Inverter shell glue groove internal defect detection system based on computer visual identification

Abstract

The invention provides a system for detecting defects in a glue groove of an inverter shell based on computer visual identification, and relates to the field of intelligent detection.The method comprises the steps that firstly, an acquired glue groove image of the inverter shell is preprocessed to improve the contrast ratio and visibility of a defect area in the image; then positioning and segmenting the preprocessed image to accurately position the ROI of the glue groove, and then carrying out feature coding on the surface of the glue groove to automatically capture fine texture abnormity to enhance the significance of a defect area, so that the defect area and a normal area form an obvious distinguishing boundary in a feature space; and finally, making a decision on the enhanced high-dimensional features to judge whether defects exist or not. Thus, the limitation that a traditional method depends on manual design characteristics can be overcome, tiny glue groove defects can still be stably recognized under the complex illumination condition, and therefore the false detection rate and the omission ratio are reduced, and the detection efficiency and the automation level are improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent detection, and more specifically, to a defect detection system in the glue groove of an inverter housing based on computer vision recognition. Background Art

[0002] With the rapid development of photovoltaic power generation and energy storage technologies, the reliability and sealing of inverters, the core devices of energy conversion, have become particularly important. The adhesive grooves in the inverter housing are crucial for protecting the internal electronic components. The quality of the adhesive filling—including defects such as cracks, bubbles, impurities, or insufficient adhesive—directly affects the device's waterproof performance, shock resistance, and long-term stability.

[0003] However, traditional inspection methods, such as manual visual inspection or automated inspection based on a single image processing algorithm, face many challenges in practical applications. Due to the reflective properties of metal materials and the colloid itself, the surface of the glue tank is prone to mirror reflections and shadow interference, which means that during the image acquisition process, key defect features may be obscured by strong reflective spots or dark areas. For example, a metal shell under direct light source illumination will produce strong reflections. At the same time, inappropriate lighting angles or ambient stray light will also cause shadows to form at the edges of the glue tank. These problems greatly affect the effectiveness of traditional threshold segmentation and edge detection algorithms, resulting in high false detection and missed detection rates. In addition, manual inspection is not only inefficient and highly subjective, but also difficult to adapt to the needs of large-scale production.

[0004] Therefore, an optimized defect detection solution for the inverter housing glue groove is desired, which can overcome complex light interference and enhance the intelligent detection system of defect feature expression, so as to improve the robustness and accuracy of defect identification and ensure the quality control of inverter products. Summary of the Invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a defect detection system for the inverter housing glue groove based on computer vision recognition.

[0006] According to one aspect of the present application, a system for detecting defects in the adhesive groove of an inverter housing based on computer vision recognition is provided, which includes:

[0007] The inverter housing adhesive groove image acquisition module is used to obtain the inverter housing adhesive groove image captured by the camera;

[0008] A pre-processing module, configured to suppress reflections and compensate shadows on the inverter housing adhesive groove image to obtain a pre-processed adhesive groove image;

[0009] A ROI extraction module is used to locate and segment the glue tank area of the pre-processed glue tank image to obtain a glue tank ROI image;

[0010] A feature enhancement module is configured to perform image feature enhancement processing on the glue slot ROI image to obtain a glue slot visual feature enhancement coding map, wherein the feature enhancement module includes: a glue slot visual feature extraction unit, configured to perform convolutional coding-based glue slot visual feature extraction on the glue slot ROI image to obtain a glue slot visual feature coding map; a glue slot visual feature receptive field enhancement unit, configured to perform feature dynamic receptive field significance enhancement on the glue slot visual feature coding map to obtain the glue slot visual feature enhancement coding map;

[0011] The defect judgment module is used to determine whether there is a defect in the glue groove of the inverter housing based on the glue groove visual feature enhancement coding map.

[0012] Compared with the existing technology, the defect detection system for the inverter housing glue groove based on computer vision recognition provided by this application first pre-processes the collected inverter housing glue groove image to improve the contrast and visibility of the defect area in the image, then locates and segments the pre-processed image to accurately locate the glue groove ROI, and then performs feature encoding on the glue groove surface to automatically capture subtle texture anomalies to enhance the significance of the defect area, so that it forms a clear boundary between it and the normal area in the feature space, and finally makes a decision on the enhanced high-dimensional features to determine whether there is a defect. In this way, the limitations of traditional methods that rely on manually designed features can be overcome, so that tiny glue groove defects can still be stably identified under complex lighting conditions, thereby reducing false detection and missed detection rates and improving detection efficiency and automation level. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0014] Figure 1 This is a system block diagram of a defect detection system in an inverter housing glue groove based on computer vision recognition according to an embodiment of the present application.

[0015] Figure 2 Schematic diagram of data flow of a system for detecting defects in the inverter housing rubber groove based on computer vision recognition according to an embodiment of the present application.

[0016] Figure 3 This is a block diagram of a feature enhancement module in a defect detection system in an inverter housing glue groove based on computer vision recognition according to an embodiment of the present application.

[0017] Figure 4 This is a block diagram of a glue groove visual feature receptive field enhancement unit in a defect detection system in a glue groove of an inverter housing based on computer vision recognition according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0019] With the rapid development of photovoltaic power generation and energy storage technology, the reliability and sealing of inverters, as core equipment for energy conversion, are becoming increasingly important. The filling quality of the inverter housing glue groove is directly related to the waterproof performance, seismic resistance and long-term stability of the equipment. However, due to problems such as reflection and shadow interference on the surface of the glue groove, traditional detection methods (such as manual detection or automated detection based on a single image processing algorithm) are difficult to accurately identify colloid defects (such as cracks, bubbles, impurities or insufficient glue). The reflective properties of metal housings and colloids can easily lead to mirror reflections and dark area obscuration, affecting the quality of image acquisition, thereby reducing the detection accuracy of traditional algorithms and resulting in high false detection and missed detection rates. In addition, manual detection is inefficient and highly subjective, making it difficult to meet the needs of large-scale production.

