Leather material surface defect detection method based on feature modeling saliency detection

Abstract

The invention discloses a method for detecting surface defects of leather materials based on feature modeling saliency detection, comprising the following steps: S01, firstly collecting video information of the leather material, and selecting a plurality of frames of images in the video as images to be detected; S02, then using a bottom-up model to perform preliminary acquisition of saliency images to obtain a first saliency image; S03, using an equal-area multi-histogram equalization algorithm to enhance the first saliency image to obtain a second saliency image; S04, processing the second saliency image based on superpixel re-optimization to obtain a third saliency image; S05, finally, marking the location of the defect according to the third saliency image, and using a global threshold method to perform adaptive thresholding on the obtained third saliency image. The invention effectively reduces the amount of algorithm calculation, and in the case where the defects are of different sizes and are scattered in multiple areas of the leather material, the detection speed is fast and the detection result is more stable and accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of material defect detection, and in particular to a leather material surface defect detection method based on feature modeling significance detection. Background Art

[0002] At present, my country's leather products industry has long relied on manual labor to detect and locate defects on the entire leather. However, due to the influence of lighting conditions, different worker experience, and changes in emotions, physical strength, working hours and other factors, it is easy to cause problems such as low efficiency in detection and sample cutting. At present, domestic and foreign researchers have proposed many different methods for leather surface defect detection, but most of them are aimed at detecting local defects on the leather surface and classifying known clear defects, and fail to effectively detect and locate defects on the entire leather at one time. Summary of the invention

[0003] Technical purpose: In view of the deficiency that unknown defects on the surface of existing leather materials cannot be effectively detected and located, the present invention discloses a method for detecting surface defects of leather materials based on feature modeling significance detection.

[0004] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0005] A method for detecting surface defects of leather materials based on feature modeling saliency detection comprises the following steps:

[0006] S01, first collect video information of leather materials, and select several frames of images in the video as images to be detected;

[0007] S02, then using a bottom-up model to perform preliminary acquisition of a saliency image to obtain a first saliency image;

[0008] S03, enhancing the first saliency image using an equal-area multi-histogram equalization algorithm to obtain a second saliency image;

[0009] S04, processing the second saliency image based on superpixel re-optimization to obtain a third saliency image;

[0010] S05. Finally, the defect locations are marked according to the third saliency image, and the third saliency image is adaptively thresholded using the global threshold method to obtain a binary image. The white area in the binary image is the detected defect area, and the defect locations on the surface of materials such as leather are marked.

[0011] Preferably, in step S02 of the present invention, the process of obtaining the first saliency image includes the steps of:

[0012] S021, first divide the central area I in the image to be detected I 1 and marginal area I 2 , Central Area I 1 Concentric with the image to be detected I, remove the central area I of the image to be detected I 1 The outer edge area I 2 , edge region I 2 The pixels inside are edge pixels;

[0013] S022, then the central area I of the image to be detected is 1 and marginal area I 2 Pixel information S i All are converted to CIELab color space to obtain new pixel information i represents the pixel number in the image to be detected;

[0014] S023, calculating the feature vector of each edge pixel, and clustering the edge pixels using the K-means++ clustering algorithm;

[0015] S024, counting the number of clustering groups after edge pixels are clustered, and the number of edge pixels in each clustering group, and calculating the central eigenvalue of each clustering group using the arithmetic mean;

[0016] S025. Finally, the distance between the feature vector of each pixel in the image to be detected and the center of each cluster group is calculated, and the distance from each pixel to the centers of all cluster groups is normalized to obtain a first saliency image.

[0017] Preferably, in step S023 of the present invention, a neighborhood block centered on each edge pixel is used to represent the features of the corresponding edge pixel, and the pixel information S of each pixel in the neighborhood block is used. * The component value is multiplied by the corresponding weight coefficient as the corresponding edge pixel feature value, and the weight coefficient is the reciprocal of the distance between the pixel in the neighborhood block and the corresponding edge pixel.

