Vehicle PVC decal defect detection method and system based on AI vision

Abstract

The application discloses a vehicle PVC decal defect detection method and system based on AI vision, which comprises the following steps: acquiring test image data of a vehicle decal; adjusting camera parameters to obtain a first image sequence with uniform brightness distribution in the case that the test image data has an overexposed area; the overexposed area is an area with a brightness value greater than a brightness threshold; performing image optimization processing on the first image sequence to obtain a second image sequence; the image optimization processing comprises at least one of the following: illumination compensation processing, denoising processing and contrast optimization processing; performing feature capture on potential low-complexity areas in the second image sequence through convolution operation to generate a feature vector set containing decal edges and surface textures; and inputting the feature vector set into a trained defect classification model to obtain a defect detection result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicles, in particular to a vehicle PVC decal defect detection method and system based on AI vision. BACKGROUND

[0002] As a key element to improve product aesthetics and brand recognition, the stability and consistency of the adhesion quality of PVC decals in vehicle appearance decoration have become an important factor affecting the overall vehicle quality. With the continuous improvement of consumers' requirements for the quality of vehicle appearance, the accuracy and efficiency of decal defect detection are also increasingly strict.

[0003] Traditional manual detection methods have obvious shortcomings when faced with mass production, and detection personnel are prone to missed detection or misdiagnosis due to fatigue, and the subjectivity of detection standards makes quality control lack uniformity. Although existing machine vision detection technology has improved detection efficiency to some extent, it has poor adaptability under complex lighting conditions and limited recognition ability for subtle defects, making it difficult to meet the demand for high-precision detection.

[0004] The core challenge of vehicle decal defect detection is caused by unstable image quality due to changes in lighting and body surface reflection during image acquisition. The uneven distribution of light sources in the workshop environment, combined with the specular reflection characteristics of the metal surface of the vehicle body, often causes the collected decal images to appear locally overexposed or shadowed. This fluctuation in image quality directly affects the accuracy of subsequent feature extraction. When image preprocessing cannot effectively eliminate these interference factors, deep learning models will be affected by noise information during feature learning, resulting in a lack of representativeness and stability of the extracted image features. The decline in feature quality further restricts the classification performance of the visual recognition network, making it difficult for the system to accurately distinguish between normal decals and various defect patterns, ultimately affecting the reliability and practicality of the entire detection system.

[0005] Therefore, how to build a complete detection method that can stably collect vehicle decal images and perform effective preprocessing, accurately extract image features using deep learning technology, and achieve high-precision defect recognition and classification through an optimized visual network has become a key problem in solving vehicle decal quality control. SUMMARY

[0006] The present application provides a vehicle PVC decal defect detection method based on AI vision to improve the reliability of decal quality detection and enable timely early warning and processing of abnormal water sources.

[0007] In a first aspect, to solve the above technical problems, the present application provides an AI vision-based vehicle PVC decal defect detection method, comprising: obtaining test image data of a vehicle decal; in the case that the test image data has an overexposed area, adjusting camera parameters to obtain a first image sequence with uniform brightness distribution; the overexposed area is an area with a brightness value greater than a brightness threshold; performing image optimization processing on the first image sequence to obtain a second image sequence; the image optimization processing includes at least one of the following: illumination compensation processing, noise reduction processing, and contrast optimization processing; capturing features of potential low complexity areas in the second image sequence through convolution operation to generate a feature vector set containing decal edges and surface textures; the potential low complexity area is an area with a texture complexity index lower than a texture complexity threshold; the texture complexity index is used to represent the complexity of the texture in the second image sequence; inputting the feature vector set into a trained defect classification model to obtain a defect detection result.

[0008] Preferably, the feature vector set containing the decal edges and the surface textures is generated by capturing features of the potential low complexity areas through the convolution operation, comprising: determining whether there is an area in the second image sequence with a contrast enhancement coefficient standard deviation lower than a standard deviation threshold; the contrast enhancement coefficient is a coefficient for modifying the contrast of different areas in the second image sequence; in the case that there is an area with a contrast enhancement coefficient standard deviation lower than a standard deviation threshold, adjusting the contrast enhancement coefficient and re-calculating the contrast enhancement coefficient standard deviation around each pixel point; iteratively performing the above steps until there is no area in the second image sequence with a contrast enhancement coefficient standard deviation lower than a standard deviation threshold, to obtain an optimized second image sequence; capturing features of the potential low complexity areas in the optimized second image sequence through the convolution operation to generate the feature vector set containing the decal edges and the surface textures.

[0009] Preferably, the image optimization processing is performed on the first image sequence to obtain the second image sequence, comprising: using a preprocessing algorithm based on histogram equalization and Gaussian filtering to perform illumination compensation and noise suppression processing on the first image sequence; calculating gradient information of the first image sequence after the illumination compensation and noise suppression processing; the gradient information includes the gradient value of each pixel point; determining the contrast enhancement coefficient of each pixel point according to a first mapping relationship and the gradient value of each pixel point, the first mapping relationship including the mapping relationship between different gradient values and different contrast enhancement coefficients; performing contrast optimization on each pixel point in the first image sequence after the illumination compensation and noise suppression processing according to the contrast enhancement coefficient of each pixel point to obtain the second image sequence.

[0010] Preferably, a preprocessing algorithm based on histogram equalization and Gaussian filtering is adopted to perform illumination compensation and noise suppression on the first image sequence, including: counting the number of pixels of each brightness level in the first image sequence, and constructing a histogram of the first image sequence; adjusting the brightness value of each pixel point according to the distribution of the histogram to obtain the first image sequence after illumination compensation; and performing convolution operation on the first image sequence after illumination compensation based on a Gaussian kernel of a preset size and a preset standard deviation to obtain the first image sequence after illumination compensation and noise suppression.

[0011] Preferably, the gradient information of the first image sequence after illumination compensation and noise suppression is calculated, including: for any pixel point in the first image sequence, determining the gradient value of the pixel point in the horizontal direction and the gradient value of the pixel point in the vertical direction; calculating the gradient value of the pixel point according to the gradient value of the pixel point in the horizontal direction, the gradient value of the pixel point in the vertical direction, and a gradient determination formula to obtain the gradient information of the first image sequence after illumination compensation and noise suppression; wherein the gradient determination formula satisfies the following relationship:

[0012] ;

[0013] denotes the gradient value of the pixel point, denotes a square root function, denotes the gradient value in the horizontal direction, denotes the gradient value in the vertical direction.

