Insulating material surface defect identification method and system based on visual detection

Abstract

The application relates to the technical field of defect identification, and particularly discloses an insulating material surface defect identification method and system based on visual detection, which performs reflection suppression and texture primitive separation based on double-image decomposition on an original surface image of insulating material to generate an insulating material defect explicit image with more prominent defect features, performs multi-scale feature extraction on the insulating material defect explicit image to capture different scale defect information of the insulating material, generates an insulating material defect multi-scale fusion feature map, and then, on the basis, uses a lightweight detection head to perform defect detection on the insulating material defect multi-scale fusion feature map to output a defect candidate region and a defect type label and confidence thereof. The method effectively solves the feature confusion problem caused by reflection and texture interference on the surface of the insulating material by performing reflection suppression and texture primitive separation on the original surface image of the insulating material, and improves the accuracy of defect identification.

Description

Technical Field

[0001] This application relates to the field of defect recognition technology, and more specifically, to a method and system for identifying surface defects in insulating materials based on visual inspection. Background Technology

[0002] Insulating materials are a critical component in power systems, electronic equipment, and numerous industrial sectors; their quality directly impacts the safe, stable operation and lifespan of equipment. Surface defects in insulating materials, such as scratches, bubbles, cracks, and impurities, not only affect the product's appearance and performance but can also lead to insulation failure, causing equipment damage, system shutdowns, and even serious safety accidents. Therefore, timely and accurate identification and assessment of surface defects in insulating materials are crucial for ensuring safe equipment operation, improving product quality, and reducing production costs. However, traditional manual visual inspection methods suffer from low efficiency, strong subjectivity, susceptibility to fatigue, and high costs, making it difficult to meet the demands of large-scale, high-precision inspections. Automated and intelligent defect identification technologies have become an urgent need for industry development.

[0003] Against this backdrop, automated defect detection methods based on machine vision have gradually attracted attention. However, in the field of surface defect detection of insulating materials, traditional image processing algorithms, such as threshold segmentation, edge detection, and morphological processing, are affected by factors such as dynamic changes in lighting conditions, interference from material surface textures, and the diversity of defect morphologies in practical applications, resulting in insufficient recognition accuracy and robustness. Furthermore, directly applying general-purpose deep learning vision models to process images of insulating material surfaces presents unique imaging challenges (such as strong reflectivity and weak texture defects). Strong noise generated by material reflections easily confuses with the substrate texture, leading to missed or false detections, which in turn affects the reliability and accuracy of the detection results.

[0004] Therefore, an optimized method and system for identifying surface defects in insulating materials based on visual inspection is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for identifying surface defects in insulating materials based on visual inspection. This method generates a more prominent image of the insulating material defects by performing reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material. Multi-scale feature extraction is then performed on this prominent image to capture defect information at different scales, generating a multi-scale fusion feature map of the insulating material defects. Furthermore, a lightweight detection head is used to detect defects on the multi-scale fusion feature map, outputting candidate defect regions, their defect type labels, and confidence scores. This method effectively solves the feature confusion problem caused by reflection and texture interference on the surface of insulating materials by performing reflection suppression and texture primitive separation on the original surface image of the insulating material, thus improving the accuracy of defect identification.

[0006] Accordingly, according to one aspect of this application, a method for identifying surface defects in insulating materials based on visual detection is provided, comprising: acquiring an original surface image of the insulating material; performing reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material to obtain a visible image of the insulating material defects; inputting the visible image of the insulating material defects into a multi-scale feature extraction model comprising a backbone network and a feature pyramid network to obtain a multi-scale fusion feature map of the insulating material defects; and inputting the multi-scale fusion feature map of the insulating material defects into a lightweight detection head to obtain candidate regions of defects and their defect type labels and confidence scores.

[0007] According to another aspect of this application, a visual inspection-based surface defect recognition system for insulating materials is provided, comprising: a surface image acquisition module for acquiring an original surface image of the insulating material; a defect explicitation module for performing reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material to obtain a defect explicitation image of the insulating material; a multi-scale feature extraction module for inputting the defect explicitation image of the insulating material into a multi-scale feature extraction model including a backbone network and a feature pyramid network to obtain a multi-scale fusion feature map of the insulating material defects; and a defect detection module for inputting the multi-scale fusion feature map of the insulating material defects into a lightweight detection head to obtain defect candidate regions and their defect type labels and confidence scores.

[0008] Compared with existing technologies, the visual inspection-based method and system for identifying surface defects in insulating materials provided in this application generates a more prominent image of the insulation material defects by performing reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material. Multi-scale feature extraction is then performed on this prominent image to capture defect information at different scales, generating a multi-scale fusion feature map of the insulation material defects. Based on this, a lightweight detection head is used to detect defects on the multi-scale fusion feature map, outputting candidate defect regions, their defect type labels, and confidence scores. This method effectively solves the feature confusion problem caused by reflection and texture interference on the surface of the insulating material by performing reflection suppression and texture primitive separation on the original surface image of the insulating material, thus improving the accuracy of defect identification. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of the data flow of a visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application.

[0012] Figure 3 This is a flowchart of step S2 in the visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application.

[0013] Figure 4 This is a flowchart of step S21 in the visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application.

[0014] Figure 5 This is a flowchart of step S22 in the visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application.

[0015] Figure 6 This is a flowchart of step S3 in the visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application.

