 1 is a schematic overall side view of the device according to the invention.

  Figure 2 is a representation of the flow chart of the image processing performed.

  Figure 3 shows an image of the support and the outline of a leather.

  FIG. 4 represents an image identifying the lines carrying the lending meanings.

  FIG. 5 represents an image identifying the quality zones of the leather.

  Figure 6 schematically shows the principle of division into a plot.

  Figure 7 shows a Tetra-tree data structure.

  Figure 8 illustrates the algorithm of the maximum of the generalized likelihood ratio.

  Figures 9 to 12 show analysis window forms for an algorithm of the maximum of the linear generalized likelihood ratio.

  Figure 13 shows the evolution of the analysis window form for an algorithm of the maximum of the circular generalized likelihood ratio.

  Figure 14 shows neighborhoods used in determining the location of circular resolved defects.

  Figures 15 to 17 show the successive steps of the polygonal statistical active contour algorithm.

  Figure 18 is a portion of an image of a leather.

  FIG. 19 is a portion of a probability image, in the context of a linear GLRT, corresponding to the processing of the image portion of FIG. 18.

  Fig. 20 is a portion of a probability image, in the context of a circular GLRT, corresponding to the processing of the image portion of Fig. 18 with a large defect mask.

  Fig. 21 is an image portion showing polygonal neighborhoods V for large circular resolved defects.

  Fig. 22 is an image portion showing large resolved fault contours obtained by a CASP algorithm.

  FIG. 25 is a bit image portion of the conserved defects corresponding to the image portion of FIG. 18.

  The fault mapping device 2, shown in FIGS. 1 and 2, comprises three main elements: the conveying module 3, the image acquisition station 4 and calculation means 5 enabling the execution of the software of image processing 6.

  The conveying module 3 comprises a translucent mat 7 driven on rollers 8, the path of this mat 7 forming a first plane 9 inclined at 45 relative to the vertical, then a horizontal portion 12 separated from the first inclined plane 9 by a roller 8 located at the upper edge of the plane 9, this horizontal portion 12 corresponding to the portion of the belt 7 located at the acquisition station 4, then a second plane inclined at 45 relative to the vertical, symmetrical to the first plane 9 with respect to the acquisition station 4, separated from the horizontal portion 12 by a roller 8 located at the upper edge of the plane 10.

  The belt 7 forms a training support in a direction of travel P for leathers 13 which can be deposited on the first inclined plane 9 and then driven to the acquisition station 4 and then to the plane 10 where they can be unloaded from the carpet 7. The inclination of the planes 9 and 10 facilitates the loading and unloading leathers 13, for viewing in a small space.

  The carpet 7 is silicone, to prevent slippage of the leather 13 on the inclined planes 9 and 10.

  The acquisition station 4 comprises an articulated strip 14 situated above the horizontal portion 12 of the belt 7 making it possible to press the leather 13 and to lay it flat before the acquisition of an image 16 by a digital camera 15 located at above the horizontal portion 12 of the belt 7 downstream of the strip 14 in the direction of travel P and directed vertically downwards.

  A first source of reflection lighting 17 of the leather 13 is located above the horizontal portion 12 of the belt 7, downstream of the camera 15, emitting a diffuse light, directed at an angle θ of 25 with respect to the plane of the leather 13. This particular orientation makes it possible to highlight both the structural defects and the appearance defects of the leather 13.

  The source 17 is covered by a cover 18 so as to avoid direct illumination of the camera 15 by this source 17 and a disturbance of the image 16 obtained.

  A second source of transmission lighting 19 placed under the translucent mat 7, facing the camera 15, makes it possible to eliminate the shadow and increase the contrast between the leather 13 and the carpet 7 in the image 16.

  In this embodiment, the camera 15 has a resolution greater than 6,000 pixels and a 35 mm objective. The camera 15 captures an image of the leather 13 with a discretization of each pixel on 1,024 levels of gray. The image obtained is thus of excellent quality, with a resolution of 0.24 mm per pixel.

  Unrepresented control means serve to slave the driving of the leather 13 by the belt 12 under the camera 15 and the acquisition of the image 16 by the camera 15.

  The camera 15 is a so-called linear sensor camera and comprises a single row of sensor cells each corresponding to a pixel disposed perpendicularly to the direction of travel P. The time displacement of the leather 13 relative to this camera makes it possible to obtain a image 16 two-dimensional.

  Once the image 16 has been obtained, the following processes are performed in several phases by the calculation means executing the image processing software 6, a flowchart of which is shown in FIG. 2, in order to obtain a mapping of the defects.