[0020] Based on this, the present application proposes a defect detection system in the inverter housing glue groove based on computer vision recognition. In the image preprocessing stage, the collected inverter housing glue groove image is preprocessed to suppress the mirror reflection of the metal surface and eliminate the interference of reflective spots and dark areas on the image quality, so as to improve the contrast and visibility of the defective area in the image. Subsequently, the preprocessed image is positioned and segmented to accurately locate the glue groove ROI and eliminate the influence of background noise on feature extraction. In the feature expression stage, the glue groove surface is feature encoded to automatically capture subtle texture anomalies (such as cracks and bubbles), and the significance of the defective area is enhanced so that it forms a clear boundary between it and the normal area in the feature space. Finally, a decision is made on the enhanced high-dimensional features to determine whether there is a defect, so as to effectively overcome the defects of traditional algorithms that are sensitive to light and have poor robustness. This application significantly suppresses the interference of reflections and shadows on imaging quality through the coordinated optimization of optics and algorithms. At the same time, it utilizes the adaptive feature extraction capabilities of deep learning to overcome the limitations of traditional methods that rely on manually designed features, so that tiny glue groove defects can still be stably identified under complex lighting conditions, thereby reducing false detection and missed detection rates and improving detection efficiency and automation levels.

[0021] Figure 1 This is a system block diagram of a defect detection system in an inverter housing glue groove based on computer vision recognition according to an embodiment of the present application. Figure 2Schematic diagram of data flow of the inverter housing glue groove defect detection system based on computer vision recognition according to an embodiment of the present application. Figure 1 and Figure 2 As shown, in the inverter housing glue groove defect detection system 100 based on computer vision recognition, it includes: an inverter housing glue groove image acquisition module 110, used to obtain the inverter housing glue groove image captured by the camera; a pre-processing module 120, used to suppress reflections and compensate for shadows on the inverter housing glue groove image to obtain a pre-processed glue groove image; an ROI extraction module 130, used to locate and segment the glue groove area on the pre-processed glue groove image to obtain a glue groove ROI image; a feature enhancement module 140, used to perform image feature enhancement processing on the glue groove ROI image to obtain a glue groove visual feature enhancement coding map; a defect judgment module 150, used to determine whether there is a defect in the inverter housing glue groove based on the glue groove visual feature enhancement coding map.

[0022] In the embodiment of the present application, the inverter housing glue groove image acquisition module 110 is used to obtain an image of the inverter housing glue groove captured by a camera. It should be understood that the inverter housing glue groove image is a visual information carrier captured by the camera that reflects the actual state of the inverter housing glue groove area, specifically including the geometric shape of the glue groove (such as the spatial structural features of the glue groove, edge position, depth, etc.), the texture details of the colloid surface (such as the presence of cracks, the distribution of bubbles, the color and shape characteristics of impurities), the interface bonding state between the colloid and the metal shell (such as the filling fullness presented by whether the glue amount is sufficient, the degree of fit between the colloid and the shell edge), and optical characteristic information such as the distribution area of reflective spots and the shadow coverage area formed due to material reflection and ambient light. In short, by collecting the inverter housing glue groove image, a data foundation can be laid for the subsequent use of computer vision recognition technology to determine whether there are defects such as cracks, bubbles, impurities or insufficient glue in the glue groove.

[0023] In particular, in one embodiment of the present application, the specific implementation process for acquiring an image of the inverter housing's adhesive groove captured by a camera includes: first, selecting a light source based on the inverter housing's metal material and the reflective properties of the adhesive. Either a diffuse light source (such as a uniform diffuser or integrating sphere) or a polarized light source can be used. A diffuse light source can distribute light more evenly across the surface, effectively reducing specular reflections. A polarized light source, when used in conjunction with a polarizer, can specifically eliminate surface reflections, creating stable lighting conditions for image capture. During light source layout, the angle and position of the light source must be carefully adjusted to avoid direct illumination onto the adhesive groove surface. Low-angle side lighting can be considered, as this helps highlight surface textures and defects while minimizing the impact of frontal reflections. Alternatively, a bar or ring light source can be considered. By continuously adjusting the light source's direction, the optimal lighting angle can be found to ensure uniform illumination while highlighting defects. Next, the camera and lens must be selected and installed. Using a polarized camera or installing a polarizer in front of the camera lens can further eliminate residual reflections. Adjust the camera and lens mounting angles to avoid facing directly at reflective metal surfaces to minimize reflection interference. A telecentric lens is also used to effectively minimize perspective distortion, ensure consistent image size, and provide standardized input conditions for subsequent image processing. To reduce interference from external, uncontrolled lighting, a light shield or baffle is placed around the inspection area. This shielding shields ambient stray light, ensuring a stable image acquisition environment and improving the reliability of the acquired image. Finally, enter the image acquisition debugging phase. Turn on the selected light source and fine-tune light parameters (such as brightness and color temperature) based on the actual lighting conditions. Simultaneously adjust camera parameters (such as exposure time, focal length, and aperture) to ensure clear imaging of the adhesive groove area. Place the inverter housing to be inspected in a fixed position and use the camera to capture the adhesive groove. This image is obtained under preliminary optimized lighting conditions, providing basic data support for subsequent data analysis.