[0016] Figure 7 This is a block diagram of a visual inspection-based surface defect identification system for insulating materials according to an embodiment of this application. Detailed Implementation

[0017] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein. It is worth noting that in this application, all data acquisition actions are performed in accordance with the relevant data protection laws and policies of the country where the application is located, and with authorization from the owner of the relevant device.

[0018] Figure 1 This is a flowchart of a visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow of a vision-based method for identifying surface defects in insulating materials according to an embodiment of this application. Figure 1 and Figure 2 As shown, the visual detection-based surface defect identification method for insulating materials according to an embodiment of this application includes the following steps: S1, acquiring an original surface image of the insulating material; S2, performing reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material to obtain a visible image of the insulating material defects; S3, inputting the visible image of the insulating material defects into a multi-scale feature extraction model containing a backbone network and a feature pyramid network to obtain a multi-scale fusion feature map of the insulating material defects; S4, inputting the multi-scale fusion feature map of the insulating material defects into a lightweight detection head to obtain candidate defect regions and their defect type labels and confidence levels.

[0019] In the aforementioned visual inspection-based method for identifying surface defects in insulating materials, step S1 involves acquiring an original surface image of the insulating material. Specifically, this application, based on the principle of machine vision inspection, uses imaging devices such as industrial cameras and line scan cameras to perform optical imaging on the surface of the insulating material during the production line or quality inspection process to obtain an original surface image. In one specific embodiment of this application, a high-resolution CCD or CMOS industrial camera, such as a 5-megapixel or higher area scan camera, is installed on the production line or at a specific inspection station. The lens is configured to adapt to the inspection distance and field of view, ensuring that each pixel can distinguish the size of the smallest target defect. Regarding the light source, a ring LED light source, a strip LED light source, or a coaxial light source can be used. The incident angle and intensity of the light source are adjusted according to the reflective characteristics of the insulating material surface to minimize the highlight areas caused by direct reflection while ensuring sufficient brightness and enhancing the contrast between the defect and the background. For example, for highly reflective surfaces, a diffuse dome light source or the angle of a strip light source is adjusted to create a dark field lighting effect to highlight surface pits or scratches. The camera acquires images under the control of a trigger signal (such as a product arrival sensor signal) and performs grayscale processing to reduce the impact of lighting changes on subsequent processing, thereby obtaining an original surface image of the insulation material that reflects its current surface state and contains information about potential defects. After acquisition, the image data is transmitted to an image processing unit (such as an industrial computer or embedded system) for storage, awaiting further processing. In this way, reliable and high-quality input data is provided for subsequent insulation material defect identification.

[0020] In the aforementioned visual detection-based method for identifying surface defects in insulating materials, step S2 involves performing reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material to obtain a visible image of the insulating material defects. It should be understood that, due to the strong reflective properties of insulating material surfaces (such as smooth materials like epoxy resin and polyester film), uneven illumination can lead to significant brightness differences in different areas of the image. Reflected light interference can mask the true grayscale features of the defects. Simultaneously, the complex or repetitive textures that may exist within the material itself can easily be confused with defect features (especially defects with similar texture orientations), severely affecting the accurate extraction and identification of subsequent defect features. Directly applying traditional algorithms or deep learning models can easily lead to missed or false detections. Therefore, in order to effectively suppress adverse factors and enhance the visual salience of defect features in the image, this application performs dual image decomposition on the original surface image of the insulating material to gradually separate and weaken light reflection and texture interference, thereby highlighting the true defect information.

[0021] Figure 3 This is a flowchart of step S2 in the visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application. Figure 3 As shown, step S2 includes: S21, performing illuminance-reflectance decomposition and reflection suppression based on multi-scale Retinex on the original surface image of the insulating material to obtain a surface reflectance image of the insulating material; S22, performing texture and defect separation based on low-rank and sparse decomposition on the surface reflectance image of the insulating material to obtain a surface texture image and an initial defect layer image of the insulating material; S23, performing noise suppression on the initial defect layer image of the insulating material to obtain a surface defect layer image of the insulating material; S24, fusing the surface reflectance image, the surface texture image, and the surface defect layer image of the insulating material to obtain a defect-revealed image of the insulating material.

[0022] Specifically, in step S21, the original surface image of the insulating material is subjected to illuminance-reflectance decomposition and reflection suppression based on multi-scale Retinex to obtain a reflectance image of the insulating material surface. It should be understood that the pixel values ​​in the original surface image of the insulating material are the result of the combined effect of the object's surface reflectivity and incident light, with the illuminance component being the main source of strong reflection and highlights. Therefore, to eliminate the influence of uneven lighting and extract the inherent reflectivity of the object, this application is based on Retinex theory, i.e., an image can be represented as the product of an illuminance component and a reflectance component, where the reflectance component contains the essential features of the target. Thus, the original surface image of the insulating material is decomposed using a multi-scale Retinex algorithm to obtain an illuminance-reflectance image that better reflects the essential characteristics of the insulating material by removing the illuminance component. Figure 4 As shown, step S21 includes: S211, using the multi-scale Retinex algorithm to estimate the illuminance of the original surface image of the insulating material to obtain an estimated illuminance component image of the insulating material surface; S212, calculating the surface reflectance image of the insulating material based on the estimated illuminance component image of the insulating material surface and the original surface image of the insulating material.