[0014] Preferably, the contrast of each pixel point in the first image sequence after illumination compensation and noise suppression is optimized according to the contrast enhancement coefficient of each pixel point to obtain a second image sequence, including: for each pixel point, determining the target brightness value of the pixel point according to the contrast enhancement coefficient and the original brightness value, the target brightness value being the product of the contrast enhancement coefficient and the original brightness value; and adjusting the brightness value of the pixel point to the target brightness value to obtain the second image sequence.

[0015] Preferably, the method further includes: performing convolution operation on the pixel points in the second image sequence through the feature extraction layer of the convolutional neural network, and generating multi-scale feature maps using convolution kernels of different sizes to obtain the texture complexity index of each region.

[0016] Preferably, the feature vector set is input into the trained defect classification model to obtain a defect detection result, including: according to variance distribution data in the feature vector set, contribution value of each dimension data is sorted, and dimension data with a ranking greater than a ranking threshold is determined as a key indicator; a data compression method is used to reduce the dimension of the vector mapping set corresponding to the key indicator to obtain a low-dimensional representation result; for the low-dimensional representation result, the low-dimensional representation result is integrated by a matrix optimization tool to obtain a final data form consistent with a matrix optimization target, and an optimized low-dimensional feature representation matrix is obtained; and the low-dimensional feature representation matrix is input into the trained defect classification model to obtain the defect detection result.

[0017] In a second aspect, the present application provides a vehicle PVC decal defect detection system, comprising:

[0018] An acquisition unit is configured to acquire test image data of a vehicle decal.

[0019] A processing unit is configured to adjust camera parameters to obtain a first image sequence with uniform brightness distribution when the test image data has an overexposed area, wherein the overexposed area is an area with a brightness value greater than a brightness threshold.

[0020] The processing unit is further configured to perform image optimization processing on the first image sequence to obtain a second image sequence, wherein the image optimization processing includes at least one of the following: illumination compensation processing, noise reduction processing, and contrast optimization processing.

[0021] The processing unit is further configured to capture features of potential low complexity areas in the second image sequence by convolution operation to generate a feature vector set containing decal edges and surface textures, wherein the potential low complexity area is an area with a texture complexity index lower than a texture complexity threshold, and the texture complexity index is used to represent the complexity of the texture in the second image sequence.

[0022] The processing unit is further configured to input the feature vector set into a trained defect classification model to obtain a defect detection result.

[0023] In a third aspect, the present application further provides a computer readable storage medium, which comprises a stored computer program, wherein when the computer program runs, the computer readable storage medium controls the device where the computer readable storage medium is located to execute the vehicle PVC decal defect detection method of any one of the above.

[0024] Compared with the prior art, the present application has the following beneficial effects:

[0025] Adaptive exposure control and image preprocessing algorithm are used to optimize image quality, convolutional neural network is used to extract multi-scale features, and finer granularity feature capture is used for potential low complexity areas. A high-differentiation defect recognition model is constructed through dimension reduction and classification algorithm to realize accurate detection and positioning of defects such as bubbles, wrinkles and shedding. The present application solves the problems of unstable image quality, difficulty in feature extraction in low contrast areas and low accuracy of defect type recognition in vehicle decal detection, improves the automation level and reliability of decal quality detection, and provides an efficient quality control means for automobile manufacturing industry. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a kind of vehicle PVC decal defect detection method flow chart based on AI vision provided by the embodiment of the present application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0028] To solve the above problems, the following specific embodiments will be described in detail and explained.

[0029] Reference Figure 1 The first embodiment of the present application provides a kind of vehicle PVC decal defect detection method based on AI vision, comprising the following steps:

[0030] S1, obtain the test image data of vehicle decal.

[0031] As a possible implementation mode, the defect detection system can adjust the brightness and angle of the multi-angle LED light source array according to the preset light source parameters to ensure uniform illumination of the vehicle decal area to obtain the test image data of the vehicle decal.

[0032] For example, in a vehicle decal detection system, first adjust the brightness and angle of the multi-angle LED light source array according to the preset light source parameters. Suppose the light source parameters include brightness level (such as 0-10 levels) and angle range (such as 0°-90°).

[0033] Specifically, the brightness can be set to level 5 and the angle can be adjusted to 45° to ensure uniform illumination of the vehicle decal area.

[0034] S2, in the case that the test image data has an overexposed area, adjusting the camera parameters to obtain a first image sequence with uniform brightness distribution.

[0035] The overexposed area is an area with a brightness value greater than a brightness threshold. The camera parameters include at least one of the following: camera light source brightness, camera light source angle, shutter speed, and aperture size.

[0036] As a possible implementation, the defect detection system can analyze the brightness distribution of the test image data through a brightness detection algorithm to determine whether there is an overexposed area with a local brightness value exceeding a preset threshold. If an overexposed area is detected, its position and range are marked. If there is an overexposed area, an adaptive exposure control module is started to dynamically adjust the exposure parameters of the camera, such as shutter speed and aperture size, according to the brightness value and range of the overexposed area, until the brightness of the overexposed area falls below the preset threshold, obtaining a first image sequence with uniform brightness distribution.

[0037] After obtaining the test image data, a brightness detection algorithm is used to analyze the brightness distribution of the image. Assuming that the preset threshold is 200 (the brightness value range is 0-255), the algorithm traverses each pixel point of the image to count the area with a brightness value exceeding 200. If an overexposed area is detected, its position and range are marked.

[0038] For example, it is detected that the brightness value in a 100x100 pixel area in the upper left corner of the image generally exceeds 200, and this area is marked as an overexposed area. If there is an overexposed area, an adaptive exposure control module is started. According to the brightness value and range of the overexposed area, the exposure parameters of the camera are dynamically adjusted. Assuming that the initial shutter speed is 1 / 100 seconds and the aperture size is F5.6, after detecting the overexposed area, the shutter speed is adjusted to 1 / 200 seconds and the aperture is reduced to F8 until the brightness of the overexposed area falls below 200. Through the above adjustment, a first image sequence with uniform brightness distribution is obtained.

