Method and system for detecting surface flaws of a v-belt based on visual inspection

Abstract

The application provides a visual detection-based surface flaw detection method and system of a poly-V belt, and belongs to the technical field of visual detection. The visual detection-based surface flaw detection method of the poly-V belt comprises the following steps: constructing a geometric structure model of a target poly-V belt, dividing a poly-V belt image into regions based on the geometric structure model, and determining a wedge tooth region, a tooth valley region, and a transition region; collecting multi-view high-resolution image data of the target poly-V belt under a preset illumination condition, and performing illumination normalization and background suppression processing on the image data to obtain a background suppression image; performing feature extraction on the processed background suppression image based on a deep learning network to generate a first feature map representing local texture abnormalities and a second feature map representing structure deformation; fusing the first feature map and the second feature map to obtain a fused feature map; based on the fused feature map, adopting a preset defect type discrimination rule to locate and classify flaws on the surface of the target poly-V belt, and outputting flaw type, coordinate position, and severity information.

Description

Technical Field

[0001] This invention belongs to the field of visual inspection technology, and in particular to a method and system for detecting surface defects on multi-wedge belts based on visual inspection. Background Technology

[0002] As the automotive industry continues to demand higher reliability and quietness from transmission systems, multi-ribbed belts are increasingly used as key power transmission components. Their manufacturing process involves multiple complex steps such as mixing, calendering, molding, and vulcanization. Even a slight deviation in any step can cause defects on the surface of the finished product, placing higher demands on traditional inspection methods. Although some companies have introduced manual visual inspection or basic photoelectric sensing devices for surface quality control, the accuracy and stability of identifying minute defects still need to be improved under high-speed continuous production conditions.

[0003] In recent years, the integration of machine vision and deep learning algorithms has brought new technical paths to industrial surface defect detection. With the trend of intelligent manufacturing and full-process quality traceability of rubber products, the demand for real-time and high-resolution evaluation of the surface condition of multi-wedge belts has become increasingly prominent. If there is uneven dispersion of rubber components during the mixing stage of multi-wedge belts, it is easy to cause local stress concentration during subsequent calendering or molding. Deviations in temperature and pressure control during vulcanization may lead to distortion of wedge tooth profile or micro-bubble residue. Dimensional deviations in post-processing directly affect the meshing and matching of the belt body and wheel groove.

[0004] Currently, due to the periodic wedge-shaped structure, material reflectivity, and complex texture of multi-wedge belts, conventional image processing methods are significantly interfered with in distinguishing real defects from structural shadows and edge reflections, resulting in a high misjudgment rate. At the same time, existing detection systems lack the ability to effectively characterize potential debonding defects at the fabric-rubber interface, making it difficult to fully identify some early hidden dangers before leaving the factory, affecting product service life and the overall reliability of the transmission system. Summary of the Invention

[0005] This invention provides a method for detecting surface defects on multi-wedge belts based on visual inspection, thereby solving the above-mentioned problems.

[0006] The first aspect of this invention provides a method for detecting surface defects on multi-wedge belts based on visual inspection, comprising:

[0007] Construct a geometric structure model of the target multi-wedge band, and divide the multi-wedge band image into regions based on the geometric structure model to determine the wedge tooth region, tooth valley region and transition region;

[0008] High-resolution multi-view image data of the target multi-wedge strip under preset lighting conditions are acquired, and the image data is subjected to illumination normalization and background suppression processing to obtain a background-suppressed image.

[0009] Based on a deep learning network, feature extraction is performed on the processed background suppression image to generate a first feature map representing local texture anomalies and a second feature map representing structural deformation.

[0010] The first feature map and the second feature map are fused to obtain a fused feature map.

[0011] Based on the fused feature map, a preset defect type discrimination rule is used to locate and classify defects on the target multi-wedge surface, and output the defect type, coordinate position and severity information.

[0012] Optionally, a geometric model of the target multi-wedge band is constructed, and the multi-wedge band image is divided into regions based on the geometric model to determine the wedge tooth region, tooth valley region, and transition region, including:

[0013] A parametric geometric model is established based on the standard cross-sectional profile parameters of the multi-wedge belt.

[0014] The parameterized geometric structure model is matched with the acquired multi-wedge strip image to determine the centerline position of each wedge tooth in the image.

[0015] Using the center line of each wedge tooth as a reference, a first predetermined offset is set along the direction perpendicular to the belt running direction to delineate the boundary of the wedge tooth area;

[0016] A second predetermined offset is set between two adjacent wedge tooth areas to define the boundary of the tooth valley area;

[0017] The area between the wedge tooth region and the tooth valley region is defined as the transition zone.