[0023] In one specific embodiment of this application, step S211 involves using a multi-scale Retinex algorithm to estimate the illuminance of the original surface image of the insulating material to obtain an estimated image of the illuminance components on the surface of the insulating material. Specifically, firstly, the original surface image of the insulating material is convolved using Gaussian filters of small, medium, and large scales, with standard deviations of the corresponding Gaussian blur kernels being sigma_s=15, sigma_m=80, and sigma_l=200, respectively, to estimate the illuminance of different frequency components. For example, for the original surface image of the insulating material... Illuminance estimation using single-scale Retinex ,in It is the first A Gaussian kernel at each scale is used. By summing and averaging the illuminance estimates at each single scale, the final estimated image of the illuminance components on the surface of the insulating material can be obtained. Here, a small-scale Gaussian kernel is used to capture local illumination changes in the image (such as small reflective areas), a medium-scale Gaussian kernel is used to capture medium-range illumination changes in the image (such as medium-sized reflective or shadow areas), and a large-scale Gaussian kernel is used to fit global illumination trends (such as overall darkness or brightness). By fusing multi-scale illumination estimates, a balance can be effectively struck between preserving local details and achieving global illumination balance, thus improving the accuracy and robustness of illumination estimation.

[0024] In one specific embodiment of this application, step S212 involves calculating the surface reflectance image of the insulating material based on the estimated illuminance component image of the insulating material surface and the original surface image of the insulating material. Specifically, firstly, based on the estimated illuminance component image of the insulating material surface and the original surface image of the insulating material, the logarithmic domain image of the surface reflectance of the insulating material is calculated using the following formula: ;in, In the original surface image of the insulating material The pixel value of the location, Estimating the illuminance components on the surface of insulating materials in the image The pixel value of the location, In the surface reflectance image of insulating material The pixel value of the location, Represents the logarithmic function with base 2. A constant hyperparameter used to prevent logarithmic zero. In the logarithmic domain image representing the surface reflectance of insulating materials The element value at position.

[0025] It should be understood that, based on Retinex theory, the original surface image of the insulating material can be represented as the product of the illuminance component and the reflectance component. Here, this application uses a logarithmic transformation to convert the multiplicative model in the image domain into an additive model in the logarithmic domain, thereby facilitating the separation of the illuminance and reflectance components of the original surface image of the insulating material.

[0026] Next, the logarithmic domain image of the reflectance of the insulating material surface is transformed back to the original domain to obtain the reflectance image of the insulating material surface. That is, the logarithmic domain image of the reflectance of the insulating material surface is transformed from the logarithmic domain back to the original image domain through exponential operations to obtain the reflectance image of the insulating material surface.

[0027] Specifically, step S22 involves performing low-rank and sparse decomposition-based texture and defect separation on the reflectivity image of the insulating material surface to obtain a texture image and an initial defect layer image of the insulating material surface. It should be understood that even after eliminating reflection interference in the original image, the inherent, regular texture background of the insulating material surface itself may still visually confuse with irregular, abrupt defect features, especially some small defects or those similar to the texture structure. Therefore, to effectively separate the texture information representing the background from the defect information representing anomalies, this application is based on the principle of low-rank and sparse decomposition of images. That is, assuming that the background texture part of the image has low-rank characteristics, while the defect part is relatively sparsely distributed in the image, the reflectivity image of the insulating material surface is decomposed using the Robust Principal Component Analysis (RPCA) algorithm, representing the background texture as a low-rank matrix and the defects as a sparse matrix. Thus, by solving this low-rank and sparse decomposition problem, the texture image and the initial defect layer image of the insulating material surface can be obtained.

[0028] Figure 5 This is a flowchart of step S22 in the visual inspection-based method for identifying surface defects in insulating materials according to an embodiment of this application. Figure 5 As shown, step S22 includes: S221, dividing the surface reflectance image of the insulating material into blocks to obtain a set of surface reflectance image blocks of the insulating material; S222, vectorizing each surface reflectance image block in the set of surface reflectance image blocks of the insulating material to obtain a surface reflectance image matrix of the insulating material; S223, performing robust principal component analysis on the surface reflectance image matrix of the insulating material to obtain a low-rank component matrix and a sparse component matrix; S224, performing image reconstruction based on the low-rank component matrix and the sparse component matrix to obtain a surface texture image of the insulating material and an initial defect layer image of the insulating material surface.

[0029] More specifically, step S221 involves dividing the reflectance image of the insulating material surface into image blocks to obtain a set of reflectance image blocks of the insulating material surface. It should be understood that, since the surface texture of insulating materials often exhibits local repetition (such as fiber weave structures or pressed mesh patterns), and defects typically exist in a locally sparse form (such as cracks or bubbles in a certain area), directly performing global low-rank sparse decomposition on the entire image is insufficient to effectively capture the spatial distribution differences between local texture patterns and defects. Therefore, in order to transform the global image decomposition problem into low-rank sparse modeling of local blocks, this application divides the entire reflectance image of the insulating material surface into a set of sub-blocks with overlapping or non-overlapping characteristics through image block division, so as to utilize the low-rank nature of the texture and the sparsity of the defects for feature separation within a local range.

