Intelligent detection system and method for appearance flaws of sports shoes based on small sample learning

Abstract

The invention relates to the technical field of visual inspection, in particular to a sneaker appearance flaw intelligent detection system and method based on small sample learning, and the system comprises a flaw intelligent detection center, a flaw analysis unit, a check acquisition unit, a difficulty evaluation unit and a visual feedback unit. According to the method, defect type identification and bounding box coordinate output of the sports shoes are realized in a mode of combining the model and the 2D image, meanwhile, along with precision analysis of the bounding box coordinates, data support is provided for follow-up, a defect area is obtained based on analysis of the bounding box coordinates, and meanwhile, 3D and spectral image analysis is performed on the defect area, so that the detection accuracy of the sports shoes is improved. The obtained flaw type is checked with the flaw type analyzed by the 2D image, so that the consistency of flaw results is ensured, the error rate of flaw detection results is reduced, and repair feasibility judgment processing is carried out, so that repair difficulty classification of the sneakers with flaws is facilitated, namely, objective classification is realized in combination with multi-index weighted fusion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual detection, and in particular to a sneaker appearance flaw intelligent detection system and method based on small sample learning. BACKGROUND

[0002] Sneaker appearance quality detection is a key process in the production link, directly affecting product quality and brand image. In the prior art, the detection method based on deep learning requires a large number of labeled samples to train the model, but the sneaker style is updated quickly (thousands of styles per year), the material is various (leather, mesh, TPU, etc.), and the new flaw (such as 3D printed shoe surface interlayer bubble) sample is scarce, resulting in insufficient model generalization ability. The small sample learning technology can quickly master new tasks from a small number of samples by learning meta-knowledge of "how to learn", providing a new idea for solving the sample scarcity problem. However, the traditional system only relies on 2D image detection, which is difficult to accurately identify flaws and positions, and has the problem of low detection accuracy. At the same time, it is difficult to verify the accuracy of the position information, resulting in a large deviation of the subsequent flaw recognition accuracy analysis result, which is not conducive to the flaw repair classification management of the sneaker. In view of the above technical defects, a solution is proposed. SUMMARY

[0003] The purpose of the present application is to provide a sneaker appearance flaw intelligent detection system and method based on small sample learning to solve the above technical defects. The present application realizes the identification of the flaw type of the sneaker and the output of the bounding box coordinates by combining the model and the 2D image, and at the same time, the accuracy analysis of the bounding box coordinates is carried out to provide data support for the subsequent analysis. At the same time, the 3D and spectral image analysis of the flaw area is carried out, the flaw type obtained is compared with the flaw type analyzed by the 2D image to ensure the consistency of the flaw result, reduce the error rate of the flaw detection result, and carry out the repair feasibility discrimination processing, which is helpful for the repair difficulty classification of the sneaker with flaws, that is, the objective classification is realized by combining multi-index weighted fusion.

[0004] The purpose of the present application can be realized by the following technical scheme: a sneaker appearance flaw intelligent detection system based on small sample learning, comprising a flaw intelligent detection center, a flaw analysis unit, a comparison and acquisition unit, a difficulty evaluation unit and a visual feedback unit. The flaw intelligent detection center is used for retrieving and storing the surface feature image set of the sneaker. The flaw analysis unit is used for flaw identification and positioning analysis of the surface feature image set to obtain the first flaw type and the bounding box coordinates. The pixel-level feature comparison and analysis process based on the bounding box coordinate accuracy verification obtains the standard signal or the trimming signal. When the standard signal is generated, the collation acquisition unit is used for coordinate transformation and defect recognition analysis on the bounding box coordinates, consistency comparison analysis is performed on the obtained second defect type and the first defect type, and a self-check signal or an output signal is obtained; The difficulty evaluation unit is used for performing repair difficulty quantitative evaluation analysis on the collected surface image of the defect area of the sports shoes, and performing discrimination processing on whether the obtained repair difficulty coefficient exceeds the preset repair difficulty coefficient threshold, to obtain a difficult-to-repair signal or a repairable signal.

[0005] Preferably, the surface feature image set represents images of the front, side and sole of the sports shoes obtained by using an industrial camera and a light source assembly.