  Beforehand, a pre-processing phase is performed, making it possible to eliminate the artifacts of the image 16. In particular, the linear sensor camera 15 causes artefacts appearing in the image in the form of columns perpendicular to the direction of movement of the image. leather, these columns appear because of the difference in sensitivity of the sensors of the camera 15. The pre-treatment consists of compensating for these differences in sensitivity. An estimate of the average per column of the image processed by the same sensor cell is calculated, and then the set of estimates of the average is grouped into a curve from which control points of a Bézier curve are calculated. Then a linear transformation of the intensities of the pixels of each column is performed, so that the new average of the column is identical to that given by the Bézier curve.

  The pre-treatment also allows a cropping of the image to center the image obtained on the leather.

  The phases a to h of the treatment, shown in FIG. 2, allowing the qualification of the leather are described below, these phases are as follows: a: geometrical analysis, b: segmentation into regions of homogeneous hue, c: fault detection , d: localization of defects, e: characterization of defects, f: identification of defects, g: selection by the user, h: classification, steps a to e constituting more particularly a mapping of defects.

  An image portion 16 is shown in FIG. 18.

  The first phase of the image processing 16 is a geometric analysis of the leather 13 for determining the quality zones and the carrier lines 27 of the lending direction. The meaning of lending corresponds to the direction of maximum elasticity of the leather at a given point. The quality zones are as follows: É The croup 23 É The collar 24 É The abutment 25 É The flanks 26 The first phase of the image processing is based on a geometrical analysis of the outline of the leather 13. In a first step , an image shown in FIG. 7 of the closed and unitary contour 28 of the leather 13 is obtained from the image of the support 29 of the leather, represented in FIG. 6.

  The support is obtained by thresholding the values of the pixels of the image 16, this operation being facilitated by the illumination by the source 19 which accentuates the contrast between the leather 13 and the carpet 7.

  Various characteristic points of the leather 13 are then determined, and in particular the center of gravity 30 of the leather 13, the extremal points of the tabs 32, 33, 34, 35, the maximum notch points of the sidewalls 36, 37 as well as the minimum points. and maximum of abutment 38, 39 and collar 40, 42.

  Geometric laws known in the state of the art make it possible to draw the boundaries of the quality zones 23, 24, 25, 26 as well as the lines bearing the direction of lending 27 as represented in FIGS. 5 and 6 from characteristic points. 30, 32, 33, 34, 35, 36, 37, 38, 39, 40, 42 described above.

  The results of this geometric analysis phase are: an image 20, an example of which is shown in FIG. 5, which groups together the different zones of quality of the leather. This image of the quality zones is useful for determining the severity of the defect because it depends on its position on the leather, and an image 22, an example of which is shown in FIG. 5, shows the carrier lines 27 of the directions of the defect. lending.

  The image 22 is generated for example for the automatic placement of the shoe parts. A piece of shoe represented in a computer format derived from computer-aided design or CAD software includes information about the meaning of lending. It is therefore possible to match the lending direction of the leather 13 given by the image 22 and that of the piece of shoe.

  The second phase b of the image processing 16 is a segmentation of regions of homogeneous color Rt of the leather 13.

  This step is an optional step intended for example to cut parts for a shoe. In this case, it is desirable not to cut a part on two regions of different hues.

  In a first step, a filtering of the original image 16 by a filter averaging over a mask, of size 15 by 15 pixels in this embodiment, is performed. This first step makes it possible to attenuate the contrast of the small defects and does not influence the result to be obtained because the purpose of this step is the detection of regions Rt of extended surface. These Rt regions have a significant contrast between them, the quality of the image 16 obtained amplifying this contrast with respect to human visualization.

  This phase b implements in a second step a Division and Fusion algorithm using a Tetra-tree. This algorithm is broken down into two stages. The first step of division consists of a recursive cutting of the image into parcels p, and the second so-called fusion step is an iterative process of grouping these parcels p into RT regions having common properties, that is, to say, in this mode of implementation, a neighboring hue. In the following description of the algorithm, a region of the image during the division process will be called parcel p, and a parcel grouping made during the merger will be called region RT.

  The division consists of a recursive cutting of the image 16 in accordance with a criterion of homogeneity of the pixels. If the criterion is not satisfied for a parcel p of the image, it is then divided into four sub-parts of equal sizes and the criterion is again evaluated in each of them. This process is repeated recursively until all parcels p satisfy the given criterion. At the start of the division, the criterion is evaluated on the entire image.

  Plot division p is shown in Figure 6 at a given stage of the algorithm. The pi parcel satisfying the homogeneity criterion is not divided on the contrary, the P2 parcel must be divided because it does not satisfy the criterion.