[0018] Optionally, high-resolution image data of the target multi-wedge strip under preset illumination conditions is acquired from multiple perspectives, and the image data is subjected to illumination normalization and background suppression processing to obtain a background-suppressed image, including:

[0019] A composite lighting system consisting of a ring diffuse light source and a coaxial directional light source was set up to acquire images under two lighting modes.

[0020] The illumination component estimation based on Retinex theory is performed on the image of the ring diffuse light source, and the reflection component is separated as the texture base image.

[0021] Edge enhancement filtering is applied to the coaxial directional light source image to highlight the edge enhancement image with contrast between the wedge tooth contour and the potential debonding area;

[0022] The texture base image and the edge enhancement image are weighted and fused to generate a preliminary processed image;

[0023] A background modeling method based on morphological opening and closing operations is used to remove periodic background interference caused by the fabric substrate from the preliminary processed image, resulting in a background-suppressed image.

[0024] Optionally, feature extraction is performed on the processed background suppression image based on a deep learning network to generate a first feature map representing local texture anomalies and a second feature map representing structural deformation, including:

[0025] Construct a two-branch convolutional neural network architecture, where the first branch is configured as a texture anomaly detection subnetwork and the second branch is configured as a geometric deformation detection subnetwork;

[0026] The background suppression image is input into the texture anomaly detection subnetwork, and high-frequency texture perturbation features are extracted through multi-layer convolution and residual connections to output the first feature map.

[0027] The background suppression image is simultaneously input into the geometric deformation detection sub-network, which includes a spatial transformation module and a key point regression head, used to predict the deviation between the ideal and actual positions of the wedge tooth profile and output a second feature map.

[0028] The texture anomaly detection subnetwork is trained using a dataset containing samples of bubble, impurity, and scratch types.

[0029] The geometric deformation detection subnetwork is trained using a dataset containing samples of wedge tooth distortion, dimensional deviation, and uneven tooth pitch.

[0030] Optionally, the first feature map and the second feature map are fused to obtain a fused feature map, including:

[0031] Channel attention weighting is applied to the first feature map to highlight high-response regions;

[0032] Spatial attention modulation is applied to the second feature map to enhance areas of significant deformation;

[0033] The two feature maps processed by the attention mechanism are concatenated along the channel dimension to form a joint feature tensor;

[0034] The joint feature tensor is input into a lightweight fusion convolutional layer to generate a single-channel fusion feature map, whose pixel value represents the overall defect confidence.

[0035] Optionally, based on the fused feature map, a preset defect type discrimination rule is used to locate and classify defects on the target multi-wedge surface, and the output defect type, coordinate location, and severity information include:

[0036] Threshold segmentation is performed on the fused feature map to generate a binarized defect mask;

[0037] Based on the defect mask, connected components are extracted as candidate defect regions;

[0038] For each candidate defect area, calculate its attribution label in the wedge tooth area, tooth valley area, or transition area;

[0039] The corresponding defect discrimination sub-rule set is invoked based on the attribution label: if it is located in the wedge tooth area, it is first determined whether it is contour distortion or microcrack; if it is located in the tooth valley area, it is first determined whether it is debonding or air bubbles; if it is located in the transition area, it is determined whether it is fiber exposure or indentation.

[0040] The severity level is determined by combining the area, shape factor, and peak response intensity of the candidate defect region in the fused feature map.

[0041] The defect type, coordinate location, and severity level are encapsulated into a structured inspection result and output.

[0042] A second aspect of the present invention provides a vision-based multi-wedge surface defect detection system, comprising:

[0043] The geometric modeling module is used to construct the geometric structure model of the target multi-wedge band, and to divide the multi-wedge band image into regions based on the geometric structure model, determining the wedge tooth region, tooth valley region and transition region;

[0044] The image acquisition and preprocessing module is used to acquire multi-view high-resolution image data of the target multi-wedge strip under preset lighting conditions, and to perform illumination normalization and background suppression processing on the image data.

[0045] The feature extraction module is used to extract features from the processed image data based on a deep learning network, and generate a first feature map representing local texture anomalies and a second feature map representing structural deformation.

[0046] The feature fusion module is used to fuse the first feature map and the second feature map to obtain a fused feature map.

[0047] The defect discrimination module is used to locate and classify defects on the target multi-wedge surface based on the fused feature map and using preset defect type discrimination rules, and output defect type, location and severity information.

