A method and system for intelligent identification of asymmetric plate defects

Abstract

This application provides an intelligent identification method and system for asymmetric plate-shaped defects, comprising: acquiring multimodal image data of the plate surface to be detected, and preprocessing the multimodal image data to obtain a standardized image; extracting features from the standardized image based on an asymmetric feature enhancement algorithm to obtain an asymmetric feature vector; inputting the asymmetric feature vector into a pre-trained asymmetric defect identification model to generate preliminary defect classification results and a defect region heatmap; segmenting the defect boundary based on the defect region heatmap and a geometric constraint optimization algorithm to determine the morphological parameters and spatial location of the asymmetric defect; and correcting the preliminary defect classification results based on the morphological parameters and spatial location using a dynamic threshold adjustment algorithm to obtain the identification result, thereby alleviating the technical problem of low accuracy in asymmetric plate-shaped defect identification in the prior art.

Description

Technical Field

[0001] This application relates to the field of plate shape defect detection technology, and in particular to an intelligent identification method and system for asymmetric plate shape defects. Background Technology

[0002] In manufacturing sectors such as steel, automotive, electronics, and aerospace, sheet metal is widely used as a critical structural component. Due to various factors during production, processing, and transportation, sheet metal surfaces are prone to asymmetric defects such as cracks, scratches, dents, and wrinkles. These defects typically exhibit irregular shapes, strong directionality, and localized ambiguity, severely impacting the product's appearance quality and structural performance. Therefore, developing an intelligent detection technology capable of efficiently and accurately identifying asymmetric defects has become an urgent technical requirement in current industrial quality inspection systems.

[0003] Currently, the industry commonly uses traditional image processing methods, general target detection models, or single-modal analysis techniques based on visible light images for defect identification. While these methods improve detection efficiency to some extent, they still suffer from low recognition accuracy and inaccurate boundary localization when faced with complex backgrounds, noise interference, and highly asymmetric defects. Furthermore, existing solutions often rely on fixed feature extraction methods, making it difficult to effectively characterize the unique spatial distribution and texture details of asymmetric defects, thus limiting their applicability and stability in real-world scenarios.

[0004] With the development of intelligent manufacturing, industrial inspection systems are increasingly demanding more intelligent and precise defect identification. Traditional methods are proving inadequate in dealing with diverse defect types and complex environmental conditions, particularly lacking effective technical support in defect morphology modeling, classification, and dynamic classification correction. Therefore, there is an urgent need to propose a novel intelligent identification method that integrates multimodal image information, possesses asymmetric feature enhancement capabilities, and can achieve precise defect boundary positioning and dynamic optimization of classification results, in order to meet the pressing needs of modern manufacturing for high-quality, automated inspection. Summary of the Invention

[0005] This application provides an intelligent identification method and system for asymmetric plate-shaped defects.

[0006] In a first aspect, this application provides an intelligent identification method for asymmetric plate defects, comprising: acquiring multimodal image data of the plate surface to be detected, and preprocessing the multimodal image data to obtain a standardized image; extracting features from the standardized image based on an asymmetric feature enhancement algorithm to obtain an asymmetric feature vector, wherein the asymmetric feature vector includes: spatial gradient distribution, local texture entropy, and curvature change; inputting the asymmetric feature vector into a pre-trained asymmetric defect identification model to generate a preliminary defect classification result and a defect region heatmap; segmenting the defect boundary based on the defect region heatmap and a geometric constraint optimization algorithm to determine the morphological parameters and spatial location of the asymmetric defect; and correcting the preliminary defect classification result based on the morphological parameters and the spatial location using a dynamic threshold adjustment algorithm to obtain an identification result, wherein the identification result includes: asymmetric defect type, level, and location information.

[0007] Furthermore, feature extraction is performed on the standardized image based on an asymmetric feature enhancement algorithm to obtain an asymmetric feature vector, including: performing multi-scale pyramid decomposition on the standardized image to generate a target image, wherein the target image includes: low-frequency sub-images, mid-frequency sub-images, and high-frequency sub-images; calculating the histogram of oriented gradients and local binary patterns for each sub-image to generate HOG feature maps and LBP feature maps; fusing the HOG feature maps and LBP feature maps pixel by pixel to generate a fused feature map; and dynamically weighting the feature channels in the fused feature map based on an asymmetric weight allocation function to generate the asymmetric feature vector, wherein the weight of the high-frequency sub-images is higher than the weight of the low-frequency sub-images.