[0039] It should be noted that dynamically adjusting the exposure parameters not only effectively avoids overexposure, but also improves the overall quality of the image, ensuring the accuracy of subsequent image processing and analysis.

[0040] In one embodiment, assuming that the vehicle decal area includes multiple colors and patterns, different colors have different reflectivities to light. Through the adjustment of the multi-angle LED light source array, appropriate illumination can be ensured for each color area, avoiding color distortion caused by uneven light.

[0041] For example, dark decal areas may require higher brightness levels, while light-colored areas require lower brightness to avoid overexposure.

[0042] Preferably, the system can also adjust the camera exposure parameters in real time according to changes in ambient light.

[0043] For example, in an outdoor environment with strong light, increasing the brightness of the light source and adjusting the camera shutter speed ensures that the image quality is not affected by the ambient light.

[0044] In one embodiment, through multiple experiments, it is found that setting the light source angle to 45° can effectively reduce the reflection of the decal surface and improve the clarity of the image. This facilitates subsequent image analysis and recognition.

[0045] It can be understood that this adaptive exposure control method is not only applicable to vehicle decal detection, but also can be extended to other scenarios that require high-precision image acquisition, such as license plate recognition and vehicle defect detection. By precisely controlling the light source and camera parameters, the image quality is ensured, and the detection accuracy and stability of the system are improved. Through the above measures, the system can always maintain uniform illumination and high-quality images under different lighting conditions, avoiding image information loss due to overexposure or underexposure, and improving the overall detection efficiency and accuracy.

[0046] S3, performing image optimization processing on the first image sequence to obtain a second image sequence.

[0047] The image optimization processing includes at least one of the following: illumination compensation processing, noise reduction processing, and contrast optimization processing.

[0048] As a possible implementation manner, the defect detection system can perform image optimization processing on the first image sequence based on the following steps to obtain a second image sequence:

[0049] S31, using a preprocessing algorithm based on histogram equalization and Gaussian filtering to perform illumination compensation and noise suppression processing on the first image sequence.

[0050] As a possible implementation manner, the defect detection system can count the number of pixels of each brightness level in the first image sequence to construct a histogram of the first image sequence; according to the distribution of the histogram, adjust the brightness value of each pixel point to obtain the first image sequence after illumination compensation; further, based on a Gaussian kernel of a preset size and a preset standard deviation, perform convolution operation on the first image sequence after illumination compensation to obtain the first image sequence after illumination compensation and noise suppression processing.

[0051] For example, assuming that the number of pixels in the dark area of the original image is relatively large, and the number of pixels in the bright area is relatively small, the histogram equalization will increase the brightness of the pixels in the dark area and decrease the brightness of the pixels in the bright area, so that the brightness distribution of the overall image tends to be uniform, and the first image sequence after illumination compensation is obtained.

[0052] For example, a 3*3 Gaussian kernel is selected, the standard deviation is set to 1.5, and the first image sequence after illumination compensation is convolved to smooth the high-frequency noise in the image, thereby obtaining the first image sequence after illumination compensation and noise suppression processing.

[0053] S32, gradient information of the first image sequence after illumination compensation and noise suppression processing is calculated.

[0054] The gradient information includes the gradient value of each pixel point.

[0055] As a possible implementation, for any pixel point in the first image sequence, the defect detection system can determine the gradient value of the pixel point in the horizontal direction and the gradient value in the vertical direction, and calculate the gradient value of the pixel point according to the gradient value of the pixel point in the horizontal direction, the gradient value in the vertical direction, and the gradient determination formula, thereby obtaining the gradient information of the first image sequence after illumination compensation and noise suppression processing.

[0056] The gradient determination formula satisfies the following relationship:

[0057] ; The gradient value of the pixel point is represented by g(x, y), The square root function is represented by sqrt(), The gradient value in the horizontal direction is represented by g x, The gradient value in the vertical direction is represented by g y.

[0058] S33, according to the first mapping relationship and the gradient value of each pixel point, the contrast enhancement coefficient of each pixel point is determined.

[0059] The first mapping relationship includes the mapping relationship between different gradient values and different contrast enhancement coefficients.

[0060] Specifically, the area with a larger gradient value represents the detail part of the image, and the contrast needs to be enhanced; the area with a smaller gradient value represents the smooth part, and the contrast needs to be weakened.

[0061] For example, a gradient threshold T is set, when the gradient value of the pixel point is greater than T, the enhancement coefficient is set to 1.2; when the gradient value is less than T, the enhancement coefficient is set to 0.8. Finally, the local contrast enhancement coefficient calculated is applied to the denoised image to perform contrast optimization processing. By adjusting the brightness value pixel by pixel, it is ensured that the detail information of the highlight and dark area in the image is effectively enhanced.

[0062] S34, according to the contrast enhancement coefficient of each pixel point, the contrast of each pixel point in the first image sequence after illumination compensation and noise suppression processing is optimized, thereby obtaining the second image sequence.

[0063] As a possible implementation, for each pixel point, a target brightness value of the pixel point is determined according to the contrast enhancement coefficient and the original brightness value, the target brightness value being a product of the contrast enhancement coefficient and the original brightness value, the brightness value of the pixel point is adjusted to the target brightness value, and a second image sequence is obtained.

[0064] For example, for a pixel point in a highlight area, the original brightness value is 250, and the enhancement coefficient is 1.2, so the adjusted brightness value is 250*1.2=300, but due to the range limitation of the brightness value, the final brightness value is 255. For a pixel point in a dark area, the original brightness value is 50, and the enhancement coefficient is 0.8, so the adjusted brightness value is 50*0.8=40.

[0065] Through the above processing, the detail part of the image is clearer, and the detail information of the highlight and dark areas is effectively enhanced, and the finally obtained second image sequence is more balanced and clear in visual effect. This processing method is particularly important in the vehicle decal detection system, because the decal area may contain various colors and patterns, and the enhancement of the detail information is helpful for subsequent image analysis and recognition.

[0066] In an embodiment, it is assumed that the vehicle decal area includes dark and light parts, the dark part has a low reflectivity to light, and the light part has a high reflectivity to light. Through the histogram equalization processing, the brightness of the dark part can be improved, and the brightness of the light part can be suppressed, so as to avoid color distortion caused by uneven light.