[0048] A computer device provided in a third aspect of the present invention includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0049] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions, when executed by a processor, implement the method described above.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] By constructing a region partitioning mechanism based on multi-wedge belt geometric priors, the interference of periodic structures on defect identification is effectively decoupled. The use of composite illumination and multi-view imaging strategies significantly improves the characterization ability of hidden defects such as debonding at the fabric-rubber interface. A dual-branch deep learning architecture is introduced to handle texture anomalies and geometric deformations separately, avoiding the performance trade-offs of a single model in multi-task learning. Combining region-based defect discrimination rules makes the classification logic more closely aligned with typical failure modes in multi-wedge belt manufacturing processes, thus achieving a balance between high recall and low false alarm rate in high-speed production line environments. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart illustrating the steps of the vision-based multi-wedge belt surface defect detection method provided by this invention;

[0054] Figure 2 A schematic diagram of the processing flow for multi-wedge band image region segmentation and geometric structure modeling;

[0055] Figure 3 The structural block diagram of the vision-based multi-wedge belt surface defect detection system provided by the present invention is shown. Detailed Implementation

[0056] Based on the differences in the surface geometric features and optical response of multi-wedge belts, this invention decomposes the complex problem of surface defect detection into four solvable technical steps: region division, illumination normalization, dual-channel feature extraction, and fusion discrimination, thereby achieving high-precision identification and classification of surface defects in multi-wedge belts.

[0057] Please see Figures 1-2 The present invention provides a method for detecting surface defects on multi-wedge belts based on visual inspection, comprising:

[0058] Step 101: Construct a geometric structure model of the target multi-wedge band, and divide the multi-wedge band image into regions based on the geometric structure model to determine the wedge tooth region, tooth valley region and transition region.

[0059] Further, step 101 may include the following sub-steps:

[0060] A parametric geometric model is established based on the standard cross-sectional profile parameters of the multi-wedge belt.

[0061] The parameterized geometric structure model is matched with the acquired multi-wedge strip image to determine the centerline position of each wedge tooth in the image.

[0062] Using the center line of each wedge tooth as a reference, a first predetermined offset is set along the direction perpendicular to the belt running direction to delineate the boundary of the wedge tooth area;

[0063] A second predetermined offset is set between two adjacent wedge tooth areas to define the boundary of the tooth valley area;

[0064] The area between the wedge tooth region and the tooth valley region is defined as the transition zone.

[0065] In this embodiment of the invention, the standard cross-sectional profile parameters of the target multi-wedge band are obtained, including the wedge tooth height. Tooth tip width Tooth root width and pitch A parametric geometric model of the target multi-wedge belt was constructed using computer-aided design software.

[0066] The model is represented on the image plane as a repetitive periodic wedge pattern along the direction of the belt's movement. This parameterized geometric model is then subjected to normalized cross-correlation template matching with the acquired multi-wedge belt images to determine the pixel coordinate sequence of the center lines of each wedge tooth in the image. Using each centerline as a reference, a first predetermined offset is set along the direction perpendicular to the belt's running direction. Define the boundaries of the wedge tooth region; set a second predetermined offset between two adjacent wedge tooth regions. The boundary of the tooth valley region is defined; the area between the wedge tooth region and the tooth valley region is defined as the transition region. Through the above geometric constraints, the entire image is divided into three functional regions with different failure mode tendencies.

[0067] Step 102: Collect multi-view high-resolution image data of the target multi-wedge strip under preset lighting conditions, and perform illumination normalization and background suppression processing on the image data to obtain a background-suppressed image.

[0068] Furthermore, step 102 may include the following sub-steps:

[0069] A composite lighting system consisting of a ring-shaped diffuse light source and a coaxial directional light source was set up to acquire images under two different lighting modes.

[0070] In this embodiment of the invention, the composite lighting system consists of a ring-shaped LED diffuse light source and a coaxial cold light source, which are triggered alternately to synchronously control an industrial camera array for image capture. The ring-shaped diffuse light source is used to uniformly illuminate the entire surface and suppress specular reflection; the coaxial cold light source illuminates along the optical axis to enhance the contrast between the surface microstructure and the potential debonding area.

[0071] The illumination component is estimated based on Retinex theory for the image of the ring diffuse light source, and the reflection component is separated as the texture base image.