[0024] In an embodiment of the present application, the pre-processing module 120 is used to suppress reflections and compensate shadows on the inverter housing glue groove image to obtain a pre-processed glue groove image. Specifically, in an embodiment of the present application, the pre-processing module is used to perform histogram equalization, morphological processing, and homomorphic filtering on the inverter housing glue groove image to obtain the pre-processed glue groove image. Accordingly, considering the reflective properties of the metal surface and colloid, mirror reflection and shadow interference are prone to occur during image acquisition. The high reflectivity of the metal housing to the light source will form a strong reflective spot, covering up the defects such as fine cracks or bubbles on the surface of the glue groove; at the same time, the shadow area formed by the position of the light source or the ambient stray light at the edge of the glue groove may cause the defects to be obscured by the dark area. Improper lighting angles or the presence of ambient stray light will cause shadows to form at the edge of the glue groove and other parts, resulting in dark areas. These shadows and dark areas will also interfere with the quality of the image, making the texture and details of the glue groove surface blurred, affecting the accuracy of defect detection. Therefore, in order to enable the subsequent defect detection algorithm to work more effectively, in the technical solution of the present application, the inverter housing glue groove image is subjected to reflection suppression and shadow compensation to obtain a pre-processed glue groove image. In particular, in the technical solution of the present application, the inverter housing glue groove image is subjected to histogram equalization, morphological processing and homomorphic filtering to obtain the pre-processed glue groove image. Specifically, since the inverter housing glue groove image is affected by the metal material and the reflective properties of the colloid, the grayscale distribution in the image may be uneven, and the grayscale values in some areas are concentrated, resulting in low image contrast and unclear defect features. For example, in areas with reflective spots or shadows, the grayscale value variation range is small, making it difficult to distinguish key defect details. Histogram equalization can redistribute the grayscale values of the image, expand the grayscale dynamic range of the image, and make the grayscale distribution of the image more uniform, thereby enhancing the overall contrast of the image. In addition, there may be some noise points in the glue groove image as well as discontinuous or irregular edges and areas caused by factors such as uneven lighting. These noise and irregular parts will interfere with the subsequent accurate identification and analysis of the glue groove area. Morphological processing (such as corrosion, dilation and other operations) can adjust the shape and structure of objects in the image, remove small noise points, connect broken edges, smooth object boundaries, and make the shape and outline of the glue groove clearer and more regular, providing a more accurate image basis for subsequent operations such as feature extraction. At the same time, the inverter housing glue groove image has mirror reflection and shadow interference, which is essentially due to the irrational distribution of low-frequency and high-frequency components in the image caused by uneven lighting. Mirror reflection mainly manifests as high-frequency noise in the image, while shadow causes changes in the low-frequency components of the image. Homomorphic filtering can separate the illumination component and reflection component of the image. By processing different frequency components, it suppresses high-frequency reflection noise and enhances low-frequency image details, thereby effectively reducing the impact of mirror reflection and shadow on image quality and improving the visual effect of the image.In this way, through the combined effects of histogram equalization, morphological processing, and homomorphic filtering, the overall quality of the inverter housing glue groove image is improved, making the edges, textures, and other features of the glue groove more clearly distinguishable, and enhancing the contrast between defective areas and normal areas. This provides high-quality image data for subsequent accurate positioning of the glue groove area and extraction of its visual features, facilitating more effective defect detection.

[0025] In an embodiment of the present application, the ROI extraction module 130 is used to locate and segment the glue groove area of the pre-processed glue groove image to obtain the glue groove ROI image. Specifically, in an embodiment of the present application, the ROI extraction module is used to: locate and segment the glue groove area of the pre-processed glue groove image based on Mask R-CNN to obtain the glue groove ROI image. Accordingly, considering that the glue groove of the inverter housing usually has an irregular geometric shape, and other structural parts or marks may be distributed on the surface of the metal housing, this makes the precise positioning of the glue groove area a challenge. Traditional image segmentation methods (such as threshold segmentation and edge detection) are easily affected by metal reflections, shadows and background noise, resulting in blurred glue groove boundaries or mis-segmentation. For example, a strong reflective area may be mistakenly judged as part of the glue groove outline, while the shadow of the recessed area may be identified as the internal structure of the glue groove. Such errors will directly affect the accuracy of subsequent feature extraction. In addition, there may be slight differences in the position and angle of the glue groove in industrial scenarios. Traditional algorithms are difficult to adapt to such dynamic changes, resulting in insufficient robustness of the detection system. Therefore, in order to detect the position of the glue tank and accurately outline the outline of the glue tank, thereby providing a high-quality image for subsequent feature extraction, the present application performs glue tank area positioning and segmentation based on Mask R-CNN on the pre-processed glue tank image to obtain the glue tank ROI image. By positioning and segmenting the glue tank area based on Mask R-CNN, the glue tank is separated from the complex image background to obtain the glue tank ROI image, effectively eliminating the interference of background noise on subsequent analysis. In this way, when performing feature extraction and defect detection on the glue tank, the algorithm can focus on the glue tank area, improve the accuracy and reliability of detection, and reduce false detection and missed detection caused by background information interference.

[0026] The following is a detailed description of a specific implementation process of "obtaining the glue slot ROI image by performing Mask R-CNN-based glue slot region positioning and segmentation on the pre-processed glue slot image":

[0027] The first step is training data preparation. To ensure the model can accurately identify the glue slot area, rich, high-quality training data is essential. Specifically, since different inverter models have different glue slot locations and shapes, the captured images should cover this diversity as much as possible, including geometric variations in the glue slots of different housing models and the filling state of the glue. After acquiring the images, professional image annotation tools such as LabelMe and CVAT are used to perform pixel-level mask annotation of the glue slot area. This process requires accurately outlining the glue slot and annotating the bounding box. This creates a training dataset containing images, bounding box coordinates, and mask vectors. To verify the model's generalization ability and avoid overfitting, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The sample distribution is balanced, covering a variety of scenarios, including normal and defective glue slots. To further enhance the model's robustness to factors such as lighting variations and camera angle deviations in industrial environments, data augmentation is also required on the training images. This includes geometric transformations of the image, such as rotating it by a certain angle, scaling it, flipping it horizontally or vertically, adjusting brightness and contrast, and simulating slight lens shake through Gaussian blur.