[0067] For example, the original brightness value of the dark part is 30, and after the equalization processing, it is improved to 60; the original brightness value of the light part is 220, and after the equalization processing, it is reduced to 180.

[0068] S4, performing feature capture on the potential low complexity area in the second image sequence through a convolution operation to generate a feature vector set containing decal edges and surface textures.

[0069] The potential low complexity area is an area with a texture complexity index lower than a texture complexity threshold; the texture complexity index is used to represent the complexity of the texture in the second image sequence.

[0070] In an example, S4 can include the following steps:

[0071] S41, determining whether there is an area in the second image sequence with a contrast enhancement coefficient standard deviation lower than a standard deviation threshold.

[0072] The contrast enhancement coefficient is a coefficient for modifying the contrast of different areas in the second image sequence.

[0073] S42, in the presence of a contrast enhancement coefficient standard deviation below the standard deviation threshold, adjusting the contrast enhancement coefficient, and re-counting the standard deviation of the contrast enhancement coefficient around each pixel point.

[0074] S43, iteratively performing the above steps until there is no region in the second image sequence with a contrast enhancement coefficient standard deviation below the standard deviation threshold, obtaining an optimized second image sequence.

[0075] In one possible implementation, extracting local contrast enhancement coefficient distribution data from the second image sequence, first need to count the contrast enhancement coefficient around each pixel point.

[0076] Specifically, assuming a certain pixel point P, the contrast enhancement coefficients around it are a1, a2, a3,..., an, calculate the mean μ and standard deviation σ of these coefficients.

[0077] For example, if the contrast enhancement coefficients around P are 1.1, 1.2, 1.0, 1.3, 1.1, then the mean μ=(1.1+1.2+1.0+1.3+1.1) / 5=1.12, and the standard deviation σ is obtained by calculating the square root of the sum of the squares of the difference between each coefficient and the mean. When constructing the regional light evaluation matrix, the mean and standard deviation of each pixel point are taken as the elements of the matrix.

[0078] For example, for a 10x10 image region, its light evaluation matrix will contain the mean and standard deviation data of 100 pixel points. According to the preset light uniformity threshold, the light uniformity of each region is judged. Assuming that the threshold is set to 0.1, if the standard deviation of the contrast enhancement coefficient of a certain region is lower than 0.1, then mark the region.

[0079] In one embodiment, by the contrast coefficient distribution graph, the region with insufficient detail appearance degree can be intuitively identified.

[0080] For example, the contrast coefficients of some regions in the distribution graph are concentrated in lower values, indicating that the detail information of these regions is not obvious. Combined with the comprehensive evaluation results of light uniformity and detail appearance degree, the range of the region that needs further feature optimization is determined. The pixel points marked as potential low complexity regions are subjected to local contrast enhancement processing.

[0081] For example, for a pixel point in a certain potential low complexity region, its original brightness value is 80, and the contrast enhancement coefficient is 1.5, then the adjusted brightness value is 80x1.5=120. In this way, the detail appearance of the region is improved. For the region with insufficient detail appearance degree, an adaptive contrast enhancement algorithm is used.

[0082] Specifically, according to the contrast coefficient distribution of the region, the enhancement coefficient is dynamically adjusted.

[0083] For example, if the mean of the contrast coefficient of a certain region is low, the enhancement coefficient can be appropriately increased, such as from 1.0 to 1.3, to ensure that the detail information is effectively improved. According to the optimized contrast coefficient distribution, the uniformity of the region light and the degree of detail expression are re-evaluated.

[0084] For example, the standard deviation of a certain region after optimization decreases from 0.15 to 0.08, indicating that the uniformity of the light is improved; at the same time, the mean of the contrast coefficient increases from 1.0 to 1.2, indicating that the degree of detail expression is enhanced. Through the iterative optimization process, until the contrast coefficient distribution of all regions meets the preset standard.

[0085] For example, the preset standard is that the standard deviation is less than 0.1 and the mean is greater than 1.1, and after multiple iterations and optimization, the contrast coefficient distribution of all regions meets this standard, and the final optimized second image sequence is determined.

[0086] S44, through a convolution operation, the feature vectors containing the decal edges and surface textures are generated by capturing the features of the potential low complexity regions in the optimized second image sequence.

[0087] As a possible implementation, the multi-scale feature mapping of the pixel points is obtained from the second image sequence, the pixel points are convolved through the feature extraction layer of the convolutional neural network, and multi-scale feature maps are generated by using convolution kernels of different sizes, and the texture complexity indicators of the regions are obtained. According to the comparison between the texture complexity indicators and the preset complexity threshold, if the texture complexity indicators are lower than the preset threshold, they are marked as potential low complexity regions, and a distribution matrix of the potential low complexity regions is generated. For the pixel points of the potential low complexity regions, a more fine-grained convolution kernel is used for secondary feature capture, and the feature vector set containing the decal edges and surface textures is generated through convolution operation. Through clustering processing of the feature vector set, the feature distribution of the decal edges and surface textures is obtained, and it is judged whether the feature distribution meets the preset texture expression standard, and the final feature optimization region range is determined.

[0088] For example, in a vehicle decal detection system, obtaining the multi-scale feature mapping of the pixel points from the second image sequence is a key step. The multi-scale feature mapping of the pixel points is realized through the feature extraction layer of the convolutional neural network.

[0089] Specifically, the convolution operation uses convolution kernels of different sizes, such as 3x3, 5x5 and 7x7, to capture various texture information within the image region. Convolution kernels of different sizes can extract features from fine edges to larger region textures, forming multi-scale feature maps.

[0090] For example, a 3x3 convolution kernel can capture the subtle changes of the decal edge, while a 7x7 convolution kernel is suitable for analyzing the overall texture distribution of the decal surface. This multi-scale approach ensures the comprehensiveness of feature extraction, which helps subsequent texture complexity analysis.

[0091] In one possible implementation, the generation of the texture complexity indicator is based on statistical analysis of the multi-scale feature maps. The system processes the feature map of each pixel point and calculates the feature intensity distribution in its local region.