[0018] Preferably, the neighboring block area of ​​the present invention is an area consisting of n pixels of width adjacent to the edge pixel centered on it, and the size of the neighboring block area is n*n, where n is an odd number. For edge pixels close to the outside, when the distance between the edge pixel and the image edge is less than (n-1) / 2, the neighborhood block is an area consisting of n pixels of width from the image edge inward.

[0019] Preferably, the process of clustering using the K-means++ clustering algorithm of the present invention comprises the steps of:

[0020] S0231. Randomly select a pixel from all edge pixels as the cluster center;

[0021] S0232. Calculate the maximum distance from each edge pixel to the existing cluster center, D i Indicates the farthest distance from the i-th pixel to the current cluster center. The largest edge pixel is the next cluster center;

[0022] S0233, calculating the distance between each edge pixel and each current cluster center and assigning the edge pixel to the cluster group corresponding to the cluster center with the smallest distance;

[0023] S0234, judging the number of samples in each cluster group, dissolving the cluster group when the number of samples is less than the set minimum value, and dividing the cluster group into two subgroups when the number of samples is greater than the set maximum value;

[0024] S0235. Determine whether the set termination condition is met. If the condition is met, the calculation is terminated. If not, repeat steps S0232-S0235.

[0025] Preferably, the central region I of the present invention 1 The width percentage relative to the image to be detected I is c, the height percentage relative to the image to be detected I is d, 0.8≤c<1, 0.8≤d<1.

[0026] Preferably, in step S03 of the present invention, the process of obtaining the second saliency image includes the steps of:

[0027] S031. First calculate the probability density function of the gray level of the first saliency image: WH is the total number of pixels in the first saliency image, n k is the number of pixels with gray value k in the first saliency image, L is the gray level of the first saliency image; the cumulative distribution function of the gray level of the first saliency image is:

[0028] S032, taking the median gray value of the first saliency image I mean As a standard, take As the threshold, the first saliency image is divided into J sub-images, and then the grayscale value F(k) after the grayscale value k of each pixel is transformed is obtained, F(k) = (L-1) × c(k), (k = 0, 1, ..., L-1), and a grayscale transformation lookup table is generated;

[0029] S033. Finally, J sub-images after histogram equalization are obtained according to the grayscale transformation lookup table, and then all the sub-images are superimposed to obtain the second saliency image enhanced by the equal-area multi-histogram equalization algorithm.

[0030] Preferably, in step S04 of the present invention, the process of superpixel re-optimization comprises the steps of:

[0031] S041, firstly perform super-pixel representation on the second saliency image;

[0032] S042, then using the SLIC algorithm to perform superpixel segmentation, dividing it into several superpixel regions, the salient value of each superpixel region is the average of the salient values ​​of all pixels in the region;

[0033] S043, the qth superpixel region, set its saliency value to S c (q), and further processed using the following algorithm, the significant value after processing is:

[0034]

[0035] Among them, j∈Θ 1 represents the set of neighboring superpixels within a distance of 1 from the qth superpixel region, is a weight parameter based on the Euclidean distance between the qth superpixel region and the superpixel j within 1 distance; m∈Θ 2 is the set of neighboring superpixels within a distance of 2 from the qth superpixel, It is a weight parameter based on the Euclidean distance between the qth superpixel region and the superpixel m within 2 distances. d(q)∈{0,1} is a binary quantity. If the superpixel region q is connected to a strong edge in the image, then d(q)=1. If the superpixel region q is not directly connected to a strong edge in the image, then d(q)=0.

[0036] Beneficial effects: The method for detecting leather material surface defects based on feature modeling significance detection provided by the present invention has the following beneficial effects:

[0037] 1. The present invention has advantages over traditional methods in terms of the accuracy of defect extraction and the operating efficiency of the algorithm. The visually significant defect detection method proposed by the present method effectively reduces the amount of algorithm calculation. When the defects are of different sizes and scattered in multiple areas of the leather material, the detection speed is fast and the detection results are more stable and accurate.