[0008] Furthermore, the multimodal image data is preprocessed to obtain a standardized image, including: using a median filtering algorithm to perform noise removal processing on the multimodal image data to generate a denoised image; using adaptive histogram equalization to enhance the contrast of the denoised image to generate an enhanced image; using a bicubic interpolation algorithm to unify the resolution of the enhanced image to generate a normalized image; and using a grayscale normalization formula to map pixel values ​​of the normalized image to generate the standardized image.

[0009] Furthermore, the standardized image is subjected to multi-scale pyramid decomposition to generate the target image, including: performing three-level downsampling processing on the standardized image based on a Gaussian kernel to generate the low-frequency sub-image; extracting high-frequency residuals from the low-frequency sub-image using the Laplacian operator to generate the high-frequency sub-image; extracting edge features from the low-frequency sub-image using a bandpass filtering algorithm to generate the mid-frequency sub-image; and stacking the low-frequency sub-image, the mid-frequency sub-image, and the high-frequency sub-image in a pyramid structure to generate the target image.

[0010] Furthermore, the HOG feature map and the LBP feature map are fused pixel by pixel to generate a fused feature map, including: performing Min-Max normalization on the HOG feature map and the LBP feature map to obtain normalized HOG feature maps and normalized LBP feature maps; performing matrix concatenation on the normalized HOG feature maps and the normalized LBP feature maps to generate a multi-channel fusion matrix; using a convolution kernel of a preset size to perform feature calibration on the fusion matrix to generate an attention feature map; and performing multi-scale feature extraction on the attention feature map through spatial pyramid pooling to generate the fused feature map.

[0011] Furthermore, the asymmetric feature vector is generated by dynamically weighting the feature channels in the fused feature map based on an asymmetric weight allocation function, including: calculating the variance of the high-frequency sub-image feature channels and using the variance as a weight benchmark value; performing attenuation processing on the weights of the mid-frequency sub-image feature channels and the low-frequency sub-image feature channels according to the weight benchmark value and a preset exponential decay function to obtain attenuated weight values; performing nonlinear mapping on the attenuated weight values ​​through a sigmoid function to obtain a mapped weight matrix; and performing a Hadamard product operation between the mapped weight matrix and the fused feature map to generate the asymmetric feature vector.

[0012] Furthermore, based on the heatmap of the defect region, the defect boundary is segmented using a geometric constraint optimization algorithm to determine the morphological parameters and spatial location of the asymmetric defect. This includes: evolving the initial contour of the defect region heatmap using a level set algorithm to obtain an initial contour curve; constructing an energy functional constraint equation for the initial contour curve based on a plate-shaped prior geometric model; solving the minimization problem of the energy functional constraint equation using the finite difference method to obtain an optimized boundary curve; and quantifying the defect morphology and spatial location based on the curvature extrema calculation and bounding box parameter extraction of the optimized boundary curve to generate the morphological parameters and spatial location.

[0013] Secondly, embodiments of this application provide an intelligent identification system for asymmetric plate-shaped defects, comprising: a preprocessing unit for acquiring multimodal image data of the plate-shaped surface to be detected and preprocessing the multimodal image data to obtain a standardized image; an extraction unit for extracting features from the standardized image based on an asymmetric feature enhancement algorithm to obtain an asymmetric feature vector, wherein the asymmetric feature vector includes: spatial gradient distribution, local texture entropy, and curvature change; a generation unit for inputting the asymmetric feature vector into a pre-trained asymmetric defect identification model to generate preliminary defect classification results and a defect region heatmap; a determination unit for segmenting the defect boundary based on the defect region heatmap and a geometric constraint optimization algorithm to determine the morphological parameters and spatial location of the asymmetric defect; and a correction unit for correcting the preliminary defect classification results based on the morphological parameters and the spatial location using a dynamic threshold adjustment algorithm to obtain an identification result, wherein the identification result includes: asymmetric defect type, level, and location information.

[0014] Thirdly, embodiments of this application provide a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement the method described in the first aspect above.

[0015] Fourthly, embodiments of this application provide a computer storage medium storing a computer program, which, when executed by a computer, implements the method described in the first aspect above.

[0016] The present invention provides an intelligent identification method and system for asymmetric plate defects, which has the following significant advantages:

[0017] By introducing multimodal image data fusion and standardized preprocessing techniques, the effects of uneven illumination and noise interference can be effectively eliminated, improving image quality and thus enhancing the accuracy of subsequent feature extraction and defect identification.