[0092] For example, for a certain pixel point, the feature map value generated by a 3x3 convolution kernel may be concentrated between 0.2 and 0.4, while the feature map value of a 5x5 convolution kernel may be between 0.3 and 0.5. By analyzing the distribution of these values, the texture complexity indicator is generated. Assuming that the preset complexity threshold is 0.35, if the texture complexity indicator of a certain region is lower than 0.35, it will be marked as a potential low complexity region. This marking method can effectively identify the part of the decal region with weak texture information, providing a basis for subsequent optimization.

[0093] Specifically, the distribution matrix of the potential low complexity region is generated by mapping the texture complexity indicator of each pixel point to the image space.

[0094] For example, for a 100x100 pixel image region, the system generates a matrix containing 10000 elements, each element recording the complexity indicator of the corresponding pixel point. This distribution matrix intuitively shows the distribution of the potential low complexity region.

[0095] For example, the edge region of the decal may be marked as a potential low complexity region due to the single texture, while the center region of the decal has higher complexity due to the rich texture. This distribution matrix provides spatial positioning basis for subsequent feature optimization.

[0096] In one possible implementation, for the pixel points in the potential low complexity region, the system uses a more fine-grained convolution kernel for secondary feature capture.

[0097] For example, a 2x2 convolution kernel is used to process the marked region to generate a feature vector set containing the decal edge and surface texture. Assuming that the original feature value of a pixel point in a potential low complexity region is 0.3, after processing by a 2x2 convolution kernel, the feature value is increased to 0.5, capturing more detailed edge information. This fine-grained feature capture can effectively enhance the expression ability of the decal edge.

[0098] For example, the clustering processing of the feature vector set is realized by the k-means algorithm, which divides the feature vectors into multiple categories to distinguish the decal edge and surface texture.

[0099] For example, the clustering results may show that the edge feature vectors are concentrated in a high-frequency feature region, while the surface texture feature vectors are distributed in a low-frequency feature region. By comparing the clustering results with the preset texture expression standard, the system determines whether the feature distribution meets the requirements.

[0100] For example, the preset standard may require that the proportion of edge feature vectors be higher than 60%, and if a certain region is only 50%, further optimization is required. The final feature optimization region range is determined through comprehensive evaluation to ensure that the edge and texture features of the decal region can be accurately expressed. This method significantly improves the feature expression capability of decal detection through multi-scale feature extraction and fine-grained optimization.

[0101] S5, input the feature vector set into the trained defect classification model to obtain a defect detection result.

[0102] In an example, S5 can include the following steps:

[0103] S51, sort the contribution value of each dimension data according to the variance distribution data in the feature vector set, and determine the dimension data with a ranking greater than a ranking threshold as a key indicator.

[0104] S52, perform dimension reduction processing on the vector mapping set corresponding to the key indicator using a data compression method to obtain a low-dimensional representation result.

[0105] S53, for the low-dimensional representation result, integrate the low-dimensional representation result through a matrix optimization tool to obtain a final data form consistent with the matrix optimization target, and obtain an optimized low-dimensional feature representation matrix.

[0106] According to the variance distribution data in the feature vector set, the contribution value of each dimension data is sorted and processed to obtain a sorted feature component list, and at least one key indicator with a higher contribution value is determined. For the feature component list, the key indicator is extracted by a data filtering tool to obtain a data subset related to the feature weight, and if the data subset meets the preset threshold range, a filtered vector mapping set is obtained. According to the vector mapping set, a dimension reduction process is processed using a data compression method to obtain a low-dimensional data structure consistent with the distribution rule, and a low-dimensional representation result is determined. For the low-dimensional representation result, the data is integrated by a matrix optimization tool to obtain a final data form consistent with the matrix optimization target, and an optimized low-dimensional feature representation matrix is obtained.

[0107] For example, in a vehicle decal detection system, variance distribution analysis of the feature vector set is a key step in optimizing feature expression. Variance distribution data reflects the degree of change of feature vectors in each dimension, which helps to identify the most important feature components for expressing decal edges and textures.

[0108] Specifically, the system first performs statistical analysis on the feature vector set, and calculates the variance value of each dimension.

[0109] For example, a set of 1000 feature vectors, each vector has 10 dimensions, the system will calculate the variance of each dimension, and generate a variance distribution list. Assuming that the variance value of the 3rd dimension is 0.6, and the variance value of the 5th dimension is 0.4, and the variance of other dimensions is less than 0.3, then the 3rd and 5th dimensions are considered to be key indicators with higher contribution. This analysis can effectively filter out dimensions that have a greater impact on the edge and texture features of the applique, improving the relevance of subsequent processing.

[0110] In one possible implementation, for the screening of the feature component list, the system uses a data screening tool to extract a data subset related to the key indicators.

[0111] For example, the system sets a contribution threshold of 0.35, and screens out dimensions with a variance value higher than 0.35 to form a data subset. Assuming that the 3rd and 5th dimensions are selected from the 10 dimensions, a new data subset is generated, containing feature vectors related to these two dimensions.

[0112] It should be noted that the data subset needs to be further verified whether it meets the pre-set threshold range.

[0113] For example, the average weight of the feature vectors in the subset is required to be between 0.4 and 0.6. If the average weight of the subset is 0.5, it meets the condition, and a screened vector mapping set is generated. This screening method can effectively focus on the core information of the applique features and reduce redundant data interference.

[0114] For example, for the dimension reduction processing of the vector mapping set, the system uses data compression methods such as principal component analysis to convert high-dimensional data into low-dimensional data structures.

[0115] Specifically, assuming that the original vector mapping set contains 1000 vectors, each vector has 10 dimensions, and the system generates a 3-dimensional low-dimensional data structure through dimension reduction processing. After dimension reduction, the data still retains the distribution rule of the edge and texture of the applique.

[0116] For example, the edge features may be concentrated in a high-value area of a certain dimension. This low-dimensional representation can simplify subsequent processing while retaining key feature information.

[0117] In one possible implementation, for the matrix optimization of the low-dimensional representation result, the system uses a matrix decomposition tool to integrate the data.