[0030] In one specific embodiment of this application, the block size is first determined based on the typical size of the surface texture of the insulating material. For example, for materials with a texture period of 10-20 pixels (such as fiberglass cloth), 32×32 pixel square blocks are used for division, ensuring that each sub-block contains at least one complete texture period. Simultaneously, to avoid information breakage at block boundaries, a 50% overlap rate is used for block division (i.e., adjacent sub-blocks overlap by 16 pixels in the horizontal and vertical directions), and the overlapping areas maintain continuity by copying edge pixels. During the block division process, areas where the image edge is less than the size of a block are zero-filled or have their edge pixels repeated to ensure that all insulating material surface reflectance image blocks are of consistent size. For example, for a reflectance image with a resolution of 512×512, after dividing into 32×32 pixel blocks with a 50% overlap rate, (512-32) / 16+1=31 rows × 31 columns, a total of 961 sub-blocks are obtained, each containing 1024 pixels.

[0031] More specifically, step S222 involves vectorizing each insulating material surface reflectance image patch in the set of insulating material surface reflectance image patches to obtain an insulating material surface reflectance image matrix. It should be understood that matrix factorization algorithms such as Robust Principal Component Analysis (RPCA) require converting two-dimensional image data into one-dimensional vector form to construct a matrix. The vectorization operation of image patches can convert local spatial structure information into column vectors of a matrix, facilitating the separation of low-rank and sparse components using linear algebra tools. Therefore, to adapt each insulating material surface reflectance image patch to the matrix factorization model, this application, based on vector space theory, expands each insulating material surface reflectance image patch into a one-dimensional column vector in row-major or column-major order, thereby arranging the vectors of all insulating material surface reflectance image patches sequentially to form a reflectance image matrix.

[0032] In one specific embodiment of this application, for each 32×32 pixel image block, a row-first vectorization method is adopted, that is, starting from the first pixel of the first row, the pixels of each row are sequentially connected to form a column vector of length 1024. Then, the vectorization results of all sub-blocks are arranged in block order (e.g., from left to right, from top to bottom) to construct a 1024×961 image matrix representing the reflectivity of the insulating material surface. Each column in the image matrix corresponds to a vector representation of an insulating material surface reflectivity image block. Due to the existence of overlapping blocks, the pixel values ​​of adjacent column vectors in the overlapping area are highly correlated, thus providing conditions for modeling the local repetition of textures using low-rank properties. For example, for an image block with a periodic grid texture, the difference between adjacent column vectors mainly lies in the translation of the grid position, and the whole can be represented by a low-dimensional subspace spanned by a few basis vectors.

[0033] More specifically, step S223 involves performing robust principal component analysis (RPCA) on the reflectivity image matrix of the insulating material surface to obtain a low-rank component matrix and a sparse component matrix. It should be understood that since the surface texture of the insulating material exhibits strong correlation (low-rank) within local blocks, while defects and noise manifest as local sparse outliers, RPC can simultaneously utilize both low-rank and sparsity properties to decompose the matrix into low-rank texture components and sparse defect components, thereby achieving the separation of texture and defects. Therefore, in order to separate regular textures and sparse defects from the reflectivity image matrix of the insulating material surface, this application, based on RPCA theory, achieves low-rank sparse decomposition of the matrix by solving a convex optimization problem.

[0034] In one specific embodiment of this application, the implementation method is as follows: utilizing the RPCA model. Image matrix of surface reflectance of the insulating material Decompose, where This is a low-rank component matrix, corresponding to the texture components; Let be a sparse component matrix, corresponding to the defect and noise components. The optimization objective function is: ;in, For matrix The nuclear norm (i.e., the sum of singular values) is used to constrain... The low-rank nature, For matrix The L1 norm (sum of absolute values ​​of elements) is used to constrain... The sparsity of the column (i.e., each column has at most a few non-zero elements). To balance the regularization parameters between low-rank and sparse terms, a value of 0.1-0.5 is typically chosen. For this objective function, the Alternating Direction Multiplier Method (ADMM) is used for iterative solution, continuously updating... and The value of is calculated until convergence, to obtain the final low-rank component matrix. and sparse component matrix .

[0035] More specifically, step S224 involves image reconstruction based on the low-rank component matrix and the sparse component matrix to obtain the surface texture image and the initial defect layer image of the insulating material. Specifically, to restore the low-rank component matrix and the sparse component matrix to a two-dimensional image for intuitive differentiation of texture and defect distribution, this application, based on the inverse process of image vectorization, rearranges the column vectors of the low-rank component matrix and the sparse component matrix into image blocks, and reconstructs the complete surface texture image and the initial defect layer image of the insulating material by stitching together these image blocks.

[0036] In one specific embodiment of this application, for the low-rank component matrix, each column vector is rearranged into two-dimensional image blocks according to the block order and block size (e.g., 32×32 pixels), and an overlapping block stitching technique is used to eliminate the block boundary effect. Specifically, in the case of overlapping blocks, each pixel is estimated multiple times in multiple overlapping blocks, and the average value is calculated based on the number of overlaps to obtain the final pixel value, thus obtaining the surface texture image of the insulating material. Similarly, the same vectorization inverse operation is performed on the sparse component matrix to obtain the initial defect layer image of the insulating material surface.

[0037] Specifically, in step S23, noise suppression is applied to the initial defect layer image of the insulating material surface to obtain the final defect layer image. It should be understood that the initial defect layer image of the insulating material surface may contain random interference such as salt-and-pepper noise and Gaussian noise due to sensor noise and environmental interference during image acquisition. This can lead to blurred defect edges and the generation of false defect points, affecting subsequent detection accuracy. Therefore, to purify the defect layer image and preserve true defect features, this application employs a filtering algorithm to suppress noise in the initial defect layer image of the insulating material surface based on the spatial distribution difference between noise and defects (noise is typically a random isolated point, while defects have a continuous structure).