[0006] Preferably, the defect recognition and positioning analysis process is as follows: S1: Preprocessing the obtained surface feature image set, which includes denoising and size normalization; S2: Extracting the surface feature image set of the sports shoes, manually labeling the defects in the surface feature image set, and the labeling content includes defect type and position information, to obtain a sample training data set containing manual labeling; S3: Preprocessing the images in the sample training data set to obtain preprocessed training samples, dividing the training samples into a support set and a query set, and finally constructing a defect recognition model based on iterative training of the support set and the query set; S4: Inputting the preprocessed surface feature image set into the defect recognition model to obtain the first defect type and the bounding box coordinates output by the defect recognition model.

[0007] Preferably, the pixel-level feature comparison analysis process is as follows: Obtaining the appearance feature images of various defects and the bounding boxes and feature labels labeled in the appearance feature images, constructing a defect feature library based on the preprocessed bounding boxes and feature labels labeled in the appearance feature images, and the defect feature library includes morphological features, color and texture features; Based on the bounding box coordinates, the pixel features in the bounding box are obtained, including texture and color, and the extracted pixel features are compared with the defect feature library, the similarity obtained after comparing the pixel features in the bounding box with the defect feature library is obtained, and whether the similarity exceeds the preset similarity threshold is discriminated, if yes, a standard signal is generated, and if not, a trimming signal is generated.

[0008] Preferably, the coordinate transformation and defect recognition analysis process is as follows: The obtained bounding box coordinates are converted into 3D point cloud coordinates by existing technology, and the area surrounded by the 3D point cloud coordinates is set as a defect area; Point cloud image information of a defect area in the sports shoes is acquired by a 3D structure light scanner, and surface spectral image data of the defect area in the sports shoes is acquired by a hyperspectral camera; The point cloud image information and the surface spectral image data are preprocessed, and features of the preprocessed point cloud image information and the surface spectral image data are extracted respectively; Features of the aligned point cloud image information and the surface spectral image data are fused by using a feature splicing method in the prior art to obtain a multi-modal feature vector.

[0009] Preferably, the multi-modal feature vector is input into a pre-set meta-knowledge graph model to obtain output geometric similarity and spectral similarity; A sum value obtained by multiplying the geometric similarity and the spectral similarity by corresponding preset weight coefficients is set as a comprehensive similarity, and whether the comprehensive similarity exceeds a preset comprehensive similarity threshold is discriminated, if yes, a second defect type is output, and if no, an artificial signal is generated.

[0010] Preferably, the first defect type and the second defect type are subjected to consistency comparison and analysis, if the first defect type and the second defect type are consistent, an output signal is generated, and if the first defect type and the second defect type are consistent, a self-check signal is generated.

[0011] Preferably, the repair difficulty quantitative evaluation and analysis process is as follows: A surface image of the defect area in the sports shoes is acquired, and the surface image is subjected to gray scale processing to acquire a total number of defect pixels in a defect boundary box in the surface image after the gray scale processing; A difference between an average gray scale value in the defect area and a preset average gray scale value threshold is acquired, and the difference between the average gray scale value in the defect area and the preset average gray scale value threshold is set as a pixel gray scale difference; A main skeleton of the defect in the defect area is acquired by a skeleton extraction algorithm (such as a Zernike moment), a pixel distance between two end points of the skeleton is calculated, and the pixel distance between the two end points of the skeleton is set as a defect length; The total number of defect pixels, the pixel gray scale difference and the defect length are respectively multiplied by corresponding set weight coefficients, a sum value obtained by the multiplication is set as a repair difficulty coefficient, and whether the repair difficulty coefficient exceeds a preset repair difficulty coefficient threshold is discriminated, if yes, a difficult-to-repair signal is generated, and if no, a repairable signal is generated.

[0012] The present application has the following advantages: (1) The small sample learning technology is adopted, the model can be trained only by a small amount of labeled samples, the dependence on a large amount of labeled data is greatly reduced, the manual labeling cost is reduced, the defect type recognition and the boundary box coordinate output of the sports shoes are realized by combining the model and the 2D image, and the pixel-level feature comparison and analysis of the accuracy of the boundary box coordinates are further carried out, so that whether the accuracy of the boundary box coordinates is qualified is judged, data support is provided for subsequent analysis, and the accuracy of subsequent analysis results is improved; (2) The defect area is obtained based on the boundary box coordinate analysis, 3D and spectral image analysis is carried out on the defect area, the obtained defect type is checked with the defect type obtained by 2D image analysis, the consistency of the defect result is ensured, and the error rate of the defect detection result is reduced; (3) The repair difficulty evaluation is carried out from the pixel-level feature angle of the defect area, and the repair feasibility discrimination processing is accompanied, which is helpful for the repair difficulty classification of the sports shoes with defects, that is, objective classification is realized by combining multi-index weighted fusion. BRIEF DESCRIPTION OF DRAWINGS