[0028] Next, we move on to the Mask R-CNN model configuration and training phase. The classic Mask R-CNN architecture is typically used for the model architecture. The backbone network can be ResNet-50 or ResNet-101, depending on computational resources and accuracy requirements. A feature pyramid network is used to integrate multi-scale features, improving the model's detection capabilities for glue grooves of varying sizes. The model output is defined as a single-class detection, specifically focusing on the glue groove region. The classification branch outputs the probability of the presence of a glue groove, the regression branch outputs the offset of the bounding box coordinates, and the mask branch outputs a binary mask of the glue groove region. The resolution of this mask must be aligned with the input image. Regarding training parameter settings, the input image must be preprocessed and resized to an appropriate size, such as 1024×1024 pixels. Mean and standard deviation normalization is performed to meet the model's input requirements. The optimizer can be either Adam or SGD. The initial learning rate is set to an appropriate value, such as 0.001, and cosine annealing or step-wise decay is used to avoid overfitting. The batch size should be set based on GPU memory availability, and the number of iterations is generally set to 20-30 epochs. After each epoch of training, loss values are evaluated on the validation set, including classification loss, regression loss, and mask loss. If the validation loss stops decreasing after three consecutive epochs, training can be terminated early to avoid overfitting.

[0029] Next comes the model validation and optimization phase. The trained model is evaluated using validation set images as input. For localization accuracy, the bounding box is checked to ensure complete coverage of the adhesive slot area, avoiding boundary shifts caused by factors such as metal reflections. For segmentation accuracy, the separation of the adhesive slot from the housing background is verified using a visual mask. Particular attention is paid to areas susceptible to shadows, such as the edges and recesses of the adhesive slot, to ensure mask integrity. For false detections, such as misidentifying metallic reflective areas as adhesive slots or missing adhesive slot portions in edge shadows, detailed analysis is performed and targeted adjustments to the data augmentation strategy are made. This can include increasing samples with appropriate lighting conditions or adjusting model hyperparameters, such as increasing the weight of the mask branch. To verify the model's generalization, a test set is used to simulate extreme production line scenarios, such as single-sided shadows caused by strong sidelights or partial coverage of the adhesive surface by reflective spots. This is done to verify that the model can consistently output a complete adhesive slot ROI. For different models of inverter housings, due to the differences in the shape and position of their glue grooves, if positioning deviation occurs, the generalization of the model can be improved by adding samples of the corresponding model or adjusting the parameters of the ROIAlign layer, such as the pooling size.

[0030] Finally, in the actual application phase, the preprocessed glue tank image is resized to the model input size, such as 1024×1024 pixels, and then fed into the trained Mask R-CNN model via the data loader. The model extracts features through the backbone network and generates candidate regions. The features of the candidate regions are then aligned in the ROIAlign layer. Finally, the classification branch determines whether the candidate region is a glue tank, the regression branch corrects the coordinates of the bounding box, and the mask branch generates a mask for the glue tank region. After obtaining the model output, post-processing is required to generate the ROI. First, non-maximum suppression is used to filter candidate regions with classification scores above a certain threshold. Filtering is performed based on bounding box overlap to retain the single best detection result and avoid duplicate detections. The floating-point mask output by the model is then binarized by setting a threshold, such as 0.5, to convert it into a binary mask, with white areas representing the glue tank and black areas representing the background. Finally, according to the binary mask, the glue groove area is cropped from the preprocessed image to obtain the glue groove ROI image, and the original pixel value must be retained so that its size is consistent with the mask.

[0031] In the embodiment of the present application, the feature enhancement module 140 is used to perform image feature enhancement processing on the glue slot ROI image to obtain a glue slot visual feature enhancement coding map. Specifically, Figure 3 FIG is a block diagram of a feature enhancement module in a defect detection system in an inverter housing glue groove based on computer vision recognition according to an embodiment of the present application. Figure 3As shown, the feature enhancement module 140 includes: a glue slot visual feature extraction unit 141, which is used to perform glue slot visual feature extraction based on convolution coding on the glue slot ROI image to obtain a glue slot visual feature coding map; a glue slot visual feature receptive field enhancement unit 142, which is used to perform feature dynamic receptive field significance enhancement on the glue slot visual feature coding map to obtain the glue slot visual feature enhancement coding map.