[0038] In one specific embodiment of this application, firstly, the noise type of the initial defect layer image on the surface of the insulating material is analyzed. If salt-and-pepper noise (manifested as isolated black and white noise points) is dominant, a median filtering algorithm is used: for each pixel, the pixel values ​​in its 3×3 or 5×5 neighborhood are sorted, and the median value is used to replace the current pixel value. This method can effectively suppress salt-and-pepper noise while preserving the defect edges. If the noise is mainly Gaussian distributed (manifested as overall grayscale fluctuations), nonlocal mean filtering (NLM) is used: multiple blocks similar to the current pixel's neighborhood are searched in the image, and noise is suppressed by weighted averaging, with the weight being the similarity between each image block and the current pixel's neighborhood image block. This method can better preserve the fine structure of the defects (such as hairline cracks) while suppressing Gaussian noise. In practical applications, the initial defect layer image on the surface of the insulating material can first be used to determine the noise type. For example, by calculating the ratio of local variance to global variance, if the ratio is greater than a certain threshold (such as 2), it is determined that Gaussian noise is dominant; otherwise, it is determined that salt-and-pepper noise is dominant. Then, based on the noise type identification result, an adaptive filtering algorithm is selected.

[0039] Specifically, in step S24, the reflectance image of the insulating material surface, the texture image of the insulating material surface, and the defect layer image of the insulating material surface are fused to obtain a visible image of the insulating material defects. It should be understood that a single image component is insufficient to comprehensively characterize the defect features of the insulating material surface. For example, while the reflectance image retains the overall grayscale information, the defect contrast is insufficient; the texture image reflects the insulating material structure but does not contain defect information; and the defect layer image highlights defect information but lacks spatial context. Therefore, this application further enhances defect contrast and improves the explicit expression of defect features by fusing the reflectance image, texture image, and defect layer image of the insulating material surface, thereby preserving the overall structural features of the insulating material surface.

[0040] In one specific embodiment of this application, firstly, the surface reflectance image, surface texture image, and surface defect layer image of the insulating material are subjected to size normalization and alignment processing (if boundary effects exist in the block processing, the original size needs to be restored by mirror filling or overlapping stitching) to ensure that the pixel coordinates of the three images correspond one-to-one. Next, a weighted fusion strategy is adopted: the surface reflectance image of the insulating material provides basic grayscale information, with a weight of 0.5; the surface texture image of the insulating material is used to characterize the background structure, with a weight of 0.3; and the surface defect layer image of the insulating material directly reflects the defect location, with a weight of 0.2. By weighted averaging of the pixel values ​​at corresponding positions in the three images, a visible image of the insulating material defects is obtained.

[0041] In the aforementioned visual detection-based method for identifying surface defects in insulating materials, step S3 involves inputting the explicit image of the insulating material defects into a multi-scale feature extraction model comprising a backbone network and a feature pyramid network to obtain a multi-scale fused feature map of the insulating material defects. It should be understood that, considering the significant differences in the size of surface defects in insulating materials, such as micrometer-level scratches and centimeter-level bubbles, single-scale convolution operations are insufficient to simultaneously capture features of defects of different sizes. Therefore, in order to extract basic features covering multiple scales from the explicit image of the insulating material defects, this application, based on the hierarchical feature extraction principle of convolutional neural networks and feature pyramid networks (FPN), extracts multi-level features through a backbone network, and then performs cross-level information interaction through the FPN network to generate a multi-scale fused feature map of the insulating material defects. Figure 6As shown, step S3 includes: S31, inputting the explicit image of the insulation material defect into the backbone network to obtain multiple feature maps of different scales; S32, inputting the multiple feature maps of different scales into the feature pyramid network to obtain a set of multi-scale feature maps; S33, fusing the set of multi-scale feature maps to obtain a multi-scale fused feature map of the insulation material defect.

[0042] In one specific embodiment of this application, the backbone network adopts a lightweight ResNet-18 architecture, which includes an input layer, four residual blocks and an output layer. Each residual block contains two 3×3 convolutional layers and a skip connection. The resolution of the extracted feature map is halved step by step through stride convolution, which are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the image of the insulation material defect manifestation, respectively, thereby obtaining multiple feature maps of different scales.

[0043] Furthermore, due to the scale gap in the feature maps output by each stage of the backbone network (e.g., a two-fold difference between 1 / 4 and 1 / 8 resolution), and the low spatial resolution of high-level feature maps, the location information of small defects may be lost, while the semantic information of low-level feature maps is insufficient, making it difficult to accurately classify defect types. Therefore, in order to bridge the semantic gap between feature maps of different scales and enhance the ability to express cross-scale defects, this application combines high-level semantic features with low-level detailed features based on the hierarchical fusion principle of feature pyramid networks through top-down paths and lateral connections. Specifically, the feature pyramid network starts from the deepest layer of the backbone network (1 / 32 resolution feature map). It enlarges the feature map size by a factor of two through nearest-neighbor upsampling (or bilinear interpolation), while simultaneously compressing the number of channels using 1×1 convolutions to generate high-level semantic feature maps. Next, the upsampled high-level semantic feature map is element-wise added to the backbone network's feature map at the same resolution (1 / 16 resolution feature map), and a 3×3 convolution is used to eliminate the aliasing effect caused by upsampling, resulting in a fused mid-to-high-level feature map. This fused mid-to-high-level feature map is then fused with the feature map of the next level, and so on, until it is fused with the shallowest layer of the backbone network (1 / 4 resolution feature map), thus obtaining a set of multi-scale feature maps covering low-level details to high-level semantics. Through this cross-layer fusion mechanism, each multi-scale feature map not only retains the defect features of its corresponding scale but also incorporates deeper-level cross-scale contextual information. This allows each multi-scale feature map to simultaneously utilize shallow positional information and deep category information, improving the model's sensitivity and recognition accuracy for defects of different sizes.