[0118] For example, the system optimizes the 3D low-dimensional data to generate an optimized low-dimensional feature representation matrix. Assuming the optimization goal is to ensure that the weight of the applique edge feature in the matrix accounts for more than 70%, the system generates the final data form by adjusting the weight of the matrix elements. This optimization method can enhance the clarity of the expression of the applique edge and texture features, providing high-quality data support for subsequent detection tasks. The entire process forms a logically rigorous feature processing chain through variance analysis, data filtering, dimension reduction, and matrix optimization, ensuring that the applique detection system can efficiently process complex image data.

[0119] S54, input the low-dimensional feature representation matrix into the trained defect classification model to obtain the defect detection result.

[0120] In one possible implementation, the feature extraction tool extracts texture and edge features from clean image data to form feature vector data. Texture features can be extracted by a local binary pattern tool, reflecting the roughness or regularity of the applique surface.

[0121] For example, the texture feature value of a normal applique may be concentrated in the range of 0.5 to 0.9, while the texture value of a defective applique such as a bubble area may be low, in the range of 0.2 to 0.4. Edge features are extracted by an edge detection tool, reflecting the continuity of the applique outline.

[0122] For example, the edge sharpness value of a normal applique may be 0.8, while the edge value of a wrinkle defect may drop to 0.3. This feature vector data provides a reliable basis for subsequent classification.

[0123] It should be noted that when the classification tool processes the feature vector data, if the classification confidence score is lower than the preset threshold, for example 0.7, further comparison is required. The comparison tool will match the feature vector with the known defect feature library to determine whether there is a bubble or wrinkle defect.

[0124] For example, the feature library of a bubble defect may record samples with a texture value below 0.3 and an edge sharpness below 0.4, while the feature library of a wrinkle defect may contain feature points with discontinuous edges. Assuming that the feature vector of a certain applique image shows a texture value of 0.25 and an edge sharpness of 0.35, after comparison it is confirmed that it is highly matched with the bubble defect feature, and it is preliminarily judged as a bubble defect. This comparison method improves the accuracy of defect discrimination.

[0125] For example, the integration tool comprehensively processes the classification confidence score and the comparison data to generate the final defect state. Assuming that the classification tool outputs a bubble defect with a confidence of 0.65, which is lower than the threshold of 0.7, but the comparison tool confirms that the feature matches the bubble defect library with a matching degree of 90%. The integration tool will integrate the results of the two to determine that the applique is a bubble defect.

[0126] Preferably, the integration tool can also introduce a weight mechanism, for example, confidence accounts for 40%, alignment results account for 60%, to balance the output of different tools. This comprehensive processing ensures the reliability of the detection conclusion and provides solid support for the quality control of decals.

[0127] It should be noted that the present application uses a support vector machine classifier to train the low-dimensional feature representation matrix for defect pattern recognition. If there is an overlapping area between the feature distribution of normal decals and defective decals in the training samples, the separation parameter of the classification hyperplane is adjusted to obtain a defect classification model with high discrimination ability.

[0128] The training process includes: according to the normal decal and defective decal data in the training samples, using a data preprocessing tool to standardize the feature matrix and obtain a normalized feature matrix. For the normalized feature matrix, a support vector machine classifier is used to train the feature distribution of normal decals and defective decals to obtain an initial classification hyperplane. If there is an overlapping area in the feature distribution of the initial classification hyperplane, the separation parameter is adjusted by a grid search tool to determine the optimized classification hyperplane. According to the optimized classification hyperplane, a cross-validation tool is used to evaluate the classification model to obtain a defect classification model with high discrimination ability.

[0129] Specifically, in the vehicle decal detection system, based on the normal decal and defective decal data in the training samples, the data preprocessing tool is used to standardize the feature matrix, which is the basis for building a high-quality classification model.

[0130] For example, suppose the training samples contain 2000 decal images, of which 1000 are normal decals and 1000 are defective decals. Each image extracts a feature vector containing 15 dimensions, such as color mean, texture contrast, and edge sharpness, etc. The data preprocessing tool will standardize these feature vectors.

[0131] Specifically, the feature values of each dimension are scaled to a range with a mean of 0 and a standard deviation of 1. Suppose the original value of the color mean of a certain dimension is 50 to 200, and after standardization it is converted to the range of -2 to 2. This standardization can eliminate the dimension difference of different dimensions and ensure the consistency of the data processed by the subsequent classifier. For the normalized feature matrix, a support vector machine classifier is used to train the feature distribution of normal decals and defective decals.

[0132] For example, the support vector machine separates the feature points of normal decals and defective decals by finding a classification hyperplane. Suppose after training, it is found that the edge sharpness feature values of normal decals are concentrated in 0.8 to 1.2, while those of defective decals are concentrated in 0.2 to 0.6. The classifier will construct an initial classification hyperplane based on these distributions. This method can effectively capture the differences between the two types of decals in the feature space.

[0133] It should be noted that if there is an overlapping area in the feature distribution of the initial classification hyperplane, for example, the feature points of normal stickers and defective stickers are mixed in the interval of 0.6 to 0.8 of the edge sharpness, the interval parameter needs to be adjusted by the grid search tool to optimize the classification hyperplane.

[0134] In a possible implementation, the grid search traverses the penalty parameter C and the kernel function parameter γ of the support vector machine, and attempts to take C as 0.1, 1, and 10, and γ as 0.01, 0.1, and 1. It is assumed that after testing, it is found that when C = 1 and γ = 0.1, the classification hyperplane can maximize the interval of the two types of stickers and reduce the number of feature points in the overlapping area. This optimization can improve the ability of the classifier to distinguish complex sticker features.

[0135] In an embodiment, for the optimized classification hyperplane, a cross-validation tool is used to evaluate the performance of the classification model.

[0136] For example, using 5-fold cross-validation, 2000 samples are divided into 5 groups, each group containing 400 samples, and 4 groups are used for training and 1 group is used for testing in turn. It is assumed that the test results show that the recognition accuracy of the classification model for normal stickers is 95%, the accuracy for defective stickers is 92%, and the overall accuracy is 93.5%. By analyzing the cross-validation results, it can be found that the model has a small amount of misjudgment on the defective stickers with low edge sharpness, which prompts further optimization for this feature. This evaluation method can ensure the stability and reliability of the model on different data subsets, and provide high-quality classification support for the sticker detection task.