  The division therefore leads to a representation of the plots p obtained in the form of a tetra-tree T, represented in FIG. 9. Each node n of the tree T has exactly four children, the root no representing the entire image, and the leaves I representing the plots, square or rectangular, can no longer be subdivided.

  When the division of the image is completed, the fusion phase of the parcels thus obtained is carried out as follows.

  First, each parcel p is assigned to a region RT. At the beginning of this phase each parcel represents a region RT. Each pair of parcels p is then considered. For each of the pairs, if the parcels p do not belong to the same region RT, the merge criterion is evaluated on the union of the two regions RT to which these two parcels belong. If this is satisfied, the two RT regions are merged into one. The two parcels p considered are now attached to this new region RT. The merger is complete when all couples have been processed.

  The algorithm uses as a criterion the average of the intensity of the pixels. In the embodiment presented here, the threshold of division and melting is identical and equal to 10 gray levels. This threshold must be configurable according to the level of quality desired by the customer. The minimum dimension on the abscissa and on the ordinate is fixed at 5 pixels for each parcel of image.

  The result of this phase b is an image 44 showing the regions of homogeneous hue RT. This image is kept for later use.

  The processing then comprises a phase c in the image 16 generally relating to the detection of defects D grouping together two phases and and c2 respectively concerning: - c, the detection of linear form defects regrouping linear resolved defects DL as well as the altered texture regions RA, the linear defects corresponding in particular to wrinkles, veins or physical defects of linear form such as tears or scars, the altered texture regions RA corresponding for example to an extensive alteration of the leather due to wrinkles, a network of veins, crimps or a concentration of defects resolved and - c2: the detection of defects resolved circular shape or circular defects Dc, corresponding in particular to structural defects such as defects caused by parasites such as lice, ringworm , or scabies, or mechanical origin as a hole or appearance defects like spots.

  These two phases c, and c2 using variants of the same algorithm of the generalized likelihood ratio or GLRT, this one is detailed below.

  The GLRT or generalized likelihood ratio algorithm 20 makes it possible to highlight the breaks of statistical magnitudes present in an image and thus to highlight the defects D. An analysis window F traverses the image 45, as represented on FIG. Figure 8. The analysis window F is divided into two regions R, and R2 of different sizes NI and N2, NI and N2 representing the number of pixels present in each region. Let Il and 12, two samples of pixels contained in each of the two regions of the window, I the meeting of these two samples. For a position of the analysis window F, the following two hypotheses are considered: É Ho: the window F covers a homogeneous zone. The characteristic parameters of the texture are therefore the same in the two regions R1 and R2 and are symbolized by the symbols.

  É H1: the window F is positioned on a fault D located in the region R,. The characteristic parameters of the texture in the two regions R1 and R2 are therefore different and are equal to pi and P2.

  The two corresponding decisions are therefore: É Ho: there is no fault.

  P (IIH0) at a threshold t. whose value is directly related to a probability of false alarm tolerated. The decision is made HI if A is greater than t, and the decision Ho is taken in the opposite case.

  By making a hypothesis on the density densities of pixel intensities, we can explain the likelihood ratio A. We have indeed P (II Ho) = P (II, u) and P (II H,) = P (I , l) .P (12 l p2) because the regions R1 and R2 are independent.

  Several choices are possible for probability densities. The most suitable for this problem is the Gaussian model. In this case the characteristic parameters taken into account are the mean and the standard deviation.

  Thus pi represents the parameters of standard deviation ai and average mi on the first region, P2 represents the parameters of standard deviation 0-2 and average m2 on the second region and p represents the parameters of standard deviation oo and of average mo on the entire F analysis window. It is then possible to express the preceding probabilities as a function of the Gaussian probability density model chosen: N '1 (if m,) 2 P (I, 1p1) = 1 i = o 1 2c61 exp 261 NZ 1 - - - - (si-m2) 2 P (I21, u2) _II exp 1 = 0 27162 262 No 1 (s, m0) 2 P (I, u) = 101 -N / 2716o exp 260 The characteristic parameters mo, u0, mi , Cree, m2, 0-2, are unknown and must be replaced by their maximum likelihood estimates m0, 60, m 61, m2, 62.

É HI: a fault is present.

  The so-called Neyman-Person strategy allows to build the optimal decision rule, because it gives the best probability of detection for a fixed false alarm probability. This rule consists of comparing the likelihood ratio A: A = Pei, in which case we obtain the following expression: log (A) = N, log- - N2 log 62 + No log 6 with: 1 N '= (s1 m,) 2 Nl, = o 1 N2 62 = E (s1 1412) 2 N2 r = o 1 No = (s, mo) 2 No, = o It is possible by using a calculation of this likelihood ratio A for a window F successively centered on each of the pixels of the image to obtain a probability image 45, giving for each pixel of the image the value of the likelihood ratio A. The higher this value, the greater the presence of a fault D is likely.