[0038] 2. The present invention uses a bottom-up model to acquire the first saliency image, which does not require prior information, does not require time consumption on parameter training, has low time complexity, and is convenient for real-time processing of images.

[0039] 3. In the process of pixel clustering, the present invention calculates the farthest distance from the edge pixel to the existing cluster center. The farthest distance means that the difference between the edge pixel and the existing cluster center is the largest, which is the place where defects are most likely to occur, thereby improving defect recognition efficiency and shortening detection time.

[0040] 4. The first saliency image of the present invention is closer to the distance map calculated by the larger cluster group. The larger cluster group has a more obvious characterization effect on the background features, can better represent the distribution of the background features, and has a stronger effect on extracting significant defects.

[0041] 5. The present invention utilizes the equal-area multi-histogram equalization algorithm to enhance the first salient image, so that the enhanced second salient image has a larger information entropy, thereby solving the problem of information loss caused by the standard histogram equalization algorithm.

[0042] 6. The present invention re-optimizes the superpixel area after superpixel segmentation. For the superpixel area that is not connected to the strong edge, the superpixel saliency value of the surrounding adjacent area is used to determine, and a third saliency image is obtained to better highlight the defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0044] Figure 1 It is a schematic diagram of dividing the image to be detected according to the present invention. DETAILED DESCRIPTION

[0045] The present invention will be described more clearly and completely below by way of a preferred embodiment in conjunction with the accompanying drawings, but the present invention is not limited to the scope of the embodiment.

[0046] The present invention provides a method for detecting surface defects of leather materials based on feature modeling significance detection, comprising the steps of:

[0047] S01, first collect video information of leather materials, and select several frames of images in the video as images to be detected;

[0048] In a specific embodiment, a frame image is acquired at intervals of 1 second in the collected video as the image to be detected;

[0049] S02, then using a bottom-up model to perform preliminary acquisition of a saliency image to obtain a first saliency image;

[0050] The process of obtaining the first saliency image includes the following steps:

[0051] S021, first divide the image to be detected into a central area and an edge area, the central area is concentric with the image to be detected, the image to be detected excluding the central area is the edge area, and the pixels in the edge area are edge pixels;

[0052] Assume that the width of the image to be detected is W pixels and the height is H pixels, then the total number of pixels in the image to be detected is N = WH, and the information S of pixel i in the image to be detected is i for:

[0053] S i =[r i g i b i ], (i=1,2,…,WH) (1)

[0054] Among them, r i , g i 、b i Represents the values ​​of the R, G, and B components of pixel i respectively.

[0055] Set the central area I in the image to be detected I 1 , Central Area I 1 The width percentage of the image to be detected I is c (0.8≤c<1), and the height percentage is d (0.8≤d<1), then the central area I can be obtained 1 The width is cW pixels and the height is dH pixels. 1 The outer edge area I 2 , according to its relationship with the central area I 1 The relative position of Figure 1 The upper edge, lower edge, left edge, and right edge shown are evenly distributed, with the upper and lower edges having the same height and the left and right edges having the same width. 2 The total number of edge pixels is N bian =(1-cd)WH.

[0056] S022, then the pixel information S of the central area and the edge area of ​​the image to be detected is i All are converted to CIELab color space to obtain new pixel information

[0057]

[0058] L i * =116f(Y i / Y n )-16,(i=1,2,…,WH)

[0059] a i * =500[f(X i / X n )-f(Y i / Y n)],(i=1,2,…,WH)

[0060] b i * =500[f(Y i / Y n )-f(Z i / Z n )],(i=1,2,…,WH)

[0061] Where f is

[0062]

[0063] Among them, X i , Y i , Z i is the intermediate variable for conversion; X n , Y n , Z n The value is 1; L * Indicates the brightness of the pixel i color, ranging from 0 to 100, L * =0 means black, L * =100 means white; a * represents the position of pixel i between red and green, a * A positive value indicates red, a * Negative values ​​indicate green; b * represents the position of pixel i between yellow and blue, b * A positive value indicates yellow, b * A negative value indicates blue. From the above formula, we can get the L of all pixels in the image I to be detected. i * 、a i * 、b i * Numeric value.