[0137] In some embodiments, based on the sticker image data, a morphological opening operation tool can be used to preliminarily separate the defect area, extract the possible abnormal part, obtain preliminary defect contour data, and determine the abnormal range for subsequent processing. For the preliminary defect contour data, a connected component analysis tool is used to segment and label the abnormal part to obtain the boundary information of the independent defect area and get the accurate defect distribution. If the boundary information of the independent defect area is confirmed, an area calculation tool is used to measure the size data of the defect area and compare it with the total area of the sticker. If the defect area ratio exceeds the preset threshold, the area is determined as an unqualified part and the defect severity level is determined. According to the distribution and severity level of the unqualified part, a coordinate mapping tool is used to mark the position of the defect area to obtain detailed position coordinate data and get the final defect positioning result.

[0138] In a possible implementation, the morphological opening operation tool separates the defect area in the sticker image through erosion and dilation operations.

[0139] For example, assume a vehicle decal image contains a bubble defect, the original image may have a blurred boundary between the defect area and the normal area due to light or material reflection. The morphological opening operation first erodes the image, weakening small noise and irregular bright spots, making the defect area more prominent; then through inflation, the overall shape of the defect area is restored, but the clear outline is preserved.

[0140] For example, if the bubble defect appears as a circular area with a diameter of about 2 mm, the erosion operation can remove noise points less than 1 mm around it, and the inflation operation ensures the integrity of the bubble outline. This operation can effectively separate the defect area and provide clear preliminary outline data for subsequent analysis.

[0141] For example, the connected component analysis tool further segments and labels independent defect areas after obtaining the preliminary defect outline. Assume that there are multiple bubbles and a wrinkle in the decal image, the connected component analysis will identify each independent defect area based on pixel connectivity, and generate boundary information for each area.

[0142] Specifically, a bubble defect may be labeled as area A, and its boundary information includes a set of closed pixel coordinates; the wrinkle defect is labeled as area B, which appears as an elongated and irregular boundary line. This labeling method decomposes complex defect distribution into independent analysis units, ensuring that subsequent processing can accurately analyze each defect area.

[0143] It should be noted that the area calculation tool measures the size of the defect area through the boundary information, and compares it with the total area of the decal to evaluate the severity of the defect.

[0144] For example, if the total area of the decal is 100 square centimeters, the area of a single bubble defect is 0.5 square centimeters, accounting for 0.5%, and the preset threshold is 1%. By comparison, the area calculation tool determines that the bubble defect is of minor grade. However, if another wrinkle defect has an area of 2 square centimeters, accounting for 2%, it is determined to be a serious defect and needs to be handled first. This quantitative method makes the defect severity evaluation more objective and provides a basis for quality control.

[0145] In one possible implementation, the coordinate mapping tool labels the location of the unqualified defect area and generates accurate coordinate data.

[0146] For example, assume that the decal is located on the door panel of a vehicle, the coordinate mapping tool maps the boundary coordinates of the defect area to the actual location of the door panel, generating a positioning result similar to "the bubble defect is located at the upper left corner of the door panel, with coordinates ranging from (50, 30) to (60, 40)". This location information facilitates subsequent repair personnel to quickly locate the defect and improves repair efficiency.

[0147] Preferably, the coordinate mapping tool can generate three-dimensional coordinate data in combination with the geometry of the decal, further adapting to the positioning needs of the curved decal.

[0148] For example, if the decal covers the curved area of the vehicle, the coordinate mapping tool can convert the two-dimensional image coordinates into three-dimensional space coordinates through the projection algorithm, ensuring that the defect position is consistent with the actual decal surface. This multi-dimensional positioning method improves the applicability in complex scenarios and provides comprehensive support for decal defect detection.

[0149] To sum up, the present application provides a vehicle PVC decal defect detection method based on AI vision, which includes: obtaining test image data of vehicle decals; in the case that the test image data has an overexposed area, adjusting the camera parameters to obtain a first image sequence with uniform brightness distribution; the overexposed area is an area with a brightness value greater than a brightness threshold; performing image optimization processing on the first image sequence to obtain a second image sequence; the image optimization processing includes at least one of the following: illumination compensation processing, denoising processing, contrast optimization processing; capturing features of potential low complexity regions in the second image sequence through convolution operation to generate a feature vector set containing decal edges and surface textures; the potential low complexity region is a region with a texture complexity index lower than a texture complexity threshold; the texture complexity index is used to represent the complexity of the texture in the second image sequence; inputting the feature vector set into a trained defect classification model to obtain a defect detection result.

[0150] The embodiment of the present application also provides a terminal device. The terminal device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a vehicle PVC decal defect detection program. The processor implements the steps in each of the vehicle PVC decal defect detection method embodiments described above when executing the computer program, such as Figure 1 The steps S11 shown. Alternatively, the processor implements the functions of each module / unit in each system embodiment described above when executing the computer program.

[0151] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present application. The one or more modules / units can be a series of computer program instruction segments that can complete a specific function, which are used to describe the execution process of the computer program in the terminal device.

[0152] The terminal device can be a desktop computer, a notebook computer, a palm computer, a smart tablet, and the like. The terminal device can include, but is not limited to, a processor, a memory. Those skilled in the art can understand that the above components are only examples of the terminal device, and do not constitute a limitation on the terminal device, and can include more or fewer components than the above, or combine certain components, or different components, for example, the terminal device can also include an input / output device, a network access device, a bus, and the like.

[0153] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, and the like. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor and the like. The processor is a control center of the terminal device, and connects various parts of the terminal device through various interfaces and lines.

[0154] The memory can be used to store the computer programs and / or modules, and the processor realizes various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory, and calling data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required by a function (such as a sound playing function, an image playing function, and the like), and the like; and the data storage area can store data created according to use of the terminal device (such as audio data, a phone book, and the like), and the like. In addition, the memory can include a high-speed random access memory, and can also include a nonvolatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0155] The modules / units integrated in the terminal device, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can implement the steps of each method embodiment when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer readable medium can include any entity or system, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium, etc. that can carry the computer program code. It should be noted that the computer readable medium can include or exclude contents according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electric carrier signals and telecommunication signals.