  This algorithm is adapted to the images 16 of leathers 13 because it makes it possible to process noisy images. Indeed, this algorithm integrates the notion of noise in its calculations, avoiding a filtering of the image. The type of noise retained follows a Gaussian law.

  In this embodiment, the GLRT algorithm is used according to two variants described below: the linear GLRT and the circular GLRT, these two variants being differentiated mainly by the shape of the chosen analysis window.

  The linear GLRT is used in phase c, the detection of resolved defects, the linear resolved defects DL or the altered texture regions RA. The circular GLRT is used in the phase C2 of detection of circular resolved defects Dc.

  The linear GLRT is characterized by the inner shape or mask of the analysis window F which decomposes, as shown in FIGS. 9 to 12, into a band-shaped central region RI passing through the center of the analysis window. F and oriented at an angle A1 relative to the abscissa axis, and two regions R2 'and R2 "located on either side of the central region RI, R2' and R2" and completing the analysis window F. The regions R2 'and R2 "are grouped into a single region R2 to calculate the likelihood L of the region R1 with respect to the union R2 of the regions R2' and R2".

  For each position of the analysis window F, several likelihood calculations L are performed by varying the angle Al formed by the band and the abscissa axis X while keeping for each position the value of the likelihood ratio A the most high. In the embodiment shown, the angle A1 varies by an angular pitch of between 0 and 180.

  The size of the analysis window F best suited to the defects of leathers 13 is a size of 51 by 51 pixels. A smaller size results in a detection of the grain of the leather 13 as a defect. A larger size gives too low and too wide probabilities for a defect because the joining of the regions R2 'and R2 "is too important and may include other defects of the neighborhood, thus decreasing the statistical difference between the defect and its background. The selected width of the band forming the RI region is 5 pixels and the angular pitch is 15.

  The mask configurations of the analysis window F are shown in FIGS. 9 to 12 for an angular pitch of 45 and an angular range of 0 to 180.

  For each dimension of the analysis window F, a bandwidth forming the region RI makes it possible to obtain a detection of all the defects. Thus, for a given analysis window dimension F, it is not necessary to vary this thickness.

  The linear GLRT algorithm provides as a result: a probability image 45 as defined above, and an orientation plane 46 indicating, for each pixel of the original image 16, the value of the angle A1 of the mask having given the value of the likelihood ratio A the highest.

  an x 47 sizing image giving for each of the pixels a constant value corresponding to the width of the region Ri.

  - A sizing image 48 giving for each of the pixels a constant value corresponding to the thickness of the band forming the R region.

  These last two images 47, 48 provide additional information only one value, but allow to obtain a homogeneous result format between the linear GLRT and the circular GLRT.

  The circular GLRT makes it possible to determine the presence in each pixel of the leather image 13 of a circular type defect Dc. The analysis window F must have an inner shape or mask adapted to the shape of the defect Dc. The window F is therefore divided into a region R1, rectangular and concentric with the analysis window F, of variable size and size, and a region R2 completing the analysis window F. For each position of the window F, La RI region is distorted.

  Indeed, the largest value of the likelihood ratio A is obtained for a given defect Dc, when the region RI is of dimensions corresponding to the dimensions of the circular defect De.

  Region RI is deformed incrementally according to evolutions EX in x and Ey in x shown in FIG. 15, where x and y represent the dimensions in abscissa and ordinate of the region Ri.

  The RI region is defined by the following parameters: - The minimum dimension of the region RI in x and y, - The maximum dimension of the region RI x and y, The pitch between two dimension values in x or y of the region RI must be weak when the IR region is small because a defect Dc of a few pixels must have a region RI very close to its real form to obtain an optimal likelihood ratio A. Indeed, a small Dc defect detected by a poorly adapted mask greatly reduces its probability. In addition, it is likely to find several small defects in a large region. In the case of a large Dc defect, a dimension of the poorly adapted RI region influences the probability less because the size of the sample I is larger.

  The step for the detection of large defects can be more important.

  In the embodiment presented here, three passes are thus made for small, medium, and large region RI dimensions. For each pass, the step has a different value, small, medium, and large, all these values being between 5 and 10 pixels.

  The maximum size of the RI region is 130 pixels for the large flaw pass.

  The analysis window F has dimensions equal to the maximum dimensions of the region RI plus 3 pixels in this embodiment.