[0072] In this embodiment of the invention, the single-scale Retinex algorithm is used to process the image of the annular diffuse light source. Processing is carried out, among which The image pixel coordinates and their reflection components The calculation formula is:

[0073]

[0074] in Standard deviation Gaussian kernel function, This represents the convolution operation. The result... As a base image for texture, it retains high-frequency abnormal information such as impurities and scratches from the material itself.

[0075] Edge enhancement filtering is applied to the coaxial directional light source image to enhance the edge image by highlighting the contrast between the wedge tooth contour and the potential debonding area.

[0076] In this embodiment of the invention, the image of the coaxial directional light source... Applying the Laplacian sharpening operator yields an edge-enhanced image. :

[0077]

[0078] in, This is the Laplacian operator, used to detect edges and abrupt changes in images. This operation enhances the local brightness abrupt changes caused by debonding at wedge edges and fabric-rubber interfaces.

[0079] The texture base image and the edge enhancement image are weighted and fused to generate a preliminary processed image.

[0080] In this embodiment of the invention, a weighting coefficient is set. , ,right and Perform linear weighted fusion:

[0081]

[0082] This fusion operation combines high-frequency detail information such as impurities and scratches in the material itself in the texture base image with edge features of the wedge tooth contour and debonding area in the edge enhancement image, providing a pre-processed image with richer information and clearer contrast for subsequent defect detection and recognition.

[0083] A background modeling method based on morphological opening and closing operations is used to remove periodic background interference caused by the fabric substrate from the preliminary processed image, resulting in a background-suppressed image.

[0084] In this embodiment of the invention, firstly... Execution radius is The background estimation image is obtained by performing a morphological opening operation on each pixel and then a closing operation with the same radius. Final background suppression image Obtained from the following formula:

[0085]

[0086] This step effectively eliminates background interference caused by the periodic texture of the fabric substrate, significantly improves the signal-to-noise ratio between the defect area and the background, and fully preserves the feature information of abnormal defects such as scratches, delamination, and impurities, providing a high-quality image foundation for subsequent defect detection and identification.

[0087] Step 103: Based on the deep learning network, perform feature extraction on the processed background suppression image to generate a first feature map representing local texture anomalies and a second feature map representing structural deformation.

[0088] In this embodiment of the invention, a dual-branch convolutional neural network architecture is constructed, wherein the first branch is configured as a texture anomaly detection subnetwork and the second branch is configured as a geometric deformation detection subnetwork.

[0089] Background suppression image Two sub-networks are input simultaneously. The texture anomaly detection sub-network consists of five convolutional layers and residual connections, and outputs a first feature map with 64 channels. Its highly activated regions correspond to local texture perturbations such as bubbles, impurities, or scratches. The geometric deformation detection subnetwork includes a Spatial Transformer Module and a keypoint regression head, which are used to predict the Euclidean distance deviation between the ideal contour position and the actual detection position of each wedge tooth, and output a single-channel second feature map. Its pixel value is proportional to the degree of local geometric distortion.

[0090] Furthermore, step 103 may include the following sub-steps:

[0091] Construct a two-branch convolutional neural network architecture, where the first branch is configured as a texture anomaly detection subnetwork and the second branch is configured as a geometric deformation detection subnetwork;

[0092] The background suppression image is input into the texture anomaly detection subnetwork, and high-frequency texture perturbation features are extracted through multi-layer convolution and residual connections to output the first feature map.

[0093] The background suppression image is simultaneously input into the geometric deformation detection sub-network, which includes a spatial transformation module and a key point regression head, used to predict the deviation between the ideal and actual positions of the wedge tooth profile and output a second feature map.

[0094] The texture anomaly detection subnetwork is trained using a dataset containing samples of bubbles, impurities, scratches, and fibers.

[0095] The geometric deformation detection subnetwork is trained using a dataset containing samples of wedge tooth distortion, dimensional deviation, and uneven tooth pitch.

[0096] Step 104: Perform feature fusion between the first feature map and the second feature map to obtain a fused feature map.

[0097] Furthermore, step 104 may include the following sub-steps:

[0098] Channel attention weighting is applied to the first feature map to highlight high-response regions.

[0099] In this embodiment of the invention, the Squeeze-and-Excitation (SE) attention mechanism is used to apply attention to the first feature map. The 64 channels were recalibrated to generate a weighted feature map. .

[0100] Spatial attention modulation is applied to the second feature map to enhance regions with significant deformation.

[0101] In this embodiment of the invention, the second feature map The spatial attention module is applied to generate spatial weight masks through global average pooling and convolution. and calculate ,in This indicates element-wise multiplication.