[0064] S023. Extract all edge pixels, where each edge pixel contains L i * 、a i * 、b i * Component, record all edge pixel information Calculate the feature vector of each edge pixel and use K-means++ clustering algorithm to cluster edge pixels;

[0065] In step S023, a neighborhood block centered on each edge pixel is used to represent the features of the corresponding edge pixel, and the pixel information S of each pixel in the neighborhood block is used. *The component value is multiplied by the corresponding weight coefficient as the corresponding edge pixel feature value. The weight coefficient is the reciprocal of the distance between the pixel in the neighborhood block and the corresponding edge pixel, specifically:

[0066] Assume that the neighborhood block size of edge pixels is n*n, where n is an odd number. For edge pixels close to the outside, when the distance between the edge pixel and the image edge is less than (n-1) / 2, the neighborhood block is an area consisting of n pixels in width from the image edge. The feature vector of edge pixel i can be expressed as:

[0067] T i =[T i L T i a T i b ] T ,(i∈I 2 )

[0068] Among them, T i L 、T i a 、T i b are L of all pixels in the n*n neighborhood blocks. i * 、a i * 、b i * Numeric value.

[0069]

[0070]

[0071]

[0072] where α r ,β r ,x r (r=1,...,n*n) are all weight factors, indicating the L of each pixel in the neighborhood block. * 、a * 、b * The weight coefficient corresponding to the component can generally be taken as the inverse of the distance between it and the corresponding edge pixel of the neighborhood block. For example, if the distance between a pixel in the neighborhood block and the corresponding edge pixel is 2, then α r ,β r ,x r All are taken as 1 / 2.

[0073] The clustering process of K-means++ clustering algorithm includes the following steps:

[0074] S0231. Randomly select a pixel from all edge pixels as the cluster center;

[0075] S0232. Calculate the maximum distance from each edge pixel to the existing cluster center, D i Indicates the farthest distance from the i-th pixel to the current cluster center. The largest edge pixel is selected as the next cluster center, I 2 Indicates the marginal area;

[0076] S0233, calculating the distance between each edge pixel and each current cluster center and assigning the edge pixel to the cluster group corresponding to the cluster center with the smallest distance;

[0077] S0234, judging the number of samples in each cluster group, dissolving the cluster group when the number of samples is less than the set minimum value, and dividing the cluster group into two subgroups when the number of samples is greater than the set maximum value;

[0078] S0235, determine whether the set termination condition is met, if the condition is met, the calculation is terminated, if not, repeat steps S0232-S0235. The termination condition can be that the internal elements of each cluster group are no longer changing, or the iteration exceeds a certain number of times.

[0079] S024. Count the number of clustering groups M after clustering edge pixels, and the number of edge pixels in each clustering group, and record the lth clustering group G l The number of edge pixels is N(l), 1≤l≤M, and the central eigenvalue is C(l), that is,

[0080] S025. Finally, the distance between the feature vector of each pixel in the image to be detected and the center of each cluster group is calculated, and the distance from each pixel to the centers of all cluster groups is normalized to obtain a first saliency image.

[0081] For each pixel in the image to be detected, first take its corresponding n*n neighborhood block to obtain the feature vector corresponding to the pixel, and then calculate the distance between the feature vector and the center of each cluster group. The calculation method is as follows:

[0082]

[0083] in, is the pixel i in the image to be detected I and the lth cluster group G l Center distance, T i is the feature vector corresponding to pixel i.

[0084] Then the distance from pixel i to the center of all cluster groups is normalized and summed. The calculation method is as follows:

[0085]

[0086] in, represents the normalized distance from pixel i to the center of all cluster groups; w l The distance weight for each cluster group is set as

[0087]

[0088] The weight value is the ratio of the number of edge pixels in each cluster group to the total number of edge pixels.