  The circular GLRT algorithm provides as a result: a probability image 45 as defined above.

  an orientation plane 46 containing a constant value of zero, which is not processed subsequently.

  an x 47 sizing image giving for each of the pixels of the image 45 an integer value of the x dimension of the region R 1 which gave the value of the highest likelihood ratio λ.

  a sizing image at y 48 giving for each of the pixels of the image 45 an integer value of the dimension in y of the region RI which gave the value of the likelihood ratio A at the highest.

  The presence of an orientation plane as a result of the circular GLRT algorithm allows a homogeneous result format to be obtained between the linear GLRT and the circular GLRT.

  The phases ci and c2 forming the defect detection phase c thus each provide as a result as described above: By way of example, FIGS. 19 and 20 represent two image portions 45 corresponding to the treatment of the portion of FIG. Figure 16 of Figure 18 by respectively a linear and circular GLRT used in phases CI and C2.

  The defect localization phase is intended to obtain a polygonal neighborhood V surrounding the circular and linear defects using the data provided by the detection step.

  The localization phase includes two phases di and d2 respectively concerning: - di: the location of the linear form defects grouping the linear resolved defects DL as well as the altered texture regions RA, and - d2: the location of the defects solved by shape circular or circular defects De, The phase d2 of location of circular defects Dc, is illustrated in Figure 14, being limited to the axis X, and described below.

  In a first step, the probability image 45 is thresholded with the threshold t set by the user to obtain a binary image grouping clusters of likelihood greater than the threshold, that is to say the set of points of the image 45 having a likelihood ratio A greater than the threshold.

  - a probability image 45 47 - an orientation plane 46 - a sizing image in x - a sizing image in y 48 A local maximum M of the likelihood ratio A per cluster is determined. A neighborhood V around this local maximum is defined whose dimensions in x and y are determined using the values corresponding to the position of the maximum in the sizing images 47, 48 in x and y.

  If a part of the cluster is outside the neighborhood V thus defined, a second local maximum M of the likelihood ratio A is sought in the residual part of the cluster, then also surrounded by a neighborhood V whose the dimensions in x and y are determined using the values corresponding to the position of the second maximum M in the sizing images 47, 48 in x and y. This approach makes it possible to isolate the large circular defects Dc in the form of a single circular fault Dc, and to isolate two small defects Dc close to each other as shown in FIG.

  The encompassing rectangular neighborhood V centered on each maximum M and dimensioned by the values present in the x and y design planes is then plotted.

  The phase di is similar to the phase d2, however in this case the polygonal neighborhood V is initially constituted by a rectangle whose dimensions, given by the dimensioning planes, correspond to the dimensions of the region RI of the analysis window F of phase c, Moreover, this rectangular neighborhood V is oriented by using the value of the orientation plane for the point corresponding to the maximum of probability.

  In addition, the phase di comprises a melting step making it possible to merge the close rectangular neighborhoods V. This melting step consists of traversing the contour of a neighborhood V by a circular shape of fixed diameter, and to merge the neighborhoods V reached by this shape with the neighborhood V whose circular shape traverses the contour. This operation is repeated for each neighborhood V and thus makes it possible to determine polygonal neighborhoods V. If a polygonal neighborhood V after the melting step is larger than a threshold set by the user, this neighborhood constitutes an altered texture region RA. In the opposite case, this neighborhood V is considered to surround a resolved linear fault DL.

  The results of the phases di and d2 are therefore: a set 49 of polygonal neighborhoods V surrounding the resolved defects DR, linear DL or circular Dc.

  a set 50 of regions of altered texture RA.

  Phase e of the image processing consists of a characterization of the resolved defects DR. This phase comprises: a first step making it possible to obtain the precise shape of the contour C of each solved defect DR, whether this defect is a linear solved DL or circular Dc defect; and a second step of calculating the photometric and geometric parameters. characteristics of each fault solved DR.

  In the first step, to obtain a precise contour C of the resolved defects DR, a polygonal statistical active contour algorithm or CASP is used. The input parameters of this algorithm are an image containing the defect DR and a contour C initialized by the neighborhood V. The CASP makes it possible to deform this contour C to obtain the precise shape of the contour C of the defect DR.

  In order to reduce the calculation time, the CASP algorithm is only used in this embodiment for large Dc defects. Indeed, it is not useful to obtain the precise form of a small Dc defect. If this defect Dc is 4 pixels and the surrounding rectangular neighborhood V surrounds perfectly in the broad sense, the approximation of the contour by the rectangular neighborhood V causes a negligible loss of 2 or 3 pixels or about 1 mm2 of the leather 13.