[0032] In an embodiment of the present application, the glue tank visual feature extraction unit 141 is used to perform convolutional coding-based glue tank visual feature extraction on the glue tank ROI image to obtain a glue tank visual feature coding map. Specifically, in an embodiment of the present application, the glue tank visual feature extraction unit is used to perform DenseNet-based glue tank visual feature extraction on the glue tank ROI image to obtain the glue tank visual feature coding map. Accordingly, considering that defects on the surface of the glue tank (such as micron-level cracks, tiny bubbles or insufficient glue) often present local, subtle textures or grayscale differences, and the high reflective properties of metal materials and the weak shadow interference that may remain after pretreatment may still cause defect features to be confused with normal areas in the image. Traditional methods rely on artificially designed features (such as edge gradients and texture statistics), which are difficult to adapt to the diversity of defect morphology under complex lighting conditions, and are not sensitive enough to minor anomalies, resulting in limited feature expression capabilities. In addition, the natural textures formed on the glue tank surface due to the colloid filling process (such as flow marks or curing ripples) may be misjudged as defects, further increasing the difficulty of classification. To this end, in the technical solution of the present application, the glue tank ROI image is subjected to DenseNet-based glue tank visual feature extraction to obtain the glue tank visual feature encoding map. In detail, DenseNet directly connects the feature map of each layer in the network to all subsequent layers through a dense connection mechanism, thereby realizing the reuse and fusion of multi-level features. In the glue tank ROI image, the local details captured by the shallow network (such as crack edges, bubble contours) and the global semantic information extracted by the deep network (such as the correlation between the defect area and the surrounding colloids) are densely transmitted and fused to form a robust feature encoding. For example, the local highlight area formed by the residual reflective material on the surface of the glue tank may interfere with the shallow features, but through the densely connected multi-scale feature fusion, the model can distinguish between real defects and reflective artifacts in combination with contextual information. At the same time, the compact structure of DenseNet reduces parameter redundancy and can still effectively learn the essential characteristics of glue tank defects with limited training data. Therefore, through the glue tank visual feature extraction based on DenseNet, the visual features related to the defects, such as the edges of the cracks, the shape of the bubbles, etc., can be accurately captured from the glue tank ROI image. These features are encoded into the glue tank visual feature coding map, providing rich and accurate information for subsequent defect detection, helping to more accurately determine whether there are defects in the glue tank and the type of defects.

[0033] In the embodiment of the present application, the glue slot visual feature receptive field enhancement unit 142 is used to enhance the dynamic receptive field significance of the glue slot visual feature coding map to obtain the glue slot visual feature enhanced coding map. Specifically, Figure 4 FIG. 1 is a block diagram of a glue groove visual feature receptive field enhancement unit in a defect detection system for a glue groove of an inverter housing based on computer vision recognition according to an embodiment of the present application. Figure 4 As shown, the glue slot visual feature receptive field enhancement unit 142 includes: a glue slot visual feature pixel-level extraction secondary subunit 142-1, which is used to extract a set of glue slot visual feature pixel-level initial vectors and the glue slot visual feature pixel-level vectors to be enhanced from the glue slot visual feature coding map; a receptive field size determination secondary subunit 142-2, which is used to perform feature distillation compression on the glue slot visual feature pixel-level vectors to be enhanced to obtain the size of the glue slot visual feature receptive field; a receptive field within the pixel-level initial vector selection secondary subunit 142-3, which is used to select the receptive field based on the glue slot visual feature receptive field. The size of the receptive field is selected from the set of pixel-level initial vectors of the glue groove visual feature local receptive field; the glue groove visual feature significant enhancement secondary sub-unit 142-4 is used to significantly enhance the glue groove visual feature pixel-level vector to be enhanced based on the set of pixel-level initial vectors in the local receptive field of the glue groove visual feature to obtain an enhanced glue groove visual feature pixel-level vector, wherein the enhanced glue groove visual feature pixel-level vector is the channel feature vector of the (i, j)th pixel position of the glue groove visual feature enhancement coding image.

[0034] It should be understood that even after preprocessing and feature extraction, the glue groove visual feature coding map may still have problems with local feature ambiguity or residual interference. Tiny defects on the glue groove surface (such as cracks and bubbles) may show weak or discontinuous texture differences under complex lighting, and the residual noise after metal reflection or shadow preprocessing may mask these differences. In addition, the morphology and scale of defects are diverse (such as elongated cracks and round bubbles), and the traditional feature enhancement method with a fixed receptive field is difficult to adapt to the contextual relevance of different defects, resulting in insufficient significance of some defect features in the coding map, affecting the discrimination accuracy of the subsequent classification model. Therefore, the present application performs feature dynamic receptive field significance enhancement on the glue groove visual feature coding map to obtain the glue groove visual feature enhancement coding map.

[0035] Specifically, the feature encoding map is first channel-decoupled, and the initial feature vector of each pixel is independently extracted to break the coupling between channels and focus on fine-grained information. The feature vector to be enhanced is then subjected to information compression and distillation to remove redundant noise and retain core features, providing a basis for dynamic receptive field prediction. Based on the compressed feature distribution, the network learns the semantic relevance of the defect area (such as the continuity of cracks or the closed contours of bubbles) and dynamically determines the optimal receptive field size for each pixel position (for example, a long and narrow crack requires a long strip receptive field to capture the extension direction, while a bubble requires a circular receptive field to integrate closed edge features). After screening the features within the local receptive field, the target pixel is significantly enhanced by aggregating contextual information (such as surrounding texture and edge continuity), amplifying the difference between the defect feature and the normal area. In this way, the expressive ability of the visual features of the glue groove is further improved, so that these features can more comprehensively and accurately describe the actual condition of the glue groove.

[0036] Specifically, in the embodiment of the present application, the glue groove visual feature pixel-level extraction secondary sub-unit is used to: perform feature decoupling on the glue groove visual feature encoding map along the channel dimension to obtain a set of pixel-level initial vectors of the glue groove visual feature. This process can be expressed as follows:

[0037] F∈R H×W×C

[0038]

[0039] Among them, F is the visual feature encoding map of the glue slot, R is a set of real numbers, H and W are the height and width of each feature matrix of F along the channel dimension, C is the number of channels of F, FeatureDecoupling is feature decoupling, v 1,1 ,v 1,W ,v H,1 …v H,W is the pixel-level initial vector of each glue slot visual feature in the set of pixel-level initial vectors of glue slot visual features;

[0040] The pixel-level initial feature vector at the (i, j)th pixel position is extracted from the set of the glue groove visual feature pixel-level initial vectors as the glue groove visual feature pixel-level vector to be enhanced. This process can be expressed by the formula:

[0041] v tbs =v i,j ∈R C

[0042] Among them, v i,j is the channel feature vector of the (i, j)th pixel position in F, v tbs Is v i,j As the pixel-level vector of visual features of the glue groove to be enhanced.