[0089] S03, enhancing the first saliency image using an equal-area multi-histogram equalization algorithm to obtain a second saliency image;

[0090] S031. First calculate the probability density function of the gray level of the first saliency image: WH is the total number of pixels in the first saliency image, n k is the number of pixels with gray value k in the first saliency image, L is the gray level of the first saliency image; the cumulative distribution function of the gray level of the first saliency image is:

[0091] S032, taking the median gray value of the first saliency image I mean As a standard, take As the threshold, the first saliency image is divided into J sub-images, and then the grayscale value F(k) after the grayscale value k of each pixel is transformed is obtained, F(k) = (L-1) × c(k), (k = 0, 1, ..., L-1), and a grayscale transformation lookup table is generated;

[0092] S033. Finally, J sub-images after histogram equalization are obtained according to the grayscale transformation lookup table, and then all the sub-images are superimposed to obtain the second saliency image enhanced by the equal-area multi-histogram equalization algorithm.

[0093] S04, processing the second saliency image based on superpixel re-optimization to obtain a third saliency image;

[0094] In step S04, the process of superpixel re-optimization includes the following steps:

[0095] S041, firstly perform super-pixel representation on the second saliency image;

[0096] S042, then using the SLIC algorithm to perform superpixel segmentation, dividing it into several superpixel regions, the salient value of each superpixel region is the average of the salient values ​​of all pixels in the region;

[0097] S043, the qth superpixel region, set its saliency value to Sc (q), and further processed using the following algorithm, the significant value after processing is:

[0098]

[0099] Among them, j∈Θ 1 represents the set of neighboring superpixels within a distance of 1 from the qth superpixel region, is a weight parameter based on the Euclidean distance between the qth superpixel region and the superpixel j within 1 distance; m∈Θ 2 is the set of neighboring superpixels within a distance of 2 from the qth superpixel, It is a weight parameter based on the Euclidean distance between the qth superpixel region and the superpixel m within 2 distances. d(q)∈{0,1} is a binary quantity. If the superpixel region q is connected to the strong edge in the image, then d(q)=1. If the superpixel region q is not directly connected to the strong edge in the image, then d(q)=0. The strong edge is the gradient edge on the image.

[0100] S05. Finally, the defect location is marked according to the third saliency image, and the third saliency image is adaptively thresholded using the global threshold method. The global median is taken and then binarized to obtain a binary image. The white area in the binary image is the detected defect area, and the defect location on the surface of the leather or other material is marked.

[0101] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting surface defects of leather materials based on feature modeling saliency detection, characterized in that: Includes steps: S01, first collect video information of leather materials, and select several frames of images in the video as images to be detected; S02, then using a bottom-up model to perform preliminary acquisition of a saliency image to obtain a first saliency image; S03, enhancing the first saliency image using an equal-area multi-histogram equalization algorithm to obtain a second saliency image; S04, processing the second saliency image based on superpixel re-optimization to obtain a third saliency image; S05. Finally, the defect location is marked according to the third saliency image, and the third saliency image is adaptively thresholded using a global threshold method to obtain a binary image. The white area in the binary image is the detected defect area, and the defect location on the surface of the leather material is marked; In step S02, the process of obtaining the first saliency image includes the following steps: S021, first divide the image to be detected I into a central area I1 and an edge area I2, the central area I1 is concentric with the image to be detected I, the image to be detected I excluding the central area I1 is the edge area I2, and the pixels in the edge area I2 are edge pixels; S022, then the pixel information S of the central area I1 and the edge area I2 of the image to be detected is i All are converted to CIELab color space to obtain new pixel information i represents the pixel number in the image to be detected; S023, calculating the feature vector of each edge pixel, and clustering the edge pixels using the K-means++ clustering algorithm; S024, counting the number of clustering groups after edge pixels are clustered, and the number of edge pixels in each clustering group, and calculating the central eigenvalue of each clustering group using the arithmetic mean; S025. Finally, the distance between the feature vector of each pixel in the image to be detected and the center of each cluster group is calculated, and the distance between each pixel and the center of all cluster groups is normalized to obtain a first saliency image. In step S023, a neighborhood block centered on each edge pixel is used to represent the characteristics of the corresponding edge pixel, and the pixel information S* component value of each pixel in the neighborhood block is multiplied by the corresponding weight coefficient as the corresponding edge pixel characteristic value, and the weight coefficient is the inverse of the distance between the pixel in the neighborhood block and the corresponding edge pixel; The neighborhood block is an area consisting of n pixels of width adjacent to the edge pixel. The size of the neighborhood block is n*n, where n is an odd number. For edge pixels close to the outside, when the distance between the edge pixel and the image edge is less than (n-1) / 2, the neighborhood block is an area consisting of n pixels of width from the image edge to the inside.