[0013] The application will be further described below with reference to the drawings; Fig. 1 is a system flow block diagram of the application; Fig. 2 is a method analysis reference diagram of the application; Fig. 3 is a partial analysis diagram of embodiment one of the application. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the application will be apparently and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0015] In this document, reference to“an embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another; Embodiment one: Please refer to Figs. 1 to 3As shown, the present application is a sports shoe appearance flaw intelligent detection system based on small sample learning, which comprises a flaw intelligent detection center, a flaw analysis unit, a verification and acquisition unit, a difficulty evaluation unit and a visual feedback unit. The flaw intelligent detection center is in one-way communication connection with the flaw analysis unit and the visual feedback unit. The flaw analysis unit is in one-way communication connection with the verification and acquisition unit. The verification and acquisition unit is in one-way communication connection with the difficulty evaluation unit. The difficulty evaluation unit is in one-way communication connection with the flaw intelligent detection center. The flaw intelligent detection center is used to retrieve the surface feature image set of sports shoes and store it; The flaw analysis unit is used to perform flaw identification and positioning analysis on the surface feature image set. The specific flaw identification and positioning analysis process is as follows: The surface feature image set represents the images of the front, side and sole of sports shoes obtained by using industrial cameras, light source assemblies and other components; S1: Preprocess the obtained surface feature image set, which includes denoising, size normalization, etc. For example, Gaussian filtering is used to remove noise in the image. Size normalization uniformly adjusts the image to a preset size (such as 224x224 pixels) for model processing. S2: Extract the surface feature image set of sports shoes, manually label the flaws in the surface feature image set, and the labeling content includes flaw type (scratches, stains, glue opening, color difference, etc.) and position information (bounding box coordinates), to obtain a sample training data set containing manual labeling; S3: Preprocess the images in the sample training data set to obtain preprocessed training samples, divide the training samples into support set and query set, and perform iterative training based on the support set and query set to finally build a flaw identification model; S4: Input the preprocessed surface feature image set into the flaw identification model to obtain the first flaw type and bounding box coordinates output by the flaw identification model; S5: Pixel-level feature comparison analysis process based on bounding box coordinate accuracy verification: obtain the appearance feature images of various flaws and the bounding box, feature label and other information labeled in the appearance feature images. Based on the bounding box, feature label and other information labeled in the appearance feature images, a flaw feature library is constructed after preprocessing, which includes morphological features, color and texture features, etc. The pixel features in the bounding box are obtained based on the bounding box coordinates, and the pixel features include texture, color, etc. The extracted pixel features are compared with the defect feature library, the similarity obtained after the comparison of the pixel features in the bounding box with the defect feature library is obtained, and whether the similarity exceeds a preset similarity threshold is determined. If yes, a standard signal is generated, and if no, a trimming signal is generated. The visual feedback unit is used to respond to the standard signal or the trimming signal, and immediately display the preset warning text corresponding to the standard signal or the trimming signal, so as to revise the obtained bounding box coordinates, thereby improving the accuracy of the bounding box coordinates.