  The principle of the CASP, illustrated in Figures 15 to 17, is to minimize the energy from an initial contour formed by 4 K nodes corresponding to the vertices of the rectangular neighborhood V. The nodes K are moved either randomly or in a sequence determined by an adjustable step on the amplitude of the deformations from 1 to 15 pixels, then the energy is calculated. As the energy decreases, the nodes K are moved and the energy calculated again. If this becomes constant or increases the previous iteration provides the best possible segmentation of the contour C of the DR fault.

  The energy calculated in this type of active contour is a function of the statistical parameters of the image, here the variance and the average.

  At first, the CASP is applied on an initial contour composed of only four K nodes. The CASP determines a contour C closer to the defect but composed of only 4 K nodes, thus of imprecise form. Nodes K are then added so that the shape of the contour C can be refined.

  The CASP algorithm is reiterated to obtain a more precise form of the contour C. This multi-step approach makes it possible to speed up the calculation time because the addition of 15 K nodes from the first step for an initial contour C away from the defect DR real would imply the displacement of 19 K nodes over large distances and therefore require a much higher computation time.

  In the approach chosen, the step of approaching the contour is performed only on 4 K nodes and the 15 K nodes added move only over small distances.

  In the second step of the phase e, the parameters calculated for each solved defect DR are: - geometrical parameters, in particular the position of the center of gravity of the defect, circularity, total area of the defect, surfaces of the defects included in the different zones of quality, perimeter, elongation, stretch, average and variance of linear defect orientation, - photometric parameters, including average, variance and local contrast, and - parameters extracted from GLRT algorithms that have highlighted the defect.

  The result of the phase e is a set 49 of the resolved defects DR, containing the information relating to their contour C and the associated parameters.

  By way of example, FIG. 22 represents an image portion comprising contours C obtained by CASP algorithm surrounding large resolved defects.

  The next phase f is a fault identification phase. The purpose of this phase is to assign each defect D a defined category, a zone of quality and a level of visibility.

  In this embodiment, the defined categories are three in number: wrinkle, vein, and - physical defects.

  This last category gathers all the structural defects causing an alteration of the flower and the defects of aspects.

  The category of a fault is determined according to minimum and maximum limits of a parameter set, consisting of the parameters calculated for each defect in the previous phase.

  The quality zone 23, 24, 25, 26, to which the fault belongs is determined from the position of the defect on the support.

  The visibility is determined by the contrast parameter defined in phase e for each defect D. The result of this phase f is an assignment to each defect of a category, a quality zone and a visibility level. .

  The optional phase g consists of a selection by the user of the defects D to be retained in the subsequent phase.

  In this embodiment, this selection can be performed by the user by means of defect sorting functions on the basis of their category, visibility and surface.

  The result of phase f is a set 52 of faults to take into account.

  Phase h consists of a classification of defects present in category: wrinkle, vein, and physical defects, and in position in a quality zone: croup 23, - collar 24, abutment 25, and - flanks 26.

  The number, area and percentage of defects present on each part is also calculated.

  The user can enter a table defining the sorting criteria defining the level of quality or choice of the leather. For example a choice 1, of better quality, is defined by a maximum number of defects of each category per zone of quality.

  From all these data, the system attributes a choice or level of quality to the leather 13.

  The result of this phase g is therefore the allocation of a level of quality to the skin, as well as an image of the preserved defects, this image can be used for a subsequent cutting phase, as a binary image 53.

  By way of example, FIG. 23 represents an image portion 53 corresponding to an image portion 16 of FIG. 18.

  Certain phases of image processing 13 are optional depending on the desired use. For example, in the tanner, during a mapping of the defects of the leather 13 in the wet-blue stage, the segmentation in regions of homogeneous hue RT is not to be realized because the inhomogeneity of hue at the wet-blue level n ' It does not have any influence at the finished leather stage. In addition, the level of quality desired by the user does not necessarily take into account regions of homogeneous hue RT. Phase b can therefore be suppressed in this case.

  Depending on the type of article made and the type of skin used, the meaning of lending does not necessarily have influence. Phase a can be simplified in this case.

  It should be noted that phases a to h do not necessarily follow sequentially.

  In particular, the phases D and D2 for the detection of defects D are performed in several passes, one pass for the detection of linear faults, and three passes for the detection of circular faults, corresponding to different window sizes F, starting with a large F analysis window size and decreasing that size.