[0102] Through spatial attention mechanisms, the model can automatically focus on areas in an image that contain key information, suppressing background noise and thus improving the accuracy of subsequent defect detection and recognition.

[0103] The two feature maps processed by the attention mechanism are concatenated along the channel dimension to form a joint feature tensor.

[0104] In this embodiment of the invention, (64 channels) and (1 channel) Concatenate along the channel dimension to obtain a 65-channel joint feature tensor. .

[0105] The joint feature tensor is input into a lightweight fusion convolutional layer to generate a single-channel fusion feature map, whose pixel value represents the overall defect confidence.

[0106] In this embodiment of the invention, a Convolutional layer pairs Dimensionality reduction is performed to output a single-channel fused feature map. This significantly reduces the computational load and network parameter count in subsequent processing, ensuring the algorithm's operational efficiency. Its pixel value range is normalized to... The higher the pixel value, the greater the probability that there is a defect at the corresponding location in the image, that is, the higher the overall defect confidence, thus realizing the quantitative characterization and precise positioning of defect areas.

[0107] Step 105: Based on the fused feature map, use the preset defect type discrimination rules to locate and classify defects on the target multi-wedge surface, and output defect type, coordinate position and severity information.

[0108] Furthermore, step 105 may include the following sub-steps:

[0109] Threshold segmentation is performed on the fused feature map to generate a binarized defect mask.

[0110] In this embodiment of the invention, a segmentation threshold is set. ,right Binarization is performed to obtain the defect mask. .

[0111] Based on the defect mask, connected components are extracted as candidate defect regions.

[0112] In this embodiment of the invention, an eight-neighbor connected component labeling algorithm is used to... Extract all independent defect candidate regions .

[0113] For each candidate defect area, calculate its affiliation label in the wedge tooth area, tooth valley area, or transition area.

[0114] In this embodiment of the invention, each defect candidate region is calculated. centroid coordinates Calculate its corresponding centroid coordinates The centroid coordinates are used to accurately characterize the spatial location of the defect candidate region in the image, providing a core localization basis for subsequent region attribution determination. The region segmentation mask generated in step 101 is queried, and the centroid coordinates are used as the basis for further analysis. The candidate region for the defect is determined by matching the corresponding pixel position in the region segmentation mask. Functional area label Determine its corresponding functional area label. .

[0115] This step achieves one-to-one matching between defect candidate areas and functional areas through centroid localization, providing a data foundation for setting differentiated defect judgment thresholds and classification standards for different functional areas. It can effectively avoid detection misjudgments caused by differences in texture and structure in different areas and improve the accuracy of defect detection.

[0116] The corresponding defect discrimination sub-rule set is invoked based on the attribution label: if it is located in the wedge tooth area, it is first determined whether it is contour distortion or microcrack; if it is located in the tooth valley area, it is first determined whether it is debonding or air bubbles; if it is located in the transition area, it is determined whether it is fiber exposure or indentation.

[0117] In this embodiment of the invention, three sets of defect discrimination sub-rules are preset. For In the area, if The mean in this region is greater than 0.6 and If the variance is less than 0.1, it is judged as contour distortion; if The appearance of thin, elongated, bright stripes indicates a microcrack. For In the area, if and If all show a circular high response, it is determined to be a bubble; if only Significant response If the flow is smooth, it is considered debonding. For In the area, if If linear highlights are observed accompanied by a sudden drop in local brightness, it is determined that the fibers are exposed.

[0118] The severity level is determined by combining the area, shape factor, and peak response intensity of the candidate defect region in the fused feature map.

[0119] In this embodiment of the invention, a severity score is defined. for:

[0120]

[0121] in, For the area, The maximum permissible defect area (e.g., 5 mm²). , , , To fuse feature maps in defect candidate regions The maximum peak response intensity within, This is the shape factor term.

[0122] Based on the above severity scores, a grading threshold is set to complete the defect level classification: when When the value is less than 0.3, the defect is judged as a minor defect; when 0.3 ≤ When the value is less than 0.7, the defect is classified as a moderate defect; when... If the value is ≥0.7, the defect is considered a severe defect.

[0123] The defect type, coordinate location, and severity level are encapsulated into a structured inspection result and output.

[0124] Step 106: Based on the structured detection results, trigger the production line graded response mechanism.

[0125] In this embodiment of the invention, after receiving the structured inspection results, the central controller performs corresponding operations according to the severity level of the defects: for severe defects, the production line is stopped immediately and the defective product is marked as scrap; for moderate defects, batch information is recorded and a re-inspection process is initiated; for minor defects, the defect is only written to the quality traceability database without interrupting production.

