Automobile leather defect detection method and system based on visual detection

Abstract

The application discloses a kind of based on visual inspection's automobile leather defect detection method and system, it is related to industrial visual inspection technical field, including, by the parallax relationship between the analysis multi-view automobile leather image and illumination direction reflection characteristic, reconstructs leather surface three-dimensional topography, generates three-dimensional point cloud diagram;Extract the curvature variation feature of three-dimensional point cloud diagram automobile leather surface, and carry out multi-scale texture analysis, identify and mark potential defect area;By three-dimensional topography measurement method, extract the defect three-dimensional morphological feature of potential defect area, calculate defect three-dimensional geometric parameter;By analyzing the defect geometry feature and spatial distribution law of defect three-dimensional geometric parameter, and utilize preset grading determination rule to divide defect grade, generate the automobile leather quality evaluation report containing defect three-dimensional coordinates;The application is combined by curvature-texture multi-scale fusion detection, and the identification ability of complex surface defect is significantly enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial visual inspection, and in particular to a vehicle leather defect detection method and system based on visual inspection. BACKGROUND

[0002] In recent years, vehicle leather surface defect detection technology has gradually developed from traditional manual visual inspection to automated detection methods based on machine vision. Current mainstream technologies mainly use two-dimensional image processing algorithms (such as edge detection, texture analysis) and single-view three-dimensional scanning techniques (such as structured light projection, laser triangulation) for defect recognition; among them, multi-spectral imaging technology can effectively identify surface color differences and texture abnormalities by analyzing the optical properties of different wavebands; and three-dimensional reconstruction methods based on structured light can achieve sub-millimeter precision surface topography measurement.

[0003] Current vehicle leather defect detection technology has obvious limitations in quantifying three-dimensional features. Existing methods mainly rely on two-dimensional image analysis techniques, which can only obtain the planar projection features of defects (such as area, perimeter, etc.), but cannot accurately represent key three-dimensional geometric parameters of defects such as depth of concave and steepness of edge. Especially when dealing with complex curved leather, this two-dimensional quantization method will result in the loss of important topographic information, such as the inability to distinguish the essential differences between surface scratches and deep cracks. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a vehicle leather defect detection method based on visual inspection to solve the problem of insufficient detection accuracy caused by the inability of existing technology to quantify three-dimensional geometric features of defects.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a vehicle leather defect detection method based on visual inspection, which includes: obtaining multi-view vehicle leather images through multi-view synchronous shooting and generating clear multi-view vehicle leather images through dynamic curved surface focusing; reconstructing the three-dimensional topography of the leather surface by analyzing the parallax relationship and illumination direction reflection characteristics between the multi-view vehicle leather images, and generating a three-dimensional point cloud map; extracting the curvature variation features of the vehicle leather surface of the three-dimensional point cloud map, and performing multi-scale texture analysis to identify and mark potential defect areas; extracting the three-dimensional morphological features of the potential defect areas through three-dimensional topography measurement methods, and calculating the three-dimensional geometric parameters of the defects; analyzing the defect geometric features and spatial distribution rules of the three-dimensional geometric parameters of the defects, and dividing the defect levels using a pre-set grading judgment rule to generate a vehicle leather quality evaluation report containing the three-dimensional coordinates of the defects.

[0008] As a preferred scheme of the automobile leather defect detection method based on visual detection, the multi-view automobile leather image comprises automobile leather images of orthographic view, oblique view and grazing view.

[0009] As a preferred scheme of the automobile leather defect detection method based on visual detection, the multi-view automobile leather image comprises automobile leather images of orthographic view, oblique view and grazing view.

[0010] Based on the multi-view automobile leather image, the surface Gaussian curvature and average curvature distribution of the automobile leather are calculated, and the focusing of the binocular camera is controlled in real time to obtain a set of locally optimal focusing automobile leather images of each view angle.

[0011] The set of locally optimal focusing automobile leather images of each view angle is subjected to frequency domain fusion, and sub-pixel level geometric correction is performed through multi-view feature matching to generate a set of multi-view automobile leather images with clear full field of view.

[0012] The set of multi-view automobile leather images with clear full field of view is subjected to polarization reflection component separation and light intensity compensation to generate enhanced multi-view automobile leather images.

[0013] The enhanced multi-view automobile leather images are subjected to parallax analysis and calculation of the depth distance from the automobile leather surface to the binocular camera to generate clear focusing multi-view automobile leather images.

[0014] As a preferred scheme of the automobile leather defect detection method based on visual detection, the multi-view automobile leather image comprises automobile leather images of orthographic view, oblique view and grazing view.

[0015] Based on the clear focusing multi-view automobile leather images, the parallax values between the view angles are calculated through the SIFT feature matching algorithm to obtain a parallax matrix with sub-pixel level accuracy.

[0016] The parallax matrix is combined with a preset light direction to reconstruct a surface normal vector field through photometric stereo vision to generate a three-dimensional gradient graph containing the inclination angle and azimuth angle of the vector.

[0017] The three-dimensional gradient graph and the parallax matrix are subjected to Poisson reconstruction to obtain an initial three-dimensional point cloud graph, and the outlying points are filtered and the holes are repaired to generate a three-dimensional point cloud graph.

[0018] As a preferred scheme of the automobile leather defect detection method based on visual detection, the multi-view automobile leather image comprises automobile leather images of orthographic view, oblique view and grazing view.

[0019] The three-dimensional point cloud map, the Gaussian curvature and the average curvature of the automobile leather surface are combined, differential geometry parameter mapping and curvature threshold segmentation are carried out, and a curvature distribution map is generated.

[0020] The curvature distribution map and the multi-view automobile leather image are subjected to wavelet multi-scale decomposition, and the texture energy distribution features of each frequency band are obtained.

[0021] The curvature distribution map and the texture energy distribution features are pixel-level weighted fusion to generate a defect probability heat map, and through adaptive threshold segmentation, a three-dimensional coordinate defect bounding box is output to mark the potential defect area.

[0022] As a preferred scheme of the automobile leather defect detection method based on visual detection, the three-dimensional morphology features of the potential defect area are extracted by the three-dimensional topography measurement method, and the three-dimensional geometric parameters of the defect are calculated.