[0016] Embodiment two: When the standard signal is generated, the acquisition unit is used for coordinate transformation and defect recognition analysis of the bounding box coordinates. The specific coordinate transformation and defect recognition analysis process is as follows: The bounding box coordinates (2D) of the image are based on the pixel coordinate system (with the origin at the upper left corner of the image, and the unit is pixel), while the 3D point cloud coordinates are based on the world coordinate system (or camera coordinate system, with the unit of millimeter / meter). The conversion of the two depends on the camera internal parameter (which describes the optical characteristics of the camera) and the external parameter (which describes the pose relationship between the camera and the world coordinate system), and the mapping is realized through the perspective projection model. The obtained bounding box coordinates are converted into 3D point cloud coordinates through the existing technology, and the area surrounded by the 3D point cloud coordinates is set as the defect area. The point cloud image information of the defect area in the sports shoes is obtained through the 3D structured light scanner, and the surface spectral image data of the defect area in the sports shoes is obtained through the hyperspectral camera. The point cloud image information and the surface spectral image data are preprocessed, including noise processing and smoothing processing. The features of the preprocessed point cloud image information and the surface spectral image data are extracted respectively, including the curvature change rate, the normal vector deviation, and the spectral angle and the reflectivity deviation value of the preprocessed surface spectral image data. The SIFT (Scale-Invariant Feature Transform) feature point matching algorithm and the ICP (Iterative Closest Point) algorithm in the existing technology are used to realize the coordinate alignment of the 3D point cloud image and the hyperspectral image. In this way, the positions of the two can be quickly aligned. Through accurate feature alignment, the effective fusion of 3D geometric features and hyperspectral features is realized, and the complementary advantages of multi-modal data are fully utilized, providing more comprehensive feature support for subsequent defect detection. The features of the aligned point cloud image information and the surface spectral image data are fused by using the feature splicing method in the existing technology to obtain a multi-modal feature vector. The extracted curvature change rate, normal vector deviation and other geometric features are spliced with spectral angle, reflectance deviation value and other spectral features to form a feature vector with higher dimension and richer information, which provides input for the small sample fusion detection model; The multi-modal feature vector is input into the pre-set meta-knowledge graph model to obtain the output geometric similarity (measured by Euclidean distance to measure the spatial difference of geometric features between the sports shoes and the standard sports shoes) and the spectral similarity (measured by cosine similarity to measure the angle difference of spectral features between the sports shoes and the standard sports shoes); The pre-set meta-knowledge graph model stores the "geometric-spectral" correlation rules of defects in the form of triples, for example: Scratch: geometric feature is "linear depression, depth 0.05-0.3mm, length >1mm", spectral feature is "reflectance decrease of 15%-30% at 450nm waveband"; Bubble: geometric feature is "circular protrusion, height 0.05-0.2mm, diameter 0.5-3mm", spectral feature is "poor uniformity of full-waveband reflectance, ΔR>10%" and so on; The sum of the geometric similarity and the spectral similarity multiplied by the corresponding preset weight coefficient is set as the comprehensive similarity, and whether the comprehensive similarity exceeds the preset comprehensive similarity threshold is determined, if yes, the second defect type is output, if not, an artificial signal is generated; The first defect type and the second defect type are compared and analyzed for consistency, if the first defect type and the second defect type are consistent, an output signal is generated, if the first defect type and the second defect type are consistent, a self-check signal is generated, and the visual feedback unit is used to respond to the artificial signal or the self-check signal or the output signal, and immediately display the preset warning text corresponding to the artificial signal or the self-check signal or the output signal, so that the defect type and the bounding box coordinates of the sports shoes can be intuitively understood according to the feedback information, and the error of the defect detection result can be understood, so that timely feedback and warning management can be carried out to ensure the accuracy and reliability of the defect detection of the sports shoes; The difficulty evaluation unit is used for quantitative evaluation and analysis of the repair difficulty of the surface image of the defect area in the collected sports shoes, and the specific repair difficulty quantitative evaluation and analysis process is as follows: The surface image of the defect area in the sports shoes is obtained, and the surface image is subjected to gray scale processing to obtain the total number of defect pixels in the defect bounding box after gray scale processing; The difference between the average gray value in the defect area and the preset average gray value threshold is obtained, and the difference between the average gray value in the defect area and the preset average gray value threshold is set as the pixel gray difference; The main skeleton of the defect in the defect area is obtained by the skeleton extraction algorithm (such as Zernike moment), the pixel distance between the two end points of the skeleton is calculated, and the pixel distance between the two end points of the skeleton is set as the defect length; The total number of defective pixels, the pixel gray scale difference and the defect length are multiplied by the corresponding set weight coefficient respectively, the multiplied sum value is set as the repair difficulty coefficient, and whether the repair difficulty coefficient exceeds the preset repair difficulty coefficient threshold is determined, if yes, a difficult repair signal is generated, if not, a repairable signal is generated, and the visual feedback unit is used to respond to the difficult repair signal or the repairable signal, and immediately display the preset warning operation corresponding to the difficult repair signal or the repairable signal, which is helpful for classifying the sneakers with defects according to repair difficulty, that is, realizing objective classification by combining multi-index weighted fusion.