[0126] This proposal presents a region-aware visual detection method for surface defects on multi-wedge tapes. The invention effectively decouples the interference of periodic structures on defect identification by constructing a region segmentation mechanism based on multi-wedge tape geometric priors. It employs a composite illumination and multi-view imaging strategy to significantly improve the characterization ability of hidden defects such as debonding at the fabric-rubber interface. A dual-branch deep learning architecture is introduced to handle texture anomalies and geometric deformations separately, avoiding the performance trade-offs of a single model in multi-task learning. Combined with defect discrimination rules based on region attribution, the classification logic better aligns with typical failure modes in multi-wedge tape manufacturing processes, thereby achieving a balance between high recall and low false alarm rate in high-speed production line environments.

[0127] Please see Figure 3 , Figure 3 This is a structural block diagram of the multi-wedge belt surface defect detection system based on visual inspection provided in Embodiment 3 of the present invention.

[0128] The present invention provides a vision-based multi-wedge belt surface defect detection system, comprising:

[0129] The geometric modeling module 301 is used to construct the geometric structure model of the target multi-wedge band, and to divide the multi-wedge band image into regions based on the geometric structure model, and to determine the wedge tooth region, tooth valley region and transition region.

[0130] During the region segmentation phase, the geometric modeling module 301 constructs a parametric model based on pre-stored multi-wedge belt model parameters (e.g., H=3.5mm, P=4mm), and locates the center lines of each wedge tooth in the original image through normalized cross-correlation matching. Since the multi-wedge belt may exhibit ±2° lateral sway during operation, the system employs a sliding window strategy to search for the maximum correlation peak along the bandwidth direction to ensure robust center line positioning. The defined wedge tooth region ( =0.35H≈1.2mm), tooth valley area ( =0.2H≈0.7mm) and the transition zone, each corresponding to a different failure mechanism: for example, the tooth valley zone is prone to bubble formation due to poor venting during vulcanization, while the transition zone is prone to fiber exposure due to concentrated calendering tension. This process-a priori regional decoupling allows subsequent discrimination rules to specifically suppress false alarms, such as excluding normal fabric texture fluctuations in the tooth valley zone from the debonding judgment.

[0131] The image acquisition and preprocessing module 302 is used to acquire multi-view high-resolution image data of the target multi-wedge strip under preset illumination conditions, and to perform illumination normalization and background suppression processing on the image data to obtain a background-suppressed image.

[0132] At the online inspection station of the multi-wedge belt continuous production line, the multi-wedge belt to be inspected passes through the inspection area at a constant speed of 1.5m / s. At this time, the ring LED diffuse light source in the composite light source unit is triggered first, and the industrial camera array synchronously acquires the first high-resolution image. Subsequently, the coaxial cold light source is triggered, and the camera array acquires the second image. The alternating triggering cycle is controlled within 20ms to ensure that the spatial position deviation between the two images is less than 0.1mm under high-speed operation, thereby meeting the subsequent pixel-level alignment requirements. This design effectively solves the problem of multi-view image misalignment caused by belt movement and provides spatial consistency guarantee for subsequent dual-channel feature extraction.

[0133] The acquired ring light source image is sent to the image processor, where a single-scale Retinex algorithm is executed. The Gaussian kernel standard deviation is used. The setting of =80 is determined based on the spectral analysis results of typical fabric texture periods (approximately 3–5 mm) and common impurity sizes (0.2–1 mm) of multi-ribbed tape: larger This value can effectively filter out low-frequency uneven lighting while retaining high-frequency defect information. The resulting separated reflection component... As the base image for texture, its grayscale dynamic range is compressed to the [0,1] range, which significantly reduces the overexposed areas caused by specular reflection on the rubber surface, allowing micro-scratches or fiber impurities that were originally covered by strong light to be revealed.

[0134] The feature extraction module 303 is used to extract features from the processed background suppression image based on a deep learning network, and generate a first feature map representing local texture anomalies and a second feature map representing structural deformation.

[0135] The dual-branch network in the feature extraction module 303 is deployed in an embedded AI inference unit, such as the NVIDIA Jetson AGX Orin. Its texture anomaly subnetwork adopts a shallow ResNet-18 structure. After training on a dataset containing 120,000 labeled samples, it achieves a recall rate of 96.2% for impurities or scratches larger than 0.3 mm². The geometric deformation subnetwork performs affine correction on the input image through a spatial transformation module, and then outputs the predicted coordinates of each wedge tooth vertex by the keypoint regression head. The Euclidean distance between these coordinates and the ideal model coordinates is mapped to the second feature map. When a wedge tooth collapses by more than 0.15 mm at the tooth tip due to wear of the vulcanizing mold, the second feature image... If the response value in this region exceeds the 0.6 threshold, it is effectively identified as a contour distortion.