[0023] Based on the marked potential defect area, a structured light three-dimensional topography measurement method is used for multi-view stripe projection and phase calculation to obtain three-dimensional point cloud data of the defect area.

[0024] The three-dimensional morphology features of the defect area three-dimensional point cloud data are extracted by a surface reconstruction algorithm.

[0025] The three-dimensional point cloud data of the defect area and the three-dimensional morphology features of the defect are fused and calculated in the spatial domain to obtain the three-dimensional geometric parameters of the defect.

[0026] As a preferred scheme of the automobile leather defect detection method based on visual detection, the three-dimensional morphology features of the potential defect area are extracted by the three-dimensional topography measurement method, and the three-dimensional geometric parameters of the defect are calculated.

[0027] The defect geometric features and spatial distribution rules of the three-dimensional geometric parameters of the defect are analyzed, and a preset grading judgment rule is used to divide the defect grade to generate an automobile leather quality evaluation report containing the three-dimensional coordinates of the defect.

[0028] Based on the three-dimensional geometric parameters of the defect and the three-dimensional point cloud data of the defect area, spatial statistical analysis is carried out to quantize the defect aggregation index.

[0029] According to the preset grading rule, the defect aggregation index is scored, and combined with the three-dimensional point cloud data of the defect area, an automobile leather quality evaluation report containing the three-dimensional coordinates of the defect is generated.

[0030] In a second aspect, the present application provides a vehicle leather defect detection system based on visual detection, comprising an imaging module, a reconstruction module, an identification module, an analysis module and a generation module; the imaging module is used to acquire multi-view vehicle leather images through multi-view synchronous shooting and generate clear multi-view vehicle leather images through dynamic curved surface focusing; the reconstruction module is used to reconstruct the three-dimensional topography of the leather surface by analyzing the parallax relationship and illumination direction reflection characteristics between the multi-view vehicle leather images and generate a three-dimensional point cloud map; the identification module is used to extract the curvature variation features of the vehicle leather surface of the three-dimensional point cloud map, perform multi-scale texture analysis, identify and mark potential defect areas; the analysis module is used to extract the three-dimensional morphological features of the potential defect areas through three-dimensional topography measurement method, calculate the three-dimensional geometric parameters of the defects; the generation module is used to analyze the defect geometric features and spatial distribution rules of the three-dimensional geometric parameters of the defects, divide the defect levels by using preset grading judgment rules, and generate a vehicle leather quality evaluation report containing three-dimensional coordinates of the defects.

[0031] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, any step of the vehicle leather defect detection method based on visual detection according to the first aspect of the present application is realized.

[0032] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, any step of the vehicle leather defect detection method based on visual detection according to the first aspect of the present application is realized.

[0033] The present application has the following beneficial effects: through dynamic curved surface focusing, clear multi-view vehicle leather images are generated, full field of view clear imaging of the vehicle leather curved surface is realized, optical focusing based on real-time curvature calculation solves the imaging defects of fixed focal length on curved surface materials, and provides a high-quality image data basis for subsequent three-dimensional reconstruction; combined with curvature-texture multi-scale fusion detection, the geometric features of the three-dimensional point cloud and the texture features of the two-dimensional image are deeply fused, the technical limitations of single modal detection are broken through, the recognition ability of complex surface defects is significantly enhanced, and the accuracy and efficiency of vehicle leather defect detection are improved. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0035] Fig. 1 A flow chart of a method for detecting defects in automotive leather based on visual inspection.

[0036] Fig. 2 A schematic diagram of a system for detecting defects in automotive leather based on visual inspection.

[0037] Fig. 3 A flow chart of generating a three-dimensional point cloud map.

[0038] Fig. 4 A flow chart of generating an automotive leather quality assessment report. DETAILED DESCRIPTION

[0039] In order to make the above objectives, features and advantages of the present application more apparent, specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0040] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details given herein, that the present application can be practiced with other than the described implementations, and that the present application can be practiced with or in conjunction with other systems, components, and / or methods. In other instances, well-known structures and / or operations are not shown or described in detail in order to avoid obscuring aspects of the present application.

[0041] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. The "in one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments.

[0042] REFERENCE Figs. 1-4 For one embodiment of the present application, the embodiment provides a method for detecting defects in automotive leather based on visual inspection, comprising the following steps:

[0043] S1, acquiring multi-view automotive leather images by multi-view synchronous shooting.

[0044] S1.1, the multi-view automotive leather images include automotive leather images of orthographic view, oblique view and grazing view.

[0045] Specifically, the binocular cameras are fixedly installed above the detection station at a normal angle of 90 degrees, an oblique angle of 45 degrees and a grazing angle of 15 degrees, respectively. The distance between the binocular cameras is exemplarily set to 300 mm. The optical axis of the binocular camera at the normal angle is perpendicular to the surface of the automotive leather, the optical axis of the binocular camera at the oblique angle forms an angle of 45 degrees with the normal, and the optical axis of the binocular camera at the grazing angle forms an angle of 15 degrees with the normal. The three binocular cameras synchronously collect images through hardware triggering, and the exposure time is exemplarily set to 10 ms. During the collection process, the automotive leather sample passes through the detection area at an exemplary speed of 50 mm / s, and the binocular cameras synchronously shoot at an exemplary frame rate of 30 fps, so as to obtain normal-angle images, oblique-angle images and grazing-angle images, and form multi-angle automotive leather images.

[0046] S2, dynamic curved surface focusing is performed to generate clear multi-angle automotive leather images.

[0047] S2.1, based on the multi-angle automotive leather images, the Gaussian curvature and average curvature distribution of the surface of the automotive leather are calculated, and the focusing of the binocular cameras is controlled in real time to obtain a set of locally optimal focused automotive leather images at different angles.