[0044] Finally, to comprehensively utilize the complementary information of the multi-scale feature maps and improve the robustness of insulation material defect identification, this application further performs a fusion process on the set of multi-scale feature maps. In a specific implementation, a positional mean pooling strategy is adopted to calculate the mean of pixel values ​​at corresponding positions in each multi-scale feature map to obtain the multi-scale fused feature map of the insulation material defect. In this way, feature information from different scales can be effectively fused, so that the final fused feature map can more comprehensively and accurately reflect the defect features on the surface of the insulation material.

[0045] In the aforementioned visual inspection-based method for identifying surface defects in insulating materials, step S4 involves inputting the multi-scale fused feature map of the insulating material defects into a lightweight detection head to obtain candidate defect regions, their defect type labels, and confidence scores. It should be understood that traditional deep learning detection heads, such as the RPN of Faster R-CNN, have high computational complexity, making it difficult to meet the demands of real-time industrial inspection. In contrast, the scenario of insulating material defect detection requires both accuracy and inference speed. Therefore, to achieve rapid localization and classification of candidate defect regions, this application employs a lightweight detection head, improving defect detection efficiency through simplified network structure and optimized anchor frame design.

[0046] In one specific embodiment of this application, the lightweight detection head is an SDD (Single Shot Detector) head. Specifically, the SDD detection head presets anchor boxes of different sizes (e.g., small anchor boxes 32×32 pixels, large anchor boxes 128×128 pixels) on the multi-scale fusion feature map of the insulation material defects. Each anchor box corresponds to a preset defect category (e.g., scratches, bubbles, dents, etc.). The detection head contains two parallel branches: one branch predicts the position offset (Δx, Δy, Δw, Δh) and size scaling factor of the anchor box through 3×3 convolution to achieve candidate region regression; the other branch classifies each anchor box through 1×1 convolution, outputting the probability distribution and confidence score of each defect type. During the training process of the SDD detection head, Focal Loss is used to alleviate the problem of positive and negative sample imbalance. In the defect detection stage, Non-Maximum Suppression (NMS) is used to remove overlapping candidate boxes, retaining detection results with a confidence score ≥0.5, and finally outputting the defect location coordinates, category label, and confidence score. In this way, based on the multi-scale fusion feature map of the insulation material defects, a lightweight SSD detection head can be used to quickly and accurately locate and identify various defects on the surface of the insulation material, and provide quantitative evaluation results, thus achieving efficient and robust defect detection.

[0047] In particular, in a preferred embodiment of this application, step S4 includes: first, performing multi-scale feature coupling noise suppression on the multi-scale fusion feature map of the insulation material defect to obtain an optimized multi-scale fusion feature map of the insulation material defect; and then inputting the optimized multi-scale fusion feature map of the insulation material defect into a lightweight detection head to obtain the defect candidate region and its defect type label and confidence level. It should be understood that for the explicit image of insulation material defects used for multi-scale image semantic local correlation feature extraction, it integrates the surface reflectance image of the insulation material obtained by performing illuminance-reflectance decomposition and reflection suppression based on multi-scale Retinex, the surface texture image of the insulation material obtained by separating texture and defects based on low-rank and sparse decomposition, and the initial defect layer image of the insulation material surface. Due to the logarithmic domain transformation-reduction of illuminance-reflectance decomposition and reflection suppression based on multi-scale Retinex and the principal component extraction-reconstruction based on texture and defects based on low-rank and sparse decomposition, the multi-scale fused feature map of insulation material defects will also have a nonlinear product residual in the logarithmic domain space of illuminance-reflectance based on multi-scale Retinex. That is, after the nonlinear product of image semantics to illumination field and material reflection field is constructed by multi-scale superposition, it is coupled in low dimension with the high-frequency scale interference retained by the principal component extraction-reconstruction based on texture and defects based on low-rank and sparse decomposition to generate coupling noise. Based on this, this application aims to suppress multi-scale feature coupling noise on the multi-scale fused feature map of the insulation material before performing defect detection, so as to optimize its feature representation and reduce the interference of coupling noise on the accuracy of defect detection.

[0048] Specifically, the multi-scale fusion feature map of the insulation material defects is, for example, represented as... First, the multi-scale fusion feature map of the insulation material defects is adjusted by multi-scale fusion feature manifold structure matching to obtain the adjusted multi-scale fusion feature map of the insulation material defects, that is, it is represented as a low-dimensional curvature adjustment mode in the logarithmic domain, i.e.: ;in, and These are the low-rank component matrix and the sparse component matrix, respectively. This is a multi-scale fusion feature map of defects in the insulating material. The Each feature matrix This indicates that the feature matrices are arranged by channels. This represents matrix multiplication. This indicates adding based on position points. This represents the multi-scale fusion feature map of defects in the adjusted insulation material. Thus, manifold geometry matching adjustment between the Retinex illuminance-reflectivity logarithmic domain and the principal component low-rank sparse domain is achieved.