[0043] It should be understandable that the coupling between channels in the glue groove visual feature encoding map may blur fine-grained defect information (such as the weak texture of cracks or the closed contours of bubbles), resulting in redundant interference of different channel features. By decoupling features along the channel dimension, the channel feature vector of each pixel can be extracted independently, breaking the dependency between channels and forcing the model to focus on the multi-dimensional feature expression of a single pixel. In other words, the set of pixel-level initial vectors of the decoupled glue groove visual features can more finely retain local details (such as the discontinuous edges of cracks or the closed shape of bubbles), reduce the masking of defect features by inter-channel noise coupling, and provide a fine-grained information basis for subsequent pixel-by-pixel processing, which is particularly beneficial for the independent characterization of tiny defects.

[0044] Accordingly, adhesive groove defect detection requires locating abnormal areas pixel by pixel, but traditional methods have difficulty adaptively adjusting the processing granularity for defects of different morphologies (such as long and thin cracks and round bubbles). Extracting the pixel-level vector of the visual features of the adhesive groove to be enhanced with the (i, j)th pixel as the center meets the detection requirements of local anchoring of defects and ensures that the processing process can be carried out around the target pixel. In other words, this step embodies a pixel-centered saliency enhancement mechanism. By anchoring the target position as the reference point for local context aggregation, it avoids feature dilution caused by global processing, and can provide a spatial positioning basis for subsequent dynamic receptive field prediction, especially adapting to the diversity and locality of defect morphology.

[0045] Specifically, in the embodiment of the present application, the receptive field size determines the secondary subunit, which is used to: compress the pixel-level vector of the visual feature of the to-be-enhanced glue tank to obtain the distilled vector of the visual feature of the to-be-enhanced glue tank. This process can be expressed by the formula:

[0046]

[0047] Among them, ||·|| is the Euclidean norm of the calculated vector, ||·|| 2 To calculate the square of the Euclidean norm of a vector, v s is the distilled vector of visual features of the tank to be enhanced;

[0048] Based on the feature distribution spatial structure characteristics of the distilled vector of the visual feature of the to-be-enhanced glue tank, the size of the glue tank visual feature receptive field of the pixel-level vector of the visual feature of the to-be-enhanced glue tank is determined. This process can be expressed as follows:

[0049]

[0050] Where r is v tbs The size of the receptive field of the visual feature of the glue groove.

[0051] It should be understood that the pixel-level vector of the visual features of the glue groove to be strengthened may contain residual noise from illumination or non-significant texture information (such as metal reflection artifacts), and direct use will interfere with the accuracy of the receptive field prediction. Through information compression operations, redundant information (such as high-frequency noise in uniform areas) can be removed, and core features that are strongly correlated with defects (such as the directional gradient of cracks or the closed edge of bubbles) can be retained. In other words, the compressed distilled vector of the visual features of the glue groove to be strengthened has a higher information density, and its distribution characteristics (such as spatial continuity or closure) can more clearly reflect the semantics of the defects (such as cracks require long-range dependence and bubbles require local closure), which can provide a low-noise, high-discriminative feature basis for dynamic receptive field size prediction.

[0052] Accordingly, considering that the traditional fixed receptive field cannot adapt to the contextual relevance of different defects (such as cracks require long strip areas to capture the extension direction, and bubbles require circular areas to integrate closed contours). Based on the spatial structural characteristics of the distilled features (such as the directionality and discreteness of the feature distribution), the model can learn a semantically driven mapping function and dynamically determine the optimal receptive field size. In other words, this step can make the receptive field content-adaptive, such as automatically expanding the lateral receptive field for slender cracks to capture continuity, and shrinking the bubble to a circular area to enhance the closed edge features. This dynamic mechanism significantly improves the pertinence of defect context modeling and can solve the problem of insufficient feature significance caused by the preset receptive field.

[0053] Specifically, in the embodiment of the present application, the secondary sub-unit for selecting pixel-level initial vectors within the receptive field is used to select a set of pixel-level initial vectors within the local receptive field of the glue groove visual feature from the set of pixel-level initial vectors of the glue groove visual feature based on the size of the glue groove visual feature receptive field. This process can be expressed by the formula:

[0054]

[0055] Among them, W is the set of pixel-level initial vectors in the local receptive field of the visual feature of the glue groove, v i-r,j ,v i+r,j ,v i,j ,v i,j+r ,v i+r,j+r are the pixel-level initial vectors in the local receptive field of the glue groove visual feature at the (ir,j)th, (i+r,j)th, (i,j), (i,j+r)th and (i+r,j+r)th pixel positions in W respectively.

[0056] It should be understood that the dynamically determined receptive field needs to accurately cover the semantically relevant areas of the target pixel (such as the texture on both sides of a crack or the gradient edge around a bubble). Centering the (i, j)th pixel, filtering the local feature vector set according to the predicted size ensures that the contextual information matches the target defect morphology. In other words, the resulting pixel-level initial vector set within the local receptive field of the glue groove visual feature constructs a contextual information library that is strongly correlated with the target pixel (such as the extension direction of the crack or the radial edge of the bubble). This provides structural support for subsequent saliency enhancement, avoids noise interference from irrelevant areas, and is particularly beneficial for feature aggregation of weak texture defects in complex backgrounds.