[0048] It should be noted that the expression for calculating the Gaussian curvature and average curvature distribution of the surface of the automotive leather based on the multi-angle automotive leather images is as follows:

[0049] ;

[0050] wherein, is the Gaussian curvature distribution of the surface of the automotive leather, is the stretching strength of the surface of the automotive leather in the first direction, is the bending strength of the surface of the automotive leather in the second direction, is the twisting degree of the surface of the automotive leather in the first and second directions, is the bending strength of the surface of the automotive leather in the first direction, is the stretching strength of the surface of the automotive leather in the second direction, is the coupling degree of the surface of the automotive leather in the first and second directions;

[0051] ;

[0052] wherein, is the average curvature distribution, is the gray change rate of the multi-angle automotive leather images in the first direction, is the mixed gray gradient of the multi-angle automotive leather images in the first and second directions, is the gray change rate of the multi-angle automotive leather images in the second direction;

[0053] Specifically, the grayscale gradient is calculated pixel-by-pixel for the input multi-view automotive leather image. The Sobel operator is used to obtain the grayscale change rate in the first and second directions of the multi-view automotive leather image. The grayscale gradient distribution characteristics are statistically analyzed within a fixed-size neighborhood window. The stretching strength of the automotive leather surface in the first and second directions, as well as the coupling degree between the two directions, are calculated. The local surface equation is fitted using the least squares method to solve for the bending strength of the automotive leather surface in the first and second directions, as well as the degree of twisting between the two directions, thus obtaining the surface feature parameters obtained from the multi-view automotive leather image.

[0054] Based on the surface feature parameters obtained from multi-view automotive leather images, a Gaussian curvature distribution reflecting the surface's unevenness and an average curvature distribution characterizing the overall degree of curvature are obtained through multi-view automotive leather image surface feature parameter mapping. A binocular camera drives a mechanical transmission device to adjust the lens position according to the curvature distribution results. During focusing, the camera monitors the changes in sharpness in local areas of the image in real time. When the sharpness reaches its optimal state, the focus is locked, and finally, a set of locally optimal focused automotive leather images from each viewpoint is output.

[0055] Furthermore, the process of determining the optimal sharpness is as follows: During the focusing process, the sharpness changes in local areas of the image are continuously monitored. The Tenengrad gradient function is used to calculate the gradient energy value of the multi-view car leather image in real time. When the gradient energy value of the multi-view car leather image reaches a peak of 15,000-20,000 (example) and begins to decrease, it is determined to be the optimal sharpness state. A control signal is immediately sent to lock the current focal length position and keep the current position of the binocular camera lens unchanged. The process is executed independently at each focusing position to ensure that the images from each perspective reach the optimal focus state.

[0056] It should be noted that the expression for calculating the gradient energy value of multi-view automotive leather images in real time using the Tenengrad gradient function is as follows:

[0057] ;

[0058] in, It is the gradient energy value of multi-view car leather images. It is a window traversal All pixel coordinates within, These are the column coordinates of pixels in a multi-view automotive leather image. These are the row coordinates of pixels in a multi-view car leather image. It is a multi-view grayscale image of car leather. It is a grayscale matrix of multi-view car leather images.

[0059] S2.2, the local optimal focusing automobile leather image set of each view angle is fused in frequency domain, and sub-pixel level geometric correction is performed through multi-view feature matching to generate a full-view clear multi-view automobile leather image set.

[0060] Specifically, the local optimal focusing automobile leather image of each view angle is converted to frequency domain through fast Fourier transform, the effective frequency band of the local optimal focusing automobile leather image of each view angle is extracted in frequency domain space using a Butterworth filter, the cutoff frequency of the Butterworth filter is set to an exemplary value of 0.6 times the Nyquist frequency, the frequency domain information of the Butterworth filter is fused by a weighted superposition method, the clarity weight coefficient is dynamically allocated according to the clarity evaluation result of the Butterworth filter, and the inverse Fourier transform is performed on the fused frequency domain data to reconstruct the spatial domain image.

[0061] Based on the reconstructed spatial domain image, not less than 500 matching feature points are extracted between the multi-view automobile leather images, a 128-dimensional descriptor of each feature point is calculated using a SIFT feature detection algorithm, an initial matching relationship is obtained through a nearest neighbor matching method, 4 pairs of matching points are randomly selected to obtain an initial homography matrix, the re-projection error of all matching points is tested using the initial homography matrix, and the matching points with an error less than an exemplary value of 1.5 pixels are determined as inliers and counted.

[0062] The random sampling is repeated for an exemplary value of 2000 times, the homography matrix with the largest number of inliers is retained as the optimal matching relationship, based on the optimal matching relationship, image registration with sub-pixel level accuracy is realized through bicubic interpolation, histogram matching processing is performed on the registered local optimal focusing automobile leather images of each view angle to unify the gray scale distribution of each image, and finally a full-view clear multi-view automobile leather image set is generated.

[0063] S2.3, the full-view clear multi-view automobile leather image set is subjected to polarization reflection component separation and light intensity compensation to generate an enhanced multi-view automobile leather image.

[0064] Specifically, a full-view clear multi-view automobile leather image set sequence with different polarization angles is obtained using a rotating polarizer, the diffuse reflection component and the specular reflection component are separated by a Stokes vector method, and a polarization degree threshold is set. The polarization degree greater than the polarization degree threshold is determined as a specular reflection dominant area and removed. The physical separation of the diffuse reflection component and the specular reflection component is realized; the diffuse reflection component is subjected to median filtering processing; a defect-free automobile leather area reference light intensity template is selected, and RGB three-channel reference values are extracted; the brightness mean value is analyzed in a local window, and when the deviation from the brightness mean value exceeds the set polarization degree threshold, the pixel brightness is adjusted, and the light intensity compensation of the full-view clear multi-view automobile leather image set is completed; the full-view clear multi-view automobile leather image set after light intensity compensation is subjected to gamma correction processing; the pixel values of the filtered diffuse reflection component multi-view automobile leather image set are subjected to three-stage mapping conversion, and finally an enhanced multi-view automobile leather image with improved defect signal-to-noise ratio is generated.

[0065] Further, the process of setting the polarization degree threshold: based on the Stokes vector method, the polarization degree numerical distribution of each pixel point is calculated, the polarization degree difference characteristics of the normal area and the defect area of the leather surface are analyzed, the critical value for distinguishing the specular reflection and the diffuse reflection component is determined by statistical histogram, and finally the polarization degree threshold is obtained.