  The phases of location and determination of the contour C must be performed between each pass to make it possible to eliminate the detected defects D of the support of the image before making a second pass, in order not to detect twice the same defect D The rectangular neighborhood V can not be used to eliminate the defect of the image, since the defect DR can be partially outside the neighborhood V. Thus the sequence can be, for an image 13: phase a, phase b , phase c2 for detecting circular defects, for a pass concerning large defects, phase d2 for locating circular defects Dc, phase e of characterization, making it possible to obtain a precise contour for each defect D ,, and then suppressing this defect of the support in the image 13, phase c, of detection of the linear defects, pass single phase d, of localization of the linear defects, phase e of characterization of the defects, allowing to obtain a con precise tact for each defect D2, removal of linear defects and altered texture areas RI of the support 29 of the image 13, phase c2 of the circular defects detection, for a pass concerning defects of medium size, phase d2, - phase e, then suppression of defects Dc of medium size of the support 29, - phase c2, for the pass concerning defects Dc of small size - phase d2, phase e, at this stage, all the defects are isolated effectively, the 20 mapping is complete, - phase f, - phase g, phase h. According to another embodiment, the image processing can be performed, following the same method, not by an image processing software 6, but by a dedicated electronic card according to a technique known elsewhere. This second embodiment makes it possible to reduce the calculation time necessary to carry out the treatment.

  For the detection of W defects in the wet-blue, the device 30 will be completed by the addition of a chemical developer, increasing the contrast of defects causing deterioration of the flower.

  Moisture differences in wet-blue leather produce large variations in hues and almost 20% of defects are not visible at this stage by a sorter such as stubble or lice stings.

  The addition of a chemical developer facilitates the automatic detection of defects D open flower significantly improving their contrast. The blue pigments only bind to the parts where the flower is more or less altered. The coloring then disappears quickly of itself, the pH of the leather being acidic at this stage.

  The configuration of the device 2 is modifiable depending on the sector 5 selected, this sector may be tannery-tanning or shoe manufacturers or leather goods.

  Thus, the tannery acquisition station is positionable at the entrance or exit of the wringer. Before spinning, the leather is wet, its weight is important and flattening is not optimal.

  However, the hue is homogeneous and allows a better identification of the defects D. After the wringer, the leather is laid flat, the folds have been crushed and the regions of hues are different according to the thickness of the leather. Detection is less efficient.

  When used for leather goods or the manufacture of footwear or furniture or automobile, the acquisition station 4 may be deported from the production line, or integrate within cutting facilities.

  In the latter case, the inclination of the planes of the belt 7 may vary for integration within the cutting facilities.

  The invention is not limited to the embodiment described, it encompasses all variants. In particular, the lighting angle θ of the leathers 13 can vary.

Claims (22)