[0136] Feature fusion module 304 is used to fuse the first feature map and the second feature map to obtain a fused feature map;

[0137] During the feature fusion stage, the SE attention mechanism focuses on the first feature map. After global pooling of the 64 channels, channel weight vectors are generated through two fully connected layers to suppress texture feature channels irrelevant to the current sample. For example, for black rubber strips, the response to red impurity channels is automatically reduced. The spatial attention module then focuses on the second feature map. A 7×7 convolution is performed to generate a spatial mask, enhancing regions with significant local deformation. The joint tensor resulting from the concatenation of these two masks is then subjected to 1×1 convolution for dimensionality reduction, and the output... The image not only integrates texture and deformation information, but also achieves selective feature enhancement through an attention mechanism, which improves the signal-to-noise ratio of the final fused image by about 40%, laying the foundation for subsequent accurate segmentation.

[0138] The defect discrimination module 305 is used to locate and classify defects on the target multi-wedge belt surface based on the fused feature map and using preset defect type discrimination rules, and output defect type, coordinate position and severity information.

[0139] The defect identification module 305 calls the corresponding rule set based on the region affiliation. For example, when a candidate region... Located in the tooth valley area and The mean is >0.5. When the variance is less than 0.08, the system classifies it as debonding rather than bubble formation. This is because debonding mainly causes localized deformation while maintaining surface texture continuity, while bubbles simultaneously disturb both texture and deformation. This logic stems from statistical analysis of real defect samples, significantly improving the classification accuracy of debonding defects. Severity scoring... In the data, the area component accounts for 40%, reflecting the risk of defect expansion, while the shape factor component... The peak response intensity reflects the stress concentration effect at the crack tip, while the peak response intensity characterizes the defect depth or protrusion degree. The weighted combination of the three is more in line with the engineering failure assessment criteria.

[0140] Ultimately, the central controller... Value triggers graded response: When severe debonding is detected ( When the defect rate is ≥0.7, the PLC immediately sends an emergency stop signal to the traction roller motor and marks "REJECT" on the surface of the belt using an inkjet printer; for moderate defects, the MES system is triggered to record the batch number, timestamp, and defect coordinates for quality engineers to trace and analyze; for minor defects, the defect is only written to the database for SPC process control. This closed-loop mechanism reduces the scrap rate of the production line and avoids capacity loss caused by excessive downtime.

[0141] A computer device provided in a third aspect of the present invention includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0142] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions, when executed by a processor, implement the method described above.

[0143] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed between each other can be through some interfaces, indirect coupling or communication connection between devices or units, and can be electrical, mechanical, or other forms. The above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0144] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

Claims

1. A method for detecting surface defects on multi-wedge belts based on visual inspection, characterized in that, include: Construct a geometric structure model of the target multi-wedge band, and divide the multi-wedge band image into regions based on the geometric structure model to determine the wedge tooth region, tooth valley region and transition region; High-resolution multi-view image data of the target multi-wedge strip under preset lighting conditions are acquired, and the image data is subjected to illumination normalization and background suppression processing to obtain a background-suppressed image. Based on a deep learning network, feature extraction is performed on the processed background suppression image to generate a first feature map representing local texture anomalies and a second feature map representing structural deformation. The first feature map and the second feature map are fused to obtain a fused feature map. Based on the fused feature map, a preset defect type discrimination rule is used to locate and classify defects on the target multi-wedge surface, and output the defect type, coordinate position and severity information.

2. The method for detecting surface defects of multi-wedge belts based on visual inspection according to claim 1, characterized in that, A geometric model of the target multi-wedge band is constructed, and the multi-wedge band image is divided into regions based on the geometric model to determine the wedge tooth region, tooth valley region, and transition region, including: A parametric geometric model is established based on the standard cross-sectional profile parameters of the multi-wedge belt. The parameterized geometric structure model is matched with the acquired multi-wedge belt image to determine the centerline position of each wedge tooth in the multi-wedge belt image. Using the center line of each wedge tooth as a reference, a first predetermined offset is set along the direction perpendicular to the belt running direction to delineate the boundary of the wedge tooth area; A second predetermined offset is set between two adjacent wedge tooth areas to define the boundary of the tooth valley area; The area between the wedge tooth region and the tooth valley region is defined as the transition zone.