[0066] S2.4, for the enhanced multi-view automobile leather image, disparity analysis and calculation of the depth distance of the automobile leather surface to the binocular camera are performed, and a clear focused multi-view automobile leather image is generated.

[0067] Specifically, SIFT feature points are extracted from the left and right view enhanced multi-view automobile leather images, the corresponding feature point relationship is obtained by a normalized cross-correlation matching method, and the matching window size is exemplarily 15x15 pixels; a semi-global stereo matching algorithm is used to obtain the disparity value, the disparity search range is exemplarily set to 64 levels, and the penalty coefficient is exemplarily set to P1=8 and P2=32; the disparity value is converted into a depth distance according to the baseline distance and the focal length parameters of the binocular camera, and the depth calculation accuracy is exemplarily controlled to be in the order of 0.1 mm; the enhanced multi-view automobile leather image fusion is guided based on the depth distance map, and the fusion weight is exemplarily distributed according to the Gaussian distribution (σ=5 pixels); the enhanced multi-view automobile leather image after fusion is subjected to non-sharpening mask processing, and the enhancement factor is exemplarily set to 0.5, and finally a clear focused multi-view automobile leather image is output;

[0068] It should be noted that the specific process of performing non-sharp mask processing on the fused enhanced multi-view automobile leather image is that a Gaussian filter is used to smooth the enhanced multi-view automobile leather image, the filter size is exemplarily set to 5*5 pixels, the high-frequency component containing the leather surface details is obtained through the difference operation of the original enhanced multi-view automobile leather image and the smoothed enhanced multi-view automobile leather image, and the original enhanced multi-view automobile leather image is superimposed back; the pixel value truncation processing is performed on the superimposition result, so that the output clear focus multi-view automobile leather image channel value is kept in the range of 0-255;

[0069] The expression for calculating the depth distance of the automobile leather surface to the binocular camera is:

[0070] ;

[0071] wherein, is the depth distance of the automobile leather surface to the binocular camera, is the baseline distance of the binocular camera, is the disparity value of the left binocular camera image and the right binocular camera image, is the horizontal pixel coordinate difference in the left binocular camera image and the right binocular camera image.

[0072] S3, by analyzing the disparity relationship between the multi-view automobile leather images and the light direction reflection characteristics, the three-dimensional topography of the leather surface is reconstructed, and a three-dimensional point cloud map is generated.

[0073] S3.1, based on the clear focus multi-view automobile leather image, the disparity value between the views is calculated through the SIFT feature matching algorithm, and a sub-pixel level precision disparity matrix is obtained.

[0074] It should be noted that the expression for calculating the disparity value between the views through the SIFT feature matching algorithm is:

[0075] ;

[0076] wherein, is the disparity value between the views, is the horizontal coordinate of the left view, is the horizontal coordinate of the right view, is the gray scale of the left view, is the gray scale of the right view;

[0077] ​Specifically, based on the clear focus of multi-view automotive leather image, SIFT feature points are extracted from the multi-view automotive leather image with clear focus of left and right views, a Gaussian difference pyramid is constructed and extreme points are located in the scale space, and an exemplary contrast of 0.04 and an exemplary edge value of 10 are set for SIFT feature point detection; 128-dimensional descriptors of the feature points are obtained and a main direction is assigned;

[0078] An initial left-right view matching relationship is obtained by using a bidirectional nearest neighbor matching method, an exemplary distance ratio of 0.8 is set, a RANSAC algorithm is executed, 4 pairs of left-right view matching points are randomly selected each time to calculate a fundamental matrix, the number of inliers satisfying a re-projection error of less than 1.5 pixels is counted, and after 2000 iterations, the fundamental matrix with the largest number of inliers is retained and the final left-right view matching point pairs are screened;

[0079] A 5x5 pixel neighborhood is extracted at the position of the left-right view matching points, the gray gradient information is calculated, and the sub-pixel level accurate coordinates are obtained by fitting a quadratic surface; the horizontal coordinate difference value is calculated by traversing all left-right view matching point pairs, and is mapped to a matrix with the same size as the left view, and the nearest neighbor interpolation is filled in the position without matching points;

[0080] A 3x3 pixel median filter is applied to smooth the disparity matrix, the corrected disparity jump region is detected, the mirror filling processing is performed on the boundary region, and the final optimized sub-pixel level accuracy disparity matrix is output.

[0081] It should be noted that the process of constructing the Gaussian difference pyramid is as follows: the input image is subjected to layer-by-layer Gaussian blur processing, an incremental Gaussian kernel parameter is used, and exemplary 5 different scale blur images are generated at each pyramid level; the blur images of adjacent scales are subtracted pixel by pixel to form a difference image group at each level; a multi-level pyramid structure is established by downsampling, and finally a Gaussian difference pyramid is obtained.

[0082] S3.2, the surface normal vector field is reconstructed by using the disparity matrix and the preset light direction through photometric stereo vision, and a three-dimensional gradient graph containing the inclination angle and azimuth angle of the vector is generated.

[0083] Specifically, the parallax matrix is converted into a depth map, and a depth value is obtained through a baseline distance and a focal length parameter of a binocular camera; a preset light direction vector is used, and an example value is X component 0.2, Y component 0.1, and Z component 0.9; a gray value of a corresponding pixel point is extracted from an automobile leather image collected under different light conditions; a surface normal vector is derived according to a relationship between pixel brightness change and light direction, and a least square optimization method is used for surface normal vector calculation; the surface normal vector is decomposed into two components of a dip angle and an azimuth angle, the dip angle represents an included angle between the normal vector and the vertical direction, and the azimuth angle represents a projection direction of the normal vector in the horizontal plane; dip angle and azimuth angle data are smoothed, and a 3*3 pixel window bilateral filter is exemplarily used; and finally, a double-channel three-dimensional gradient map composed of the dip angle and the azimuth angle is output.

[0084] Further, in the process of converting the parallax matrix into the depth map, according to a geometric principle of binocular stereo vision, a baseline distance (an example value is 100 mm) and a focal length parameter (an example value is 1200 pixels) of a binocular camera calibrated in advance are used, a depth value is obtained through a depth calculation formula, linear normalization processing is performed on the depth value, the depth value is mapped to a 0-255 gray scale range, and finally a depth map is generated.