[0049] Then, based on the residual response between the adjusted insulation material defect multi-scale fusion feature map and the original insulation material defect multi-scale fusion feature map, the adjusted insulation material defect multi-scale fusion feature map is subjected to high-frequency interference orthogonal scaling to obtain a denoised insulation material defect multi-scale fusion feature map. More specifically, for the adjusted insulation material defect multi-scale fusion feature map... Each feature matrix, for example denoted as The low-rank curvature orthogonal coupling of the residual disturbance is expressed as: ;in, Denotes the F-norm of a matrix. Indicates the difference based on location. This indicates dot product by position. Represents a matrix Perform the reciprocal operation position by position. This represents the multi-scale fusion feature map of insulation material defects after noise reduction. Specifically, it utilizes the commutative property of the F-norm of the residual response matrix to generate a normative transformation from high-frequency interference to a low-dimensional space, thereby solving the low-dimensional coupling problem of multi-scale residual interference.

[0050] Finally, the noise-reduced multi-scale fusion feature map of insulation material defects and the original multi-scale fusion feature map of insulation material defects are orthogonally fused to obtain the optimized multi-scale fusion feature map of insulation material defects, namely: ;in, It is a feature map The Each feature matrix This represents the optimization of the multi-scale fusion feature map of insulation material defects. This corrects the coupling noise between the multi-scale Retinex illuminance-reflectance logarithmic domain and the principal component low-rank sparse domain of the multi-scale fusion feature map of insulation material defects, thereby optimizing the multi-scale fusion feature map of insulation material defects.

[0051] In summary, the visual detection-based surface defect identification method for insulating materials according to the embodiments of this application is explained. It generates a more prominent defect-revealing image of the insulating material by performing reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material. Furthermore, it extracts multi-scale features from this revealed defect image to capture defect information at different scales, generating a multi-scale fusion feature map of the insulating material defects. Then, based on this, a lightweight detection head is used to detect defects on the multi-scale fusion feature map of the insulating material defects, outputting candidate defect regions, their defect type labels, and confidence scores. This method effectively solves the feature confusion problem caused by reflection and texture interference on the surface of the insulating material by performing reflection suppression and texture primitive separation on the original surface image of the insulating material, thus improving the accuracy of defect identification.

[0052] Furthermore, this application also provides a visual inspection-based system for identifying surface defects in insulating materials.

[0053] Figure 7 This is a block diagram of a vision-based surface defect identification system for insulating materials according to an embodiment of this application. Figure 7 As shown, the visual detection-based surface defect recognition system 100 for insulating materials according to an embodiment of this application includes: a surface image acquisition module 110 for acquiring an original surface image of the insulating material; a defect explicitation module 120 for performing reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material to obtain an explicit image of the insulating material defect; a multi-scale feature extraction module 130 for inputting the explicit image of the insulating material defect into a multi-scale feature extraction model containing a backbone network and a feature pyramid network to obtain a multi-scale fusion feature map of the insulating material defect; and a defect detection module 140 for inputting the multi-scale fusion feature map of the insulating material defect into a lightweight detection head to obtain defect candidate regions and their defect type labels and confidence levels.

[0054] Here, those skilled in the art will understand that the specific operation of each module in the above-described vision-based insulation material surface defect identification system has been described in the above-described manner. Figures 1 to 6 The description of the visual inspection-based surface defect identification method for insulating materials is detailed here, and therefore, its repeated description will be omitted.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying surface defects in insulating materials based on visual inspection, characterized in that, include: Obtain the original surface image of the insulating material; The original surface image of the insulating material is subjected to reflection suppression and texture primitive separation based on dual image decomposition to obtain a visible image of the insulating material defects, including: The original surface image of the insulating material is subjected to illuminance-reflectance decomposition and reflection suppression based on multi-scale Retinex to obtain the surface reflectance image of the insulating material, including: The illuminance estimation image of the original surface image of the insulating material is obtained by using the multi-scale Retinex algorithm; Based on the estimated illuminance component image of the insulating material surface and the original surface image of the insulating material, the surface reflectance image of the insulating material is calculated, including: based on the estimated illuminance component image of the insulating material surface and the original surface image of the insulating material, the logarithmic domain image of the surface reflectance of the insulating material is calculated using the following formula: ； in, In the original surface image of the insulating material The pixel value of the location, Estimating the illuminance components on the surface of insulating materials in the image The pixel value of the location, In the surface reflectance image of insulating material The pixel value of the location, Represents the logarithmic function with base 2. A constant hyperparameter used to prevent logarithmic zero. In the logarithmic domain image representing the surface reflectance of insulating materials The element value of the position; convert the logarithmic domain image of the surface reflectance of the insulating material back to the original domain to obtain the surface reflectance image of the insulating material; The surface reflectance image of the insulating material is subjected to texture and defect separation based on low-rank and sparse decomposition to obtain the surface texture image and the initial defect layer image of the insulating material. Noise suppression is applied to the initial defect layer image on the surface of the insulating material to obtain the surface defect layer image of the insulating material; The surface reflectance image of the insulating material, the surface texture image of the insulating material, and the surface defect layer image of the insulating material are fused to obtain a defect-revealed image of the insulating material; The explicit image of the insulation material defects is input into a multi-scale feature extraction model containing a backbone network and a feature pyramid network to obtain a multi-scale fused feature map of the insulation material defects. The multi-scale fusion feature map of the insulation material defects is input into a lightweight detection head to obtain candidate defect regions, their defect type labels, and confidence levels, including: The optimization of the insulation material defect multi-scale fusion feature map by performing multi-scale feature coupling noise suppression on the multi-scale fusion feature map includes: adjusting the multi-scale fusion feature manifold structure matching of the multi-scale fusion feature map to obtain an adjusted multi-scale fusion feature map; based on the residual response between the adjusted multi-scale fusion feature map and the original multi-scale fusion feature map, performing high-frequency interference orthogonal scaling on the adjusted multi-scale fusion feature map to obtain a noise-reduced multi-scale fusion feature map; and orthogonally fusing the noise-reduced multi-scale fusion feature map and the original multi-scale fusion feature map to obtain the optimized multi-scale fusion feature map. The optimized multi-scale fusion feature map of insulation material defects is input into a lightweight detection head to obtain the defect candidate regions, their defect type labels, and confidence levels.