[0156] It should be noted that the above-described system embodiments are only illustrative, and the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. In addition, the connection relationship between the modules in the system embodiments provided by the present application indicates that there is a communication connection between them, which can be realized as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0157] The above-described specific embodiments further illustrate the purpose, technical solutions, and beneficial effects of the present application. It should be understood that the above-described specific embodiments are only examples of the present application and are not intended to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An AI vision-based vehicle PVC decal defect detection method, characterized in that, The method comprises the following steps: acquiring test image data of vehicle decals; in the case that the test image data has overexposed areas, adjusting camera parameters to obtain a first image sequence with uniform brightness distribution; the overexposed areas are areas with brightness values greater than a brightness threshold; performing image optimization processing on the first image sequence to obtain a second image sequence; the image optimization processing comprises at least one of the following: illumination compensation processing, noise reduction processing, and contrast optimization processing; performing feature capture on potential low-complexity areas in the second image sequence through convolution operation to generate a feature vector set containing decal edges and surface textures; the potential low-complexity areas are areas with texture complexity indicators lower than a texture complexity threshold; the texture complexity indicator is used to represent the complexity of the textures in the second image sequence; inputting the feature vector set into a trained defect classification model to obtain a defect detection result; wherein the feature capture on the potential low-complexity areas in the second image sequence through convolution operation to generate the feature vector set containing decal edges and surface textures comprises: determining whether there are areas in the second image sequence with contrast enhancement coefficient standard deviations lower than a standard deviation threshold; the contrast enhancement coefficient is a coefficient for modifying the contrast of different areas in the second image sequence; in the case that there are areas with contrast enhancement coefficient standard deviations lower than the standard deviation threshold, adjusting the contrast enhancement coefficient and re-calculating the contrast enhancement coefficient standard deviation around each pixel point; iteratively performing the above steps until there are no areas in the second image sequence with contrast enhancement coefficient standard deviations lower than the standard deviation threshold, to obtain an optimized second image sequence; performing feature capture on the potential low-complexity areas in the optimized second image sequence through convolution operation to generate the feature vector set containing decal edges and surface textures; wherein the image optimization processing on the first image sequence to obtain the second image sequence comprises: applying a preprocessing algorithm based on a combination of histogram equalization and Gaussian filtering to perform illumination compensation and noise suppression processing on the first image sequence; calculating gradient information of the first image sequence after the illumination compensation and noise suppression processing; the gradient information comprises gradient values of each pixel point; determining contrast enhancement coefficients of each pixel point according to a first mapping relationship and the gradient values of each pixel point; the first mapping relationship comprises a mapping relationship between different gradient values and different contrast enhancement coefficients; areas with larger gradient values represent detailed parts of the image and need to enhance the contrast; areas with smaller gradient values represent smooth parts and need to weaken the contrast; for each pixel point, determining a target brightness value of the pixel point according to the contrast enhancement coefficient and the original brightness value, the target brightness value being the product of the contrast enhancement coefficient and the original brightness value, adjusting the brightness value of the pixel point to the target brightness value to obtain the second image sequence.

2. The method of claim 1, wherein, the preprocessing algorithm based on a combination of histogram equalization and Gaussian filtering is applied to perform illumination compensation and noise suppression processing on the first image sequence, comprising: counting the number of pixels of each brightness level in the first image sequence to construct a histogram of the first image sequence; adjusting the brightness value of each pixel point according to the distribution of the histogram to obtain a first image sequence after illumination compensation; performing convolution operation on the first image sequence after illumination compensation based on a preset size of Gaussian kernel and a preset standard deviation to obtain a first image sequence after illumination compensation and noise suppression processing.

3. The method of claim 1, wherein, The calculation of the gradient information of the first image sequence after illumination compensation and noise suppression processing comprises: For any pixel point in the first image sequence, determining the gradient value of the pixel point in the horizontal direction and the gradient value in the vertical direction; According to the gradient value of the pixel point in the horizontal direction, the gradient value in the vertical direction, and the gradient determination formula, the gradient value of the pixel point is calculated to obtain the gradient information of the first image sequence after illumination compensation and noise suppression processing; Wherein, the gradient determination formula satisfies the following relationship: ; represents a gradient value of a pixel point, represents a square root function, represents a gradient value in a horizontal direction, represents a gradient value in a vertical direction.

4. The method of claim 1, wherein, The method further comprises: performing convolution operation on the pixel points in the second image sequence through the feature extraction layer of the convolution neural network to generate multi-scale feature maps using convolution kernels of different sizes to obtain texture complexity indicators of each region.

5. The method of claim 1, wherein, The input of the feature vector set into the trained defect classification model to obtain a defect detection result comprises: According to the variance distribution data in the feature vector set, the contribution value of each dimension data is sorted, and the dimension data with a ranking greater than a ranking threshold is determined as a key indicator; using a data compression method to reduce the dimension of the vector mapping set corresponding to the key indicator to obtain a low-dimensional representation result; integrating the low-dimensional representation result through a matrix optimization tool to obtain an optimized low-dimensional feature representation matrix in the form consistent with the matrix optimization target; inputting the low-dimensional feature representation matrix into the trained defect classification model to obtain the defect detection result.

6. A vehicle PVC decal defect detection system for implementing the vehicle PVC decal defect detection method according to any one of claims 1 to 5, characterized by, Comprise: acquisition unit, processing unit; The acquisition unit is used to acquire test image data of vehicle decals; The processing unit is configured to adjust camera parameters to obtain a first image sequence with uniform brightness distribution when the test image data has an overexposed area; the overexposed area is an area with a brightness value greater than a brightness threshold; The processing unit is further configured to perform image optimization processing on the first image sequence to obtain a second image sequence; the image optimization processing comprises at least one of the following: illumination compensation processing, noise reduction processing, and contrast optimization processing; The processing unit is further configured to capture features in potential low complexity regions in the second image sequence through convolution operation to generate a feature vector set containing decal edges and surface textures; the potential low complexity region is a region with a texture complexity indicator lower than a texture complexity threshold; the texture complexity indicator is used to represent the complexity of the texture in the second image sequence; The processing unit is further configured to input the feature vector set into a trained defect classification model to obtain a defect detection result.

7. A computer readable storage medium characterized in that, The computer readable storage medium comprises a stored computer program, wherein the computer readable storage medium controls a device where the computer readable storage medium is located to execute the vehicle PVC decal defect detection method according to any one of claims 1 to 5 when the computer program runs.

Images