[0085] It should be noted that the preset process of the light direction: in the light experiment, an array of point light sources is erected, a calibration ball or a calibration object with a known geometric shape is placed at the position of the automobile leather sample; images of the calibration object under different light sources are collected, the spatial direction vector of each light source is calculated through the positioning of the mirror reflection highlight points; the relative position parameters of the light source and the binocular camera are recorded, including the elevation angle and the azimuth angle; the elevation angle and the azimuth angle direction vector are normalized to a unit vector and stored as a light parameter, and an example main light direction vector is (0.2, 0.1, 0.9); the angle error is controlled within an example range of ±1°.

[0086] S3.3, Poisson reconstruction is performed on the three-dimensional gradient map and the parallax matrix to obtain an initial three-dimensional point cloud map, and outlier filtering and hole repair are performed to generate a three-dimensional point cloud map.

[0087] Specifically, the dip angle data of the three-dimensional gradient map is checked to ensure that the value is within the range of 0 to 90 degrees, the azimuth angle data is checked to ensure that the value is within the range of 0 to 360 degrees, and abnormal points are removed; an octree spatial index structure is created, the voxel resolution is set to an example 0.1 mm, the Poisson equation is discretized in three-dimensional space, and the conjugate gradient method is used for iterative solution, and the isosurface is extracted from the scalar field to generate a triangular mesh;

[0088] Perform statistical outlier removal, obtain the average distance of each coordinate point to the 30 nearest neighbor coordinate points to obtain the distance distribution histogram, identify the hole region with more than 20 closed boundary edges, generate the Delaunay triangular mesh based on the three-dimensional gradient map; according to the 0.83 times interval of the average point distance of the original three-dimensional gradient map, new sampling points are inserted uniformly inside the mesh, so that the total number of sampling points reaches 1.2 times of the original three-dimensional gradient map; define a vertex neighborhood with a radius of 2 mm and perform Laplace smoothing for 3 times with a displacement weight of 0.2; finally output the three-dimensional point cloud map of the car leather surface;

[0089] It should be noted that the process of creating an octree spatial index structure is as follows: by traversing the X / Y / Z coordinates of all points in the three-dimensional gradient map, the minimum and maximum values of the three dimensions X / Y / Z are extracted, and the six extreme point coordinates of X / Y / Z are used as boundaries to determine the space bounding box range. The cubic space is equally divided into eight sub-cubes along the X / Y / Z axis; for each sub-cube, the same division operation is recursively performed until the preset depth level or voxel resolution requirement is reached; a spatial index number is established for each final level voxel unit, and the included point cloud data pointer is recorded; the parent-child relationship pointer between levels is constructed to form a complete tree index structure, and finally the octree spatial index structure is obtained.

[0090] S4, extract the curvature variation features of the three-dimensional point cloud map of the car leather surface, and perform multi-scale texture analysis to identify and mark potential defect areas.

[0091] S4.1, combine the three-dimensional point cloud map, the Gaussian curvature and the average curvature of the car leather surface, and perform differential geometry parameter mapping and curvature threshold segmentation to generate a curvature distribution map.

[0092] Specifically, based on the Gaussian curvature and the average curvature of the car leather surface, the Gaussian curvature and the average curvature of the car leather surface are converted into visual color coding, wherein the Gaussian curvature is represented using the red-green color spectrum, and the average curvature is represented using the blue-yellow color spectrum; set the curvature threshold range to distinguish different feature areas, for example, the Gaussian curvature value in the range of -0.5 to +0.5 mm⁻² corresponds to different color intensities; identify continuous curvature feature areas by region growing method, for example, select points with absolute curvature value exceeding 0.3 mm⁻¹ as region growing starting points; superimpose and fuse the visual results of the Gaussian curvature and the average curvature of the car leather surface to form a composite image that comprehensively reflects the convex-concave features of the leather surface; finally output the curvature distribution map of the car leather surface containing color coding and feature area marking;

[0093] It should be noted that the process of setting the curvature threshold is: analyzing the Gaussian curvature and average curvature value distribution of all points on the surface of the automotive leather, generating a double curvature statistical histogram; using Otsu automatic threshold calculation method to separate the histogram mode, determining the critical interval of curvature distribution; finally output the curvature threshold.

[0094] S4.2, the curvature distribution graph and the multi-view automotive leather image are subjected to wavelet multi-scale decomposition, and the texture energy distribution characteristics of each frequency band are obtained.

[0095] Specifically, the curvature distribution graph of the concave-convex features of the automotive leather is spatially registered with the multi-view automotive leather image of the corresponding view to ensure accurate correspondence of pixel positions; a Daubechies wavelet basis function is selected, and an exemplary decomposition layer number is set to 3 layers; a two-dimensional discrete wavelet transform is performed to decompose LL low-frequency subbands and LH / HL / HH high-frequency subbands; texture features of the high-frequency subbands are extracted at each scale, and the low-frequency subbands are recursively decomposed; the feature response values of each frequency band are recorded, and after normalization, multi-scale feature descriptions are formed, and finally the texture energy distribution characteristics of each frequency band are obtained.

[0096] It should be noted that the specific process of performing two-dimensional discrete wavelet transform to decompose the subband is: performing one-dimensional wavelet filtering operation on the registered multi-view automotive leather image along the row and column directions respectively, performing low-pass and high-pass filter set processing on each row of the image by row filtering, and performing low-pass and high-pass filtering on the row filtering result again by column filtering; the results of row filtering and column filtering are cross combined to generate LL low-frequency subband (low-pass filtering in both row and column), LH subband (low-pass filtering in row and high-pass filtering in column), HL subband (high-pass filtering in row and low-pass filtering in column), and HH subband (high-pass filtering in both row and column); the size of each subband is reduced to 1 / 4 of the original multi-view automotive leather image, the LL subband retains the main contour information of the multi-view automotive leather image, the LH subband contains horizontal details, the HL subband contains vertical details, and the HH subband contains diagonal details.

[0097] S4.3, the curvature distribution graph and the texture energy distribution characteristics are pixel-level weighted fused to generate a defect probability heat map, and through adaptive threshold segmentation, a three-dimensional coordinate defect bounding box is output to mark the potential defect area.