[0018] By employing an asymmetric feature enhancement algorithm based on spatial gradient distribution, local texture entropy, and curvature variation, we can accurately capture the unique characteristics of asymmetric defects in shape, edge, and surface texture, overcoming the problem of insufficient description of complex defect features by traditional methods.

[0019] By combining geometric constraint optimization algorithms to segment the boundary of the defect heatmap, the accuracy of defect region location is improved, and the morphological parameters (such as length, width, curvature) and spatial location information of the defect can be obtained, providing a reliable basis for subsequent quality assessment and decision-making.

[0020] By using a dynamic threshold adjustment algorithm based on morphological parameters and spatial location, the preliminary classification results are corrected through feedback, which effectively reduces the false detection rate and the false negative rate, and improves the overall stability and adaptability of the system.

[0021] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating an intelligent method for identifying asymmetric plate defects provided in an embodiment of this application;

[0024] Figure 2 A schematic diagram of an intelligent recognition system for asymmetric plate defects provided in an embodiment of this application;

[0025] Figure 3 This application provides a schematic diagram of a computing device. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1: According to an embodiment of the present invention, an embodiment of an intelligent identification method for asymmetric plate shape defects is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0028] Figure 1 This is a flowchart of an intelligent identification method for asymmetric plate shape defects according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0029] Step S101: Collect multimodal image data of the surface of the plate to be detected, and preprocess the multimodal image data to obtain a standardized image;

[0030] Step S102: Based on the asymmetric feature enhancement algorithm, feature extraction is performed on the standardized image to obtain an asymmetric feature vector, wherein the asymmetric feature vector includes: spatial gradient distribution, local texture entropy, and curvature change;

[0031] Step S103: Input the asymmetric feature vector into the pre-trained asymmetric defect recognition model to generate preliminary defect classification results and defect region heatmaps.

[0032] Step S104: Based on the heat map of the defect area, the defect boundary is segmented using a geometric constraint optimization algorithm to determine the morphological parameters and spatial location of the asymmetric defect.

[0033] Step S105: Based on the morphological parameters and the spatial location, the preliminary defect classification result is corrected by a dynamic threshold adjustment algorithm to obtain the identification result, wherein the identification result includes: asymmetric defect type, level and location information.

[0034] In this embodiment of the invention, multimodal image data (including visible light, laser contour, infrared thermal imaging, and other multi-source images) of the plate surface to be detected are acquired using an industrial linear scan camera (such as SICK Ranger E). The images are then denoised using the NL-Means algorithm, and illumination differences are eliminated by grayscale normalization. Finally, dynamic threshold segmentation is used to separate the background and target regions, resulting in a standardized image with uniform size and enhanced contrast (such as a grayscale image with a resolution of 0.14 mm / pixel), providing high-quality input for subsequent feature extraction.

[0035] Next, the gradient magnitudes in the X / Y directions of the image are calculated using the Sobel operator to generate a gradient direction histogram (HOG) to describe edge asymmetry. The entropy of the gray-level co-occurrence matrix (GLCM) is calculated within a sliding window to quantify texture disorder and identify chipped edges and scratches. The surface contour is fitted using B-spline curves, and the location and rate of change of curvature extrema are calculated by differentiation. The above features are fused to generate an asymmetric feature vector (dimension 1024×1), that is, the gradient distribution reflects stress distortion, the texture entropy captures microscopic damage, and the curvature quantifies macroscopic deformation.

[0036] Next, the asymmetric feature vector is input into the pre-trained asymmetric defect recognition model, which uses hybrid nonlinear convolution (standard convolution + Gaussian convolution) to extract cross-scale features; preliminary defect classification results (such as the probability of categories like sickle bends and edge-in-center composite waves) are generated through the Softmax output layer; at the same time, gradient-weighted class activation mapping (Grad-CAM) is used to generate a heat map of the defect region, and the probability distribution of defects is identified by the heat value.

[0037] Then, using the peak point of the heat map as the initial contour, a horizontal energy functional equation is constructed; prior constraints on plate stress distribution (such as the rolling tension model) are introduced, and the contour evolution is solved iteratively through the SLSQP optimizer; a smooth defect boundary is output, and the morphological parameters (curvature extrema, aspect ratio) and spatial position (center coordinates of the bounding box) are calculated, thereby locating the error.