[0057] Specifically, in the embodiment of the present application, the glue groove visual feature significantly enhances the secondary sub-unit, which is used to: perform body-boundary conformal regularization on the set of pixel-level initial vectors within the local receptive field of the glue groove visual feature to determine the first modulation weighting coefficient and the second modulation weighting coefficient; based on the first modulation weighting coefficient and the second modulation weighting coefficient, perform weighted fusion based on attention weight on the pixel-level vector of the glue groove visual feature to be enhanced and the set of pixel-level initial vectors within the local receptive field of the glue groove visual feature to obtain the enhanced glue groove visual feature pixel-level vector. The above process can be expressed by the formula:

[0058]

[0059] Among them, v m,n is the pixel-level initial vector within the local receptive field of the visual feature of the (m, n)th pixel position in W, v T is the scoring weight vector of the visual features of the glue slot, is matrix multiplication, Softmax is the softmax function, θ(v m,n ) is v m,n The corresponding attention weight of the visual feature of the glue groove, α and β are the first modulation weighting coefficient and the second modulation weighting coefficient, respectively, v' i,j It is v i,j The enhanced pixel-level vector of the enhanced glue groove visual feature is the channel feature vector of the (i, j)th pixel position of the glue groove visual feature enhanced encoding image.

[0060] In particular, first, the first modulation weighting coefficient and the second modulation weighting coefficient need to be defined. Specifically, for the visual feature pixel vector v of the glue groove to be enhanced i,j The corresponding pixel-level initial vector v in the local receptive field of the glue groove visual feature m,nFor the set of (m=ir~i+r,n=jr~j+r), in order to enhance the amplification effect of the characteristic components related to saliency and the suppression / weakening effect of the characteristic components related to non-saliency, it is expected that the set of pixel-level initial vectors within the local receptive field of the groove visual feature can have conformal representation, that is, it is expected that there can be a high correspondence between its volume space representation and boundary representation.

[0061] Therefore, we first determine the pixel-level volume space representation vector within the local receptive field of the visual feature of the glue groove as:

[0062]

[0063] The pixel-level boundary representation vector within the local receptive field of the glue groove visual feature is:

[0064]

[0065] Then, by modulating the weighting coefficients α and β, the boundary surface-volume space tensor has the regularity that satisfies the commutation relation, that is, the pixel-level volume space representation vector v within the local receptive field of the visual feature of the glue groove is m,n (3) The pixel-level boundary representation vector v within the local receptive field of the visual feature and the glue groove m,n (2) The spatial two-norm representation of the difference vector tends to the product of coefficients α and β:

[0066] ||v m,n (3) -v m,n (2) ||2=ω×α×β

[0067] where ω is the scaling factor.

[0068] In this way, by appropriately selecting the boundary surface conditions and utilizing the conformal commutativity of the volume space, the regularity specification standard is met, thereby achieving a highly conformal representation fusion of the pixel-level initial vectors within the local receptive field of the glue groove visual feature, and enhancing the context-aware feature saliency expression of the pixel-level vectors of the enhanced glue groove visual feature.

[0069] Since the sum of α and β is 1 when weighted fusion is performed, the first modulation weighting coefficient and the second modulation weighting coefficient can be obtained by combining the above processing.

[0070] It should be understandable that the difference between defects and normal areas may be masked by illumination noise or low contrast (such as shallow cracks and reflective areas). By aggregating features within the local receptive field (such as edge continuity and texture consistency), defect-specific patterns (such as the linear response of cracks or the annular gradient of bubbles) can be amplified in the feature space. In other words, the pixel-level vector of the enhanced visual feature of the glue groove obtained after enhancement highlights the semantic difference between the defect and the surrounding area (such as the directional enhancement of the crack centerline or the response focus of the closed contour of the bubble), so that the enhanced coding map of the visual feature of the glue groove can retain fine-grained information while improving the intra-class consistency (such as complete cracks) and inter-class discrimination (such as bubbles and artifacts) of the defect area, and ultimately improve the robustness of the downstream classification model to the discrimination of fuzzy defects.

[0071] In an embodiment of the present application, the defect judgment module 150 is configured to determine whether a defect exists within the inverter housing's rubber groove based on the rubber groove visual feature enhancement coding map. Specifically, in an embodiment of the present application, the defect judgment module is configured to input the rubber groove visual feature enhancement coding map into a support vector machine-based defect detection model to obtain a defect detection result, wherein the defect detection result is used to indicate whether a defect exists within the inverter housing's rubber groove. It should be understood that the rubber groove visual feature enhancement coding map contains rich, high-dimensional feature information after multiple rounds of processing. Support vector machines (SVMs) have excellent performance in solving binary classification problems. For inverter housing rubber grooves, there are only two states: "defective" and "non-defective," which perfectly fits the binary classification application scenario of SVMs. By finding an optimal classification hyperplane, it can accurately distinguish between normal features and defect features in the rubber groove visual feature enhancement coding map. Even when faced with complex nonlinear feature distributions, SVMs, through kernel function techniques, can transform nonlinear problems in low-dimensional space into linearly separable problems in high-dimensional space, thereby achieving high-precision classification. The model analyzes and processes input features to produce clear defect detection results, providing a key basis for quality control during inverter production. Whether it's cracks, bubbles, impurities, or insufficient glue, the model can identify various defects based on learned characteristic patterns, helping manufacturers promptly identify substandard products, preventing defective inverters from entering the market and ensuring product reliability and sealing.