[0098] Specifically, the curvature distribution graph of the convex and concave features of the automobile leather and the texture feature value of the texture energy distribution feature are normalized to unify the numerical range to the interval of 0-1; an exemplary curvature-texture fusion weight coefficient is set as the curvature feature of the convex and concave features of the automobile leather 0.6, and the texture energy distribution feature 0.4, pixel-level weighted summation is performed to generate a fusion feature map; Gaussian smoothing processing is performed on the fusion feature map, and the kernel size is exemplarily set to 5*5 pixels; the Otsu algorithm is used to automatically calculate the segmentation threshold, and an exemplary segmentation threshold of 0.45 is obtained; the regions exceeding the segmentation threshold are subjected to connected domain analysis, and the minimum defect area is exemplarily set to 25 square millimeters; the two-dimensional defect region is mapped back to the three-dimensional point cloud space, and the three-dimensional convex hull vertex coordinates of the defect region are extracted; an axial alignment three-dimensional bounding box is generated according to the vertex coordinates, and the boundary box expansion coefficient is exemplarily taken as 1.1 times; finally, the three-dimensional coordinate bounding box information of the marked potential defect region is output, and the potential defect region is marked.

[0099] S5, by a three-dimensional topographic measurement method, extracting the defect three-dimensional morphological features of the potential defect region, calculating the defect three-dimensional geometric parameters.

[0100] S5.1, based on the marked potential defect region, using a structured light three-dimensional topographic measurement method, performing multi-view fringe projection and phase unwrapping to obtain three-dimensional point cloud data of the defect region.

[0101] Specifically, a structured light projection device is erected within the defect region boundary box, and an exemplary 12-step phase-shifted sinusoidal fringe pattern is projected, and the fringe period is exemplarily set to 30 pixels; an industrial camera is synchronously triggered to collect deformed fringe images, and the resolution of the industrial camera is exemplarily 2048*2048 pixels; the collected deformed fringe images are subjected to phase unwrapping, and a four-step phase shift method is used to calculate the wrapped phase image; the phase is unfolded by a multi-frequency heterodyne method, and the frequency selection is exemplarily 1 / 60, 1 / 50, 1 / 40; the unfolded absolute phase and the pre-calibrated industrial camera-height mapping relationship are converted to obtain three-dimensional height information; the three-dimensional height information is converted into three-dimensional point cloud coordinates in combination with the internal and external parameters of the industrial camera; the defect region data collected by multiple views is subjected to ICP registration, and the registration error is exemplarily controlled within 0.05 millimeters; finally, the three-dimensional point cloud data of the defect region is output;

[0102] Further, the calibration process of the pre-calibrated industrial camera-height mapping relationship is as follows: a standard plane calibration board is used to perform precise displacement along the vertical direction in the measurement space with an exemplary step distance of 0.1 millimeters, a sinusoidal fringe is projected at each height position, and a deformed fringe image is collected; the four-step phase shift method is used to obtain the phase value corresponding to each height, and a cubic spline interpolation is used to fit a continuous mapping function, and the calibration accuracy reaches an exemplary 0.01 millimeters, and finally an industrial camera height-phase correspondence relationship is generated;

[0103] It should be noted that the process of ICP registration of the defect area data collected by multiple views is: finding the spatial corresponding points between the three-dimensional point clouds of the defect area at different angles, removing the abnormal corresponding relationship by RANSAC algorithm; optimizing the rigid transformation parameters to minimize the distance of the corresponding points, and gradually aligning the point clouds in the iteration process; setting an exemplary maximum number of iterations of 50 times and an exemplary distance of 0.1 millimeter; verifying that the registration accuracy meets the exemplary error requirement of 0.05 millimeter; finally obtaining the three-dimensional point cloud fused with the multi-view data, and completing the ICP registration.

[0104] S5.2, by a surface reconstruction algorithm, extracting the defect three-dimensional morphological features of the defect area three-dimensional point cloud data.

[0105] Specifically, the normal vector of the registered defect area three-dimensional point cloud is estimated by the moving least squares method, and the neighborhood radius is exemplarily set to 5 millimeters; the Poisson surface reconstruction algorithm is used to generate a triangular mesh surface, and the reconstruction depth is exemplarily set to 9 layers; the morphological feature parameters are extracted from the reconstructed surface, including the vertical distance from the defect depth measurement reference surface to the lowest point, the defect volume is exemplarily calculated by the number of occupied voxels, the surface area change rate is exemplarily compared with the surface area ratio of the defect area and the surrounding normal area, and the curvature mutation index is exemplarily counted as the proportion of Gaussian curvature points exceeding 0.25; the feature parameters are normalized by the min-max method to generate a defect morphological descriptor containing an exemplary 12-dimensional feature vector; and finally output the defect three-dimensional morphological features.

[0106] S5.3, spatial domain fusion calculation of the defect area three-dimensional point cloud data and the defect three-dimensional morphological features is performed to obtain the defect three-dimensional geometric parameters.

[0107] It should be noted that the expression of the spatial domain fusion calculation to obtain the defect three-dimensional geometric parameters is:

[0108] ;

[0109] wherein, is the defect three-dimensional geometric parameter, is the defect three-dimensional morphological feature weight, is the defect area three-dimensional point cloud data weight, is the defect three-dimensional morphological feature, is the defect area three-dimensional point cloud data;

[0110] Specifically, the three-dimensional morphological characteristics of the defects are normalized so that each three-dimensional morphological characteristic value of the defects is in the interval of 0-1; the three-dimensional point cloud data of the defect area is voxelized, and the voxel resolution is exemplarily set to 0.1 millimeter; a feature weight parameter is set, and the three-dimensional morphological feature weight of the defects is exemplarily taken as 0.6, and the three-dimensional point cloud data weight of the defect area is exemplarily taken as 0.4; the three-dimensional point cloud data of the defect area and the three-dimensional morphological characteristics of the defects are subjected to a weighted fusion operation in the spatial domain, and the defect morphological characteristic value and the point cloud density value at the corresponding position in each voxel unit are combined; the fused three-dimensional morphological characteristics of the defects are extracted, including the maximum depth, the volume deviation, and the curvature integral value of the defects; the parameter consistency is verified, and the spatial position error is exemplarily controlled within 0.05 millimeters; and finally, the three-dimensional geometric parameters of the defects containing the size accuracy and the shape accuracy are output.