2. The method for detecting leather material surface defects based on feature modeling saliency detection according to claim 1, characterized in that: The clustering process of K-means++ clustering algorithm includes the following steps: S0231. Randomly select a pixel from all edge pixels as the cluster center; S0232. Calculate the maximum distance from each edge pixel to the existing cluster center, D i Indicates the farthest distance from the i-th pixel to the current cluster center. The largest edge pixel is the next cluster center; S0233, calculating the distance between each edge pixel and each current cluster center and assigning the edge pixel to the cluster group corresponding to the cluster center with the smallest distance; S0234, judging the number of samples in each cluster group, dissolving the cluster group when the number of samples is less than the set minimum value, and dividing the cluster group into two subgroups when the number of samples is greater than the set maximum value; S0235. Determine whether the set termination condition is met. If the condition is met, the calculation is terminated. If not, repeat the steps S0232-S0235.

3. The method for detecting leather material surface defects based on feature modeling saliency detection according to claim 1, characterized in that: The width percentage of the central area I1 relative to the image to be detected I is c, and the height percentage relative to the image to be detected I is d, 0.8≤c<1, 0.8≤d<1.

4. The method for detecting leather material surface defects based on feature modeling saliency detection according to claim 1, characterized in that: In step S03, the process of obtaining the second saliency image includes the following steps: S031. First calculate the probability density function of the gray level of the first saliency image: WH is the total number of pixels in the first saliency image, n k is the number of pixels with gray value k in the first saliency image, L is the gray level of the first saliency image; the cumulative distribution function of the gray level of the first saliency image is: S032, taking the median gray value of the first saliency image I mean As a standard, take As the threshold, the first saliency image is divided into J sub-images, and then the grayscale value F(k) after the grayscale value k of each pixel is transformed is obtained, F(k) = (L-1) × c(k), (k = 0, 1, ..., L-1), and a grayscale transformation lookup table is generated; S033. Finally, J sub-images after histogram equalization are obtained according to the grayscale transformation lookup table, and then all the sub-images are superimposed to obtain a second saliency image enhanced by the equal-area multi-histogram equalization algorithm.

5. The method for detecting surface defects of leather materials based on feature modeling saliency detection according to claim 1, characterized in that: In step S04, the process of superpixel re-optimization includes the following steps: S041, firstly perform super-pixel representation on the second saliency image; S042, then use the SLIC algorithm to perform superpixel segmentation, dividing it into several superpixel regions, and the salient value of each superpixel region is the average of the salient values ​​of all pixels in the region; S043, the qth superpixel region, set its saliency value to S c (q), and further processed using the following algorithm, the significant value after processing is: Among them, j∈Θ 1 represents the set of neighboring superpixels within a distance of 1 from the qth superpixel region, is a weight parameter based on the Euclidean distance between the qth superpixel region and the superpixel j within 1 distance; m∈Θ 2 is the set of neighboring superpixels within a distance of 2 from the qth superpixel, It is a weight parameter based on the Euclidean distance between the qth superpixel region and the superpixel m within 2 distances; d(q)∈{0,1} is a binary quantity. If the superpixel region q is connected to the strong edge in the image, then d(q)=1; if the superpixel region q is not directly connected to the strong edge in the image, then d(q)=0.

Images