[0072] In summary, the defect detection system 100 for the inverter housing glue groove based on computer vision recognition according to the embodiment of the present application is explained. It first pre-processes the collected inverter housing glue groove image to improve the contrast and visibility of the defect area in the image, then locates and segments the pre-processed image to accurately locate the glue groove ROI, and then performs feature encoding on the glue groove surface to automatically capture subtle texture anomalies to enhance the significance of the defect area, so that it forms a clear boundary with the normal area in the feature space, and finally makes a decision on the enhanced high-dimensional features to determine whether there is a defect. In this way, the limitations of traditional methods that rely on manually designed features can be overcome, so that tiny glue groove defects can still be stably identified under complex lighting conditions, thereby reducing false detection and missed detection rates, and improving detection efficiency and automation level.

Claims

1. A defect detection system for inverter housing rubber groove based on computer vision recognition, characterized in that: include: The inverter housing adhesive groove image acquisition module is used to obtain the inverter housing adhesive groove image captured by the camera; A pre-processing module, configured to suppress reflections and compensate shadows on the inverter housing adhesive groove image to obtain a pre-processed adhesive groove image; A ROI extraction module is used to locate and segment the glue tank area of the pre-processed glue tank image to obtain a glue tank ROI image; A feature enhancement module is configured to perform image feature enhancement processing on the glue slot ROI image to obtain a glue slot visual feature enhancement coding map, wherein the feature enhancement module includes: a glue slot visual feature extraction unit, configured to perform convolutional coding-based glue slot visual feature extraction on the glue slot ROI image to obtain a glue slot visual feature coding map; a glue slot visual feature receptive field enhancement unit, configured to perform feature dynamic receptive field significance enhancement on the glue slot visual feature coding map to obtain the glue slot visual feature enhancement coding map; The defect judgment module is used to determine whether there is a defect in the glue groove of the inverter housing based on the glue groove visual feature enhancement coding map.

2. The inverter housing glue groove defect detection system based on computer vision recognition according to claim 1 is characterized in that: The preprocessing module is used to perform histogram equalization, morphological processing and homomorphic filtering on the inverter housing glue groove image to obtain the preprocessed glue groove image.

3. The inverter housing glue groove defect detection system based on computer vision recognition according to claim 2 is characterized in that: The ROI extraction module is used to perform mask R-CNN-based glue tank region positioning and segmentation on the pre-processed glue tank image to obtain the glue tank ROI image.

4. The inverter housing glue groove defect detection system based on computer vision recognition according to claim 1 is characterized in that: The glue slot visual feature extraction unit is used to perform glue slot visual feature extraction based on DenseNet on the glue slot ROI image to obtain the glue slot visual feature coding map.

5. The inverter housing glue groove defect detection system based on computer vision recognition according to claim 1 is characterized in that: The glue groove visual feature receptive field enhancement unit includes: A secondary sub-unit for extracting pixel-level glue groove visual features is used to extract a set of initial pixel-level glue groove visual feature vectors and pixel-level glue groove visual feature vectors to be enhanced from the glue groove visual feature coding map; A receptive field size determination secondary subunit is used to perform feature distillation compression on the pixel-level vector of the visual feature of the glue groove to be enhanced to obtain the size of the receptive field of the visual feature of the glue groove; A secondary sub-unit for selecting pixel-level initial vectors within the receptive field, configured to filter out a set of pixel-level initial vectors within the local receptive field of the glue groove visual feature from the set of pixel-level initial vectors of the glue groove visual feature based on the size of the receptive field of the glue groove visual feature; The secondary sub-unit for significant enhancement of the glue groove visual feature is used to significantly enhance the pixel-level vector of the glue groove visual feature to be enhanced based on the set of pixel-level initial vectors within the local receptive field of the glue groove visual feature to obtain an enhanced glue groove visual feature pixel-level vector, wherein the enhanced glue groove visual feature pixel-level vector is the channel feature vector of the (i, j)th pixel position of the glue groove visual feature enhancement coding map.

6. The inverter housing glue groove defect detection system based on computer vision recognition according to claim 5 is characterized in that: The glue tank visual feature pixel-level extraction secondary sub-unit is used to: Performing feature decoupling on the glue groove visual feature encoding map along the channel dimension to obtain a set of pixel-level initial vectors of the glue groove visual feature; The pixel-level initial feature vector at the (i, j)th pixel position is extracted from the set of the glue groove visual feature pixel-level initial vectors as the glue groove visual feature pixel-level vector to be enhanced.

7. The inverter housing glue groove defect detection system based on computer vision recognition according to claim 6 is characterized in that: The receptive field size determines the secondary subunit, which is used to: Performing information compression on the pixel-level vector of the visual feature of the to-be-enhanced glue tank to obtain a distilled vector of the visual feature of the to-be-enhanced glue tank; Based on the feature distribution spatial structure characteristics of the distilled vector of the visual feature of the glue tank to be strengthened, the size of the glue tank visual feature receptive field of the pixel-level vector of the visual feature of the glue tank to be strengthened is determined.

8. The inverter housing glue groove defect detection system based on computer vision recognition according to claim 7 is characterized in that: The visual characteristics of the glue tank significantly enhance the secondary sub-unit, which is used to: Performing volume-boundary conformal regularization on a set of pixel-level initial vectors within a local receptive field of the glue groove visual feature to determine a first modulation weighting coefficient and a second modulation weighting coefficient; Based on the first modulation weighting coefficient and the second modulation weighting coefficient, the pixel-level vector of the glue groove visual feature to be enhanced and the set of pixel-level initial vectors within the local receptive field of the glue groove visual feature are weighted fused based on the attention weight to obtain the enhanced glue groove visual feature pixel-level vector.

9. The inverter housing glue groove defect detection system based on computer vision recognition according to claim 8 is characterized in that: The defect judgment module is used to: input the glue groove visual feature enhancement coding map into the defect detection model based on the support vector machine to obtain a defect detection result, and the defect detection result is used to indicate whether there is a defect in the glue groove of the inverter housing.

Images