[0017] Embodiment three: The intelligent detection method for appearance defects of sneakers based on small sample learning comprises the following steps: Step one: retrieval and storage of the surface feature image set of the sneakers; Step two: defect recognition and positioning analysis process based on the surface feature image set, that is, defect recognition and positioning analysis are performed on the surface feature image set to obtain the first defect type and the boundary box coordinates; Step three: pixel-level feature comparison analysis process based on boundary box coordinate precision verification, that is, the pixel features in the obtained boundary box are compared and analyzed to obtain a standard signal or a trimming signal; Step four: coordinate transformation and defect recognition analysis process of the boundary box coordinates based on information progression, that is, whether the obtained comprehensive similarity exceeds the preset comprehensive similarity threshold is determined to obtain a manual signal or a self-check signal or an output signal; Step five: repair difficulty quantitative evaluation and analysis process of the surface image of the defect area in the sneakers, that is, whether the obtained repair difficulty coefficient exceeds the preset repair difficulty coefficient threshold is determined to obtain a difficult repair signal or a repairable signal; In summary, the small sample learning technology is adopted in the present application, and only a small amount of labeled samples are needed to train the model, which greatly reduces the dependence on a large amount of labeled data and reduces the cost of manual labeling. At the same time, the model and the 2D image are combined to realize the defect type recognition and boundary box coordinate output of the sneakers, and further pixel-level feature comparison analysis is performed on the precision of the boundary box coordinates to determine whether the precision of the boundary box coordinates is qualified, so as to provide data support for the subsequent analysis, thereby improving the accuracy of the subsequent analysis results. Based on the boundary box coordinate analysis, the defect area is obtained, and 3D and spectral image analysis is performed on the defect area. The obtained defect type is checked with the defect type obtained by 2D image analysis to ensure the consistency of the defect result and reduce the error rate of the defect detection result. In addition, the repair difficulty is evaluated from the pixel-level feature angle of the defect area, and accompanied by repair feasibility determination, which is helpful for classifying the sneakers with defects according to repair difficulty, that is, realizing objective classification by combining multi-index weighted fusion.

[0018] The setting of the threshold is for the result comparison analysis, so as to determine whether it is good or bad, and the size of the threshold is set according to the large model analysis of sample data and artificial experience, and the sample data is also recorded and stored, and appropriate adjustment can be made through seasonal or reasonable influence conditions.

[0019] The size of the coefficient is a specific value obtained by quantifying each parameter, which is convenient for subsequent comparison. The size of the coefficient depends on the amount of sample data and the corresponding running coefficient preliminarily set by the person skilled in the art for each group of sample data; as long as it does not affect the proportional relationship between the parameter and the quantized value.

[0020] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A smart detection system for appearance defects in athletic shoes based on few-sample learning, characterized in that, It includes a defect intelligent detection center, a defect analysis unit, a verification and acquisition unit, a difficulty assessment unit, and a visual feedback unit; The intelligent defect detection center is used to retrieve and store the surface feature image set of sports shoes; The defect analysis unit is used to identify and locate defects in the surface feature image set to obtain the first defect type and bounding box coordinates. A pixel-level feature comparison and analysis process based on bounding box coordinate accuracy verification yields standard or trimmed signals. When a standard signal is generated, the verification and acquisition unit is used to perform coordinate transformation and defect identification analysis on the bounding box coordinates, and performs consistency comparison analysis between the obtained second defect type and the first defect type to obtain a self-test signal or output signal. The difficulty assessment unit is used to quantitatively assess and analyze the repair difficulty of the surface images of the defective areas in the collected sports shoes, and to determine whether the obtained repair difficulty coefficient exceeds the preset repair difficulty coefficient threshold, so as to obtain a difficult repair signal or a repairable signal.

2. The intelligent detection system for sports shoe appearance defects based on few-sample learning according to claim 1, characterized in that, The surface feature image set represents images of the front, side, and sole of the athletic shoe acquired using an industrial camera and light source assembly.

3. The intelligent detection system for sports shoe appearance defects based on few-sample learning according to claim 1, characterized in that, The defect identification and location analysis process is as follows: S1: Preprocess the acquired surface feature image set, including noise reduction and size normalization; S2: Extract the surface feature image set of the sports shoes, manually annotate the defects in the surface feature image set, and the annotation content includes the defect type and location information to obtain the sample training dataset containing manual annotations; S3: Preprocess the images in the sample training dataset to obtain preprocessed training samples. Divide the training samples into support set and query set. Iterate and train based on the support set and query set to finally build a defect recognition model. S4: Input the preprocessed surface feature image set into the defect recognition model to obtain the first defect type and bounding box coordinates output by the defect recognition model.