  1. A process for qualifying a leather (13) using a leather defect mapping (D) (13), implementing a mapping device (2), comprising means for acquiring a digital image (16). ) leather (13), and an image processing (16) obtained comprising at least a phase (ci, c2) of applying to the image (16) of the leather (13), after any prior steps , a generalized likelihood ratio statistical algorithm or GLRT, for calculating, in an analysis window (F) centered on a given pixel of the image (16) and comprising a mask with at least two regions (RI, R2 ), the likelihood ratio (A) of the presence of a defect (D) in a first region (RI) with respect to a second region (R2) and thereby obtaining a probability image (45) indicating for each pixel the value of this likelihood ratio (A).   2. Method according to claim 1, characterized in that it comprises a phase (ci) for detecting linear type defects (DL, RA), and a phase (c2) for detecting circular type defects (Dc).   3. Method according to claim 2, characterized in that the phase (ci) for detecting linear defects (DL, RA) uses a statistical algorithm of the generalized likelihood ratio or GLRT in its linear form, that is to say using a mask form for the analysis window (F) having a band-shaped central region (RI) passing through the center of the analysis window (F) and oriented by an angle (AI) relative to to the x-axis, and two lateral regions (R2) and (R2 ") located on both sides of central region (RI), the two lateral regions (R2) and (R2") completing the window of analysis (F), performing several likelihood ratio calculations (A) according to several values of the angle (Al) and retaining for each pixel: the largest value of the likelihood ratio (A) in an image of probability (45) and - the value of the angle (Al) giving the largest value of the likelihood ratio nce (A) in an orientation plane (46).   4. Method according to one of claims 2 and 3, characterized in that the phase (c2) for detecting circular defects (Dc) uses a statistical algorithm of the generalized likelihood ratio or GLRT in its circular form, that is, that is to say using a mask form for the analysis window (F) having a central region (RI) concentric with the analysis window (F), of rectangular shape and a region (R2) completing the window of analysis (F), by carrying out several calculations of the likelihood ratio (A) by deforming the central region (RI) to increase its size, in the direction of the abscissa and the ordinate, keeping for each pixel: - the most important value of the likelihood ratio (A) in a probability image (45), and - the abscissa and ordinate dimensions of the central region (RI) having given the largest value of the likelihood ratio (A) respectively stored in two images of dimensi on the abscissa and ordinate (47, 48).   5. Method according to one of claims 3 and 4, characterized in that it comprises a phase (d2) for locating circular defects (Dc), comprising the steps of: - looking for local maxima (M) of the report likelihood (A) in clusters or sets of image points 45 having a likelihood ratio (A) greater than a threshold value (t), - defining a neighborhood (V) around this local maximum (M) of which the dimensions in abscissa (x) and ordinate (y) are determined using values from a detection phase (c ,, c2) and - looking for a second local maximum (M) of the likelihood ratio (A) in the residual part of the cluster, if a part of the cluster is outside the neighborhood (V) thus defined, - a neighborhood (V) also being defined around this second cluster whose dimensions on the abscissa ( x) and ordinate (y) are determined using values from a detection phase (cl, c2).   6. Method according to claim 5, characterized in that the values from a detection phase (ci, c2) for determining the dimensions in abscissa (x) and ordinate (y) of a neighborhood (V) around of this local maximum (M) are the values corresponding to the position of the maximum (M) in the sizing images (47, 48).   7. Method according to one of claims 3 to 6, characterized in that it comprises a phase (di) of location of linear defects (DL RA), comprising the steps of: - looking for local maxima (M) of likelihood ratio (A) in clusters or sets of image points 45 having a likelihood ratio (A) greater than a threshold value (t), - defining a neighborhood (V) around this local maximum (M) whose dimensions correspond to those of the region (RI) of the phase (d1), this neighborhood (V) being oriented by using the value corresponding to the position of the maximum (M) in the orientation plane (46) and searching for a second local maximum (M) of the likelihood ratio (A) in the residual part of the cluster, if a part of the cluster 10 is outside the neighborhood (V) thus defined, - a neighborhood (V) is also defined around this second cluster whose dimensions correspond to those of the region (RI) of the phase (c1), this neighborhood (V) being oriented by using the value corresponding to the position of the maximum (M) in the orientation plane (46).   8. The method as claimed in claim 7, characterized in that the phase (di) for locating the linear defects (DL RA) also comprises a step of merging the close neighborhoods V and isolating the neighborhoods V with a surface greater than a threshold. given as an altered texture region (RA).   9. Method according to one of claims 1 to 8, characterized in that it comprises a characterization phase (e) comprising a step of performing a determination of the precise shape of the contour (C) of the resolved defects (DR). using a polygon statistical active contour algorithm or CASP.   10. The method of claim 9, characterized in that the phase (e) of characterization comprises a step of calculating geometric parameters and photometric characteristics of each defect (D).   11. Method according to one of claims 1 to 10, characterized in that it comprises an identification phase (f) consisting of classifying the defects (D) into categories.   12. Method according to one of claims 1 to 11, characterized in that it comprises a phase (g) for selection of defects (D) to be taken into account by a user.   13. Method according to one of claims 1 to 12, characterized in that it comprises a phase (h) classification of the leather, to assign a level of quality, depending on the position, the type and the defect surface (D) on the leather (13).   14. Method according to one of claims 1 to 13, characterized in that the characteristic parameters used in the calculation of the likelihood ratio (A) are the mean (m) and the standard deviation (a).   15. Method according to one of claims 1 to 14, characterized in that the model chosen to represent the densities of probabilities of the intensity of the pixels is the Gaussian model.   16. Method according to one of claims 1 to 15, characterized in that it further comprises a phase (a) of performing a geometric analysis of the leather (13) to identify the quality zones (23, 24, 25, 26) and the carrier lines (27) elasticity of the leather.   17. A method according to claims 1 to 16, characterized in that it further comprises a phase (b) of determining a leather segmentation in regions of homogeneous hues (RT) using Division and Fusion algorithm using a Tetra- tree (T).   18. Method according to one of claims 2 to 17, characterized in that it comprises several passes in the phase (c2) for detecting circular defects corresponding to different analysis window sizes (F).   19. Mapping device (2) for implementing the method according to one of claims 1 to 18 comprising conveying means (7, 8) leathers (13) for routing them to a position acquisition device (4) comprising a digital camera (15) as well as lighting means, characterized in that the lighting means comprise a transmission lighting source (19) through the part of the conveying means constituting the leather support (13) and a reflection lighting source (17).   20. Device according to claim 19, characterized in that the angle (9) of the direction of the reflection lighting source (17) with respect to the plane of the leather (13) is between 15 and 35.   21. Device according to one of claims 19 and 20, characterized in that it comprises calculation means (5) for software image processing.   22. Device according to one of claims 19 and 20, characterized in that it comprises dedicated electronic calculation means for performing image processing.