[0111] S6. The three-dimensional geometric parameters of the defects are analyzed to obtain the defect geometric characteristics and the spatial distribution law, and a preset grading rule is used to divide the defect levels to generate an automobile leather quality evaluation report containing the three-dimensional coordinates of the defects.

[0112] S6.1. Spatial statistical analysis is performed based on the three-dimensional geometric parameters of the defects and the three-dimensional point cloud data of the defect area to quantize the defect aggregation index.

[0113] Specifically, the three-dimensional point cloud data of the defect area is divided into a cubic grid unit exemplarily with a size of 5mmx5mmx5mm; the number and intensity value of the three-dimensional geometric parameters of the defects in each cubic grid unit are recorded; Moran's I index is applied to adjacent grid units to analyze the spatial autocorrelation; the spatial distribution pattern of the defect parameters is identified; an exemplary aggregation threshold of 0.7 is used to distinguish the significantly aggregated area; the volume proportion and the average aggregation intensity of the aggregated area are counted; and finally, the defect aggregation index is output.

[0114] Further, the process of applying Moran's I index to analyze the spatial autocorrelation of adjacent grid units: a distance inverse weight function is used to describe the spatial relationship between adjacent grid units, and the neighborhood radius is exemplarily set to 15 millimeters; the defect intensity values of each adjacent grid unit are standardized; the parameter similarity of each adjacent grid unit and its neighborhood grid units is analyzed; the Moran's I index result varies in the interval of -1 to 1, a positive value reflects the defect aggregation trend, and a negative value reflects the discrete distribution characteristics; the reliability of the spatial pattern is evaluated through significance test; and an analysis result containing the spatial autocorrelation degree and the distribution characteristics is generated.

[0115] It should be noted that the setting process of the aggregation threshold value: analyze the numerical distribution of defect aggregation in historical detection data, draw the aggregation probability density curve; observe the inflection point position of the curve to determine the mutation interval of the aggregation value; combine the actual requirements of defect sensitivity in specific application scenarios; through statistical distribution characteristics and engineering experience verification, finally obtain the aggregation threshold value.

[0116] S6.2, according to the preset grading rule, scoring the defect aggregation index, and combining the three-dimensional point cloud data of the defect area, generating an automobile leather quality evaluation report containing the three-dimensional coordinates of the defect.

[0117] Specifically, the defect aggregation index is divided into 5 levels according to the preset grading rule, 0-0.3 is A level, 0.3-0.5 is B level, 0.5-0.7 is C level, 0.7-0.9 is D level, and 0.9-1.0 is E level; The spatial coordinate information of the three-dimensional point cloud data of each defect area is extracted, including the X / Y / Z axis coordinate range and the center of mass position; According to the grade division result, mark the quality score of each defect area; Integrate the defect three-dimensional geometric parameters, aggregation score and coordinate information; Generate a structured report document containing defect distribution diagram, parameter statistical table and coordinate details; Finally output the automobile leather quality evaluation report containing the three-dimensional coordinates of the defect;

[0118] It should be noted that the preset process of the grading rule is: analyze the numerical distribution characteristics of the defect aggregation index in the historical detection data, and count the frequency of different interval values; Observe the influence degree of different numerical intervals on product pass rate in actual production environment; Verify the rationality of each level interval through batch test data; Finally determine the defect aggregation grading rule containing 5 levels (0-0.3 is A level, 0.3-0.5 is B level, 0.5-0.7 is C level, 0.7-0.9 is D level, and 0.9-1.0 is E level).

[0119] The embodiment also provides an automobile leather defect detection system based on visual detection, comprising: an imaging module, a reconstruction module, an identification module, an analysis module and a generation module.

[0120] The imaging module is used for acquiring multi-view automobile leather images through multi-view synchronous shooting and generating clear multi-view automobile leather images through dynamic curved surface focusing.

[0121] The reconstruction module is used for reconstructing the three-dimensional topography of the leather surface by analyzing the parallax relationship and light direction reflection characteristics between the multi-view automobile leather images, and generating a three-dimensional point cloud map.

[0122] The identification module is used for extracting the curvature variation characteristics of the automobile leather surface of the three-dimensional point cloud map, and performing multi-scale texture analysis to identify and mark potential defect areas.

[0123] an analysis module configured to extract a three-dimensional morphological feature of a potential defect area by a three-dimensional topography measurement method, and calculate a three-dimensional geometric parameter of the defect;

[0124] a generation module configured to divide a defect level by analyzing a defect geometric feature and a spatial distribution rule of the three-dimensional geometric parameter of the defect, and using a preset grading judgment rule, and generate an automobile leather quality evaluation report containing a three-dimensional coordinate of the defect.

[0125] The embodiment also provides a computer device suitable for the automobile leather defect detection method based on visual detection, which comprises a memory and a processor.

[0126] The computer device can be a terminal, and the computer device comprises a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, a trackball or a touchpad arranged on the shell of the computer device. In addition, the input device can be an external keyboard, a touchpad or a mouse, etc.

[0127] The embodiment also provides a storage medium on which a computer program is stored, the computer program being executed by a processor to implement the method for detecting automobile leather defects based on visual detection proposed in the above embodiment; and the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk.