4. The intelligent detection system for sports shoe appearance defects based on few-sample learning according to claim 1, characterized in that, The pixel-level feature comparison and analysis process is as follows: The appearance feature images of various historical defects and the bounding boxes and feature labels marked in the appearance feature images are obtained. Based on the bounding boxes and feature labels marked in the appearance feature images, a defect feature library is constructed after preprocessing. The defect feature library includes morphological features, color and texture features. The pixel features within the bounding box are obtained based on the bounding box coordinates. The pixel features include texture and color. The extracted pixel features are compared with the defect feature library to obtain the similarity between the pixel features within the bounding box and the defect feature library. The similarity is then judged to see if it exceeds a preset similarity threshold. If it does, a standard signal is generated; otherwise, the signal is adjusted.

5. The intelligent detection system for sports shoe appearance defects based on few-sample learning according to claim 1, characterized in that, The coordinate transformation and defect identification analysis process is as follows: The obtained bounding box coordinates are converted into 3D point cloud coordinates using existing technology, and the area enclosed by the 3D point cloud coordinates is set as the defect area. Point cloud image information of the defective area in the sports shoe was obtained by using a 3D structured light scanner, and surface spectral image data of the defective area in the sports shoe was obtained by using a hyperspectral camera. Preprocessing is performed on point cloud image information and surface spectral image data, and feature extraction is performed on the preprocessed point cloud image information and surface spectral image data respectively; The feature stitching method in the existing technology is used to fuse the features of the aligned point cloud image information and the features of the surface spectral image data to obtain a multimodal feature vector.

6. The intelligent detection system for sports shoe appearance defects based on few-sample learning according to claim 5, characterized in that, The multimodal feature vectors are input into a pre-set meta-knowledge graph model to obtain the output geometric similarity and spectral similarity. The sum of the geometric similarity and spectral similarity multiplied by their respective preset weight coefficients is set as the comprehensive similarity. The comprehensive similarity is then judged to see if it exceeds the preset comprehensive similarity threshold. If it does, the second defect type is output; otherwise, an artificial signal is generated.

7. The intelligent detection system for sports shoe appearance defects based on few-sample learning according to claim 6, characterized in that, A consistency comparison analysis is performed on the first defect type and the second defect type. If the first defect type and the second defect type are consistent, an output signal is generated; otherwise, a self-test signal is generated.

8. The intelligent detection system for sports shoe appearance defects based on few-sample learning according to claim 1, characterized in that, The quantitative assessment and analysis process for the repair difficulty is as follows: The surface image of the defective area in the sports shoe is obtained, and the surface image is processed into grayscale. The total number of defective pixels within the defect boundary box in the grayscale processed surface image is obtained. The difference between the average gray value in the defective area and the preset average gray value threshold is obtained, and the difference between the average gray value in the defective area and the preset average gray value threshold is set as the pixel gray value difference; The main skeleton of the defect within the defect area is obtained by using a skeleton extraction algorithm (such as Zernike moments), the pixel distance between the two ends of the skeleton is calculated, and the pixel distance between the two ends of the skeleton is set as the defect length. The total number of defective pixels, pixel grayscale difference, and defect length are multiplied by the corresponding set weight coefficients, and the sum of the multiplication is set as the repair difficulty coefficient. The repair difficulty coefficient is then judged to see if it exceeds the preset repair difficulty coefficient threshold. If it does, a difficult repair signal is generated; otherwise, a repairable signal is generated.

9. A few-shot learning-based intelligent detection system for sports shoe appearance defects, wherein the method is applied to any one of claims 1-8, characterized in that, Its characteristic is that it includes the following steps: Step 1: Retrieving and storing the surface feature image set of the athletic shoe; Step 2: Defect identification and localization analysis based on surface feature image set Step 3: Pixel-level feature comparison and analysis process based on bounding box coordinate accuracy verification; Step 4: Coordinate transformation and defect identification analysis of bounding box coordinates based on information progression; Step 5: Quantitatively assess and analyze the difficulty of repairing the surface images of the defective areas in the sneakers.

Images