3. The method for detecting surface defects of multi-wedge belts based on visual inspection according to claim 1, characterized in that, High-resolution multi-view image data of the target multi-wedge strip under preset illumination conditions are acquired, and the image data is subjected to illumination normalization and background suppression processing to obtain a background-suppressed image, including: A composite lighting system consisting of a ring diffuse light source and a coaxial directional light source was set up to acquire high-resolution image data under two lighting modes. The illumination component estimation based on Retinex theory is performed on the image of the ring diffuse light source, and the reflection component is separated as the texture base image. Edge enhancement filtering is applied to the coaxial directional light source image to highlight the edge enhancement image with contrast between the wedge tooth contour and the potential debonding area; The texture base image and the edge enhancement image are weighted and fused to generate a preliminary processed image; A background modeling method based on morphological opening and closing operations is used to remove periodic background interference caused by the fabric substrate from the preliminary processed image, resulting in a background-suppressed image.

4. The method for detecting surface defects of multi-wedge belts based on visual inspection according to claim 1, characterized in that, Based on a deep learning network, feature extraction is performed on the processed background suppression image to generate a first feature map representing local texture anomalies and a second feature map representing structural deformation, including: Construct a two-branch convolutional neural network architecture, where the first branch is configured as a texture anomaly detection subnetwork and the second branch is configured as a geometric deformation detection subnetwork; The background suppression image is input into the texture anomaly detection subnetwork, and high-frequency texture perturbation features are extracted through multi-layer convolution and residual connections to output the first feature map. The background suppression image is simultaneously input into the geometric deformation detection sub-network, which includes a spatial transformation module and a key point regression head, to predict the deviation between the ideal and actual positions of the wedge tooth profile and output a second feature map.

5. The method for detecting surface defects of multi-wedge belts based on visual inspection according to claim 1, characterized in that, The first feature map and the second feature map are fused to obtain a fused feature map, which includes: Channel attention weighting is applied to the first feature map to highlight high-response regions; Spatial attention modulation is applied to the second feature map to enhance areas of significant deformation; The two feature maps processed by the attention mechanism are concatenated along the channel dimension to form a joint feature tensor; The joint feature tensor is input into a lightweight fusion convolutional layer to generate a single-channel fusion feature map, whose pixel value represents the overall defect confidence.

6. The method for detecting surface defects of multi-wedge belts based on visual inspection according to claim 1, characterized in that, Based on the fused feature map, a preset defect type discrimination rule is used to locate and classify defects on the target multi-wedge surface, and the output defect type, coordinate location, and severity information include: Threshold segmentation is performed on the fused feature map to generate a binarized defect mask; Based on the defect mask, connected components are extracted as candidate defect regions; For each candidate defect area, calculate its attribution label in the wedge tooth area, tooth valley area, or transition area; The corresponding defect discrimination sub-rule set is invoked based on the attribution label: if it is located in the wedge tooth area, it is first determined whether it is contour distortion or microcrack; if it is located in the tooth valley area, it is first determined whether it is debonding or air bubbles; if it is located in the transition area, it is determined whether it is fiber exposure or indentation. The severity level is determined by combining the area, shape factor, and peak response intensity of the candidate defect region in the fused feature map. The defect type, coordinate location, and severity level are encapsulated into a structured inspection result and output.

7. A vision-based multi-wedge belt surface defect detection system, comprising: The geometric modeling module is used to construct the geometric structure model of the target multi-wedge band, and to divide the multi-wedge band image into regions based on the geometric structure model, determining the wedge tooth region, tooth valley region and transition region; The image acquisition and preprocessing module is used to acquire multi-view high-resolution image data of the target multi-wedge strip under preset lighting conditions, and to perform illumination normalization and background suppression processing on the image data. The feature extraction module is used to extract features from the processed image data based on a deep learning network, and generate a first feature map representing local texture anomalies and a second feature map representing structural deformation. The feature fusion module is used to fuse the first feature map and the second feature map to obtain a fused feature map. The defect discrimination module is used to locate and classify defects on the target multi-wedge surface based on the fused feature map and using preset defect type discrimination rules, and output defect type, location and severity information.

8. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method as described in any one of claims 1-6.

9. A computer-readable medium, characterized in that, It stores a computer program thereon, wherein the program, when executed by a processor, implements the method as described in any one of claims 1-6.

Images