2. The method for identifying surface defects in insulating materials based on visual inspection according to claim 1, characterized in that, The lightweight detection head is an SDD detection head.

3. The method for identifying surface defects in insulating materials based on visual inspection according to claim 1, characterized in that, The surface reflectance image of the insulating material is subjected to texture and defect separation based on low-rank and sparse decomposition to obtain a surface texture image and an initial defect layer image of the insulating material, including: The surface reflectance image of the insulating material is divided into blocks to obtain a set of surface reflectance image blocks of the insulating material; Vectorize each surface reflectance image patch in the set of surface reflectance image patches of the insulating material to obtain the surface reflectance image matrix of the insulating material; Robust principal component analysis was performed on the surface reflectance image matrix of the insulating material to obtain the low-rank component matrix and the sparse component matrix; Image reconstruction is performed based on the low-rank component matrix and the sparse component matrix respectively to obtain the surface texture image of the insulating material and the initial defect layer image of the insulating material surface.

4. The method for identifying surface defects in insulating materials based on visual inspection according to claim 1, characterized in that, The explicit image of the insulation material defects is input into a multi-scale feature extraction model containing a backbone network and a feature pyramid network to obtain a multi-scale fused feature map of the insulation material defects, including: The image of the insulation material defects is input into the backbone network to obtain feature maps of multiple different scales; The feature maps of different scales are input into the feature pyramid network to obtain a set of multi-scale feature maps; The set of multi-scale feature maps is fused to obtain the multi-scale fused feature map of the insulation material defect.

5. A visual inspection-based surface defect identification system for insulating materials, characterized in that, include: The surface image acquisition module is used to acquire the original surface image of the insulating material; The defect explicitation module is used to perform reflection suppression and texture primitive separation based on dual image decomposition on the original surface image of the insulating material to obtain a defect explicitation image of the insulating material, including: The original surface image of the insulating material is subjected to illuminance-reflectance decomposition and reflection suppression based on multi-scale Retinex to obtain the surface reflectance image of the insulating material, including: The illuminance estimation image of the original surface image of the insulating material is obtained by using the multi-scale Retinex algorithm; Based on the estimated illuminance component image of the insulating material surface and the original surface image of the insulating material, the surface reflectance image of the insulating material is calculated, including: based on the estimated illuminance component image of the insulating material surface and the original surface image of the insulating material, the logarithmic domain image of the surface reflectance of the insulating material is calculated using the following formula: ； in, In the original surface image of the insulating material The pixel value of the location, Estimating the illuminance components on the surface of insulating materials in the image The pixel value of the location, In the surface reflectance image of insulating material The pixel value of the location, Represents the logarithmic function with base 2. A constant hyperparameter used to prevent logarithmic zero. In the logarithmic domain image representing the surface reflectance of insulating materials The element value of the position; convert the logarithmic domain image of the surface reflectance of the insulating material back to the original domain to obtain the surface reflectance image of the insulating material; The surface reflectance image of the insulating material is subjected to texture and defect separation based on low-rank and sparse decomposition to obtain the surface texture image and the initial defect layer image of the insulating material. Noise suppression is applied to the initial defect layer image on the surface of the insulating material to obtain the surface defect layer image of the insulating material; The surface reflectance image of the insulating material, the surface texture image of the insulating material, and the surface defect layer image of the insulating material are fused to obtain a defect-revealed image of the insulating material; A multi-scale feature extraction module is used to input the explicit image of the insulation material defects into a multi-scale feature extraction model containing a backbone network and a feature pyramid network to obtain a multi-scale fused feature map of the insulation material defects. The defect detection module is used to input the multi-scale fusion feature map of the insulation material defects into a lightweight detection head to obtain candidate defect regions, their defect type labels, and confidence levels, including: The optimization of the insulation material defect multi-scale fusion feature map by performing multi-scale feature coupling noise suppression on the multi-scale fusion feature map includes: adjusting the multi-scale fusion feature manifold structure matching of the multi-scale fusion feature map to obtain an adjusted multi-scale fusion feature map; based on the residual response between the adjusted multi-scale fusion feature map and the original multi-scale fusion feature map, performing high-frequency interference orthogonal scaling on the adjusted multi-scale fusion feature map to obtain a noise-reduced multi-scale fusion feature map; and orthogonally fusing the noise-reduced multi-scale fusion feature map and the original multi-scale fusion feature map to obtain the optimized multi-scale fusion feature map. The optimized multi-scale fusion feature map of insulation material defects is input into a lightweight detection head to obtain the defect candidate regions, their defect type labels, and confidence levels.

Images