[0128] In summary, the present application achieves full field clear imaging of automobile leather surface by generating clear focused multi-view automobile leather images through dynamic curved surface focusing, and solves the imaging defects of fixed focus on curved surface materials based on real-time curvature calculation optical focusing, thereby providing high-quality image data basis for subsequent three-dimensional reconstruction; and the curvature-texture multi-scale fusion detection is combined to deeply fuse the geometric features of three-dimensional point cloud and the texture features of two-dimensional image, thereby breaking through the technical limitations of single modal detection, significantly enhancing the recognition ability of complex surface defects, and achieving the improvement of accuracy and efficiency of automobile leather defect detection.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A method for detecting defects in automotive leather based on visual inspection, characterized in that: The application relates to a method for evaluating the quality of automobile leather. Multi-view automobile leather images are acquired through multi-view synchronous shooting, and dynamic curved surface focusing is performed to generate clear multi-view automobile leather images; Through analysis of the parallax relationship and illumination direction reflection characteristics among the multi-view automobile leather images, the three-dimensional topography of the leather surface is reconstructed to generate a three-dimensional point cloud map, and the specific steps are as follows, Based on the clear multi-view automobile leather images, the parallax values among the views are calculated through a SIFT feature matching algorithm to obtain a sub-pixel level parallax matrix; The parallax matrix is combined with a preset illumination direction to reconstruct the surface normal vector field through photometric stereo vision to generate a three-dimensional gradient map containing the inclination angle and azimuth angle of the vector; The three-dimensional gradient map and the parallax matrix are subjected to Poisson reconstruction to obtain an initial three-dimensional point cloud map, and the initial three-dimensional point cloud map is subjected to outlier filtering and hole repairing to generate a three-dimensional point cloud map; The curvature variation features of the automobile leather surface of the three-dimensional point cloud map are extracted, and multi-scale texture analysis is performed to identify and mark potential defect regions; Through a three-dimensional topography measurement method, the defect three-dimensional morphological features of the potential defect regions are extracted, and the defect three-dimensional geometric parameters are calculated. The defect geometric features and spatial distribution rules of the defect three-dimensional geometric parameters are analyzed, preset grading judgment rules are used to divide the defect grades, and an automobile leather quality evaluation report containing the three-dimensional coordinates of the defects is generated.

2. The visual inspection based automotive leather defect detection method as claimed in claim 1, wherein: The multi-view automobile leather images include automobile leather images in orthographic, oblique and grazing views.

3. The visual inspection based automotive leather defect detection method as claimed in claim 2, wherein: The dynamic curved surface focusing is performed to generate clear multi-view automobile leather images, and the specific steps are as follows, Based on the multi-view automobile leather images, the Gaussian curvature and average curvature distribution of the automobile leather surface are calculated, and the focusing of the binocular camera is controlled in real time to obtain a set of locally optimal focusing automobile leather images in each view; The set of locally optimal focusing automobile leather images in each view is subjected to frequency domain fusion, and sub-pixel level geometric correction is performed through multi-view feature matching to generate a set of clear multi-view automobile leather images in the whole field of view; The set of clear multi-view automobile leather images in the whole field of view is subjected to polarization reflection component separation and light intensity compensation to generate enhanced multi-view automobile leather images; The enhanced multi-view automobile leather images are subjected to parallax analysis and calculation of the depth distance of the automobile leather surface to the binocular camera to generate clear multi-view automobile leather images.

4. The visual inspection based automotive leather defect detection method as claimed in claim 1, wherein: The curvature variation features of the automobile leather surface of the three-dimensional point cloud map are extracted, and multi-scale texture analysis is performed to identify and mark potential defect regions, and the specific steps are as follows, The three-dimensional point cloud map, the Gaussian curvature and the average curvature of the automobile leather surface are combined, and differential geometric parameter mapping and curvature threshold segmentation are performed to generate a curvature distribution map; The curvature distribution map and the multi-view automobile leather images are subjected to wavelet multi-scale decomposition to obtain the texture energy distribution features in each frequency band; The curvature distribution map and the texture energy distribution features are subjected to pixel level weighted fusion to generate a defect probability heat map, and a three-dimensional coordinate defect bounding box is output through adaptive threshold segmentation to mark the potential defect regions.

5. The visual inspection based automotive leather defect detection method as claimed in claim 4, wherein: The defect three-dimensional morphological features of the potential defect regions are extracted through a three-dimensional topography measurement method, and the defect three-dimensional geometric parameters are calculated, and the specific steps are as follows, Based on the marked potential defect area, a structured light three-dimensional topography measurement method is adopted to perform multi-view fringe projection and phase calculation to obtain three-dimensional point cloud data of the defect area; Through a curved surface reconstruction algorithm, three-dimensional morphological features of the defect area three-dimensional point cloud data are extracted; The defect area three-dimensional point cloud data and the defect three-dimensional morphological features are fused and calculated in a spatial domain to obtain three-dimensional geometric parameters of the defect.

6. The visual inspection based automotive leather defect detection method as claimed in claim 5, wherein: The defect geometric features and spatial distribution rules of the defect three-dimensional geometric parameters are analyzed, and a preset grading rule is used to divide the defect grade to generate an automobile leather quality evaluation report containing three-dimensional coordinates of the defect, and the specific steps are as follows, Based on the defect three-dimensional geometric parameters and the defect area three-dimensional point cloud data, spatial statistical analysis is performed to quantize the defect aggregation index; According to the preset grading rule, the defect aggregation index is scored, and combined with the defect area three-dimensional point cloud data, an automobile leather quality evaluation report containing three-dimensional coordinates of the defect is generated.

7. A visual inspection based automotive leather defect detection system based on the visual inspection based automotive leather defect detection method according to any one of claims 1 to 6, characterized in that: It includes an imaging module, a reconstruction module, an identification module, an analysis module and a generation module; The imaging module is used to acquire multi-view automobile leather images by multi-view synchronous shooting and generate clear and focused multi-view automobile leather images by dynamic curved surface focusing; The reconstruction module is used to reconstruct the three-dimensional topography of the leather surface by analyzing the parallax relationship and light direction reflection characteristics between the multi-view automobile leather images to generate a three-dimensional point cloud map; The identification module is used to extract the curvature variation features of the automobile leather surface of the three-dimensional point cloud map and perform multi-scale texture analysis to identify and mark the potential defect area; The analysis module is used to extract the defect three-dimensional morphological features of the potential defect area by a three-dimensional topography measurement method and calculate the defect three-dimensional geometric parameters; The generation module is used to analyze the defect geometric features and spatial distribution rules of the defect three-dimensional geometric parameters, and use a preset grading rule to divide the defect grade to generate an automobile leather quality evaluation report containing three-dimensional coordinates of the defect.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor executes the computer program to realize the steps of the automobile leather defect detection method based on visual detection according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the automobile leather defect detection method based on visual detection according to any one of claims 1-6.

Images