[0038] Finally, based on the curvature extreme value, the category determination threshold is calculated according to the formula: threshold = k × curvature extreme value + b (k and b are fitted from historical data); when the initial classification confidence is lower than the threshold, the adaptive genetic algorithm is triggered to reassign the category weights; the final identification result is output, including defect type (such as "1 / 4 wave"), level (level 1-5), and location information (pixel coordinates + physical coordinate transformation).

[0039] In this application, multimodal image data refers to multidimensional information of a plate surface collected by different sensors (visible light camera, laser scanner, infrared thermal imager), including texture, depth, and temperature field data, which is used to complement and enhance defect features.

[0040] Standardized image: refers to a uniform format image after denoising, normalization, and size cropping, with a size of 1024×1024 pixels and a grayscale range of [0,255], eliminating environmental interference.

[0041] Asymmetric feature vector: A feature tensor with a dimension of 1024×1, composed of spatial gradient distribution (HOG vector), local texture entropy (GLCM entropy matrix), and curvature change (B-spline derivative sequence), used to quantify asymmetric deformation.

[0042] Defect area heat map: A color probability map generated by Grad-CAM. The heat value (0~1) reflects the probability that a pixel belongs to a defect area, and the red highlighted area is the core of the defect.

[0043] Geometric constraint optimization algorithm: The level set contour evolution algorithm integrates the mechanical stiffness equation and the thermal expansion coefficient, and achieves physical-driven segmentation through energy functional minimization.

[0044] Morphological parameters include curvature extrema (maximum curvature point value) and aspect ratio (defect bounding box width / height), used to quantify defect geometric properties.

[0045] Spatial location: The coordinates of the defect centroid in the plate coordinate system (unit: mm), obtained by affine transformation of pixel coordinates.

[0046] Dynamic threshold adjustment algorithm: an adaptive threshold calculation model based on morphological parameters.

[0047] In this embodiment of the invention, the directionality of stress distortion (such as the unidirectional aggregation of gradients in a single-sided wave) is captured by spatial gradient distribution; the microscopic disorder of edge breakage / scratches is quantified by local texture entropy; and the distribution of macroscopic curvature extrema points of wave shape is analyzed by curvature change. This solves the problems of traditional defect identification methods that rely on symmetric basis functions for polynomial fitting, which cannot describe asymmetric patterns such as 1 / 4 waves and edge-in-center composite waves, and that conventional features (such as grayscale statistics) ignore curvature gradient changes.

[0048] In this embodiment of the invention, feature extraction is performed on the standardized image based on an asymmetric feature enhancement algorithm to obtain an asymmetric feature vector, including:

[0049] The standardized image is decomposed into a multi-scale pyramid to generate a target image, wherein the target image includes: a low-frequency sub-image, a mid-frequency sub-image, and a high-frequency sub-image;

[0050] For each sub-image, calculate the histogram of oriented gradients and the local binary pattern to generate HOG feature maps and LBP feature maps;

[0051] The HOG feature map and the LBP feature map are fused pixel by pixel to generate a fused feature map;

[0052] The asymmetric feature vector is generated by dynamically weighting the feature channels in the fused feature map based on an asymmetric weight allocation function, wherein the weight of the high-frequency sub-image is higher than the weight of the low-frequency sub-image.

[0053] In this embodiment of the invention, the standardized image is continuously subjected to Gaussian filtering and downsampling five times to generate five low-frequency sub-images, with the resolution of each layer decreasing by 50%. A Laplacian pyramid is constructed: the i-th layer Gaussian image is upsampled to the size of the (i-1)-th layer, and then subtracted pixel by pixel from the (i-1)-th layer Gaussian image to generate a mid-frequency sub-image (edge ​​contour) and a high-frequency sub-image (texture details). The target image is output as follows: low-frequency sub-image (5th layer Gaussian image), mid-frequency sub-image (3rd-4th layer Laplacian image), and high-frequency sub-image (1st-2nd layer Laplacian image), forming an image group covering the entire frequency band.

[0054] Next, the sub-image is divided into 8×8 pixel cells, and the pixel gradient direction is calculated in each cell. The gradient magnitude is used as the weight to generate a direction histogram. Every 4 cells form a 16×16 block and L2-norm normalization is performed. All block histograms are concatenated to generate the HOG feature map.

[0055] use The operator compares the gray values ​​of the neighborhood for each pixel to generate an 8-bit binary code, calculates the 59-dimensional histogram of each 16×16 window, concatenates all window histograms, and generates an LBP feature map.

[0056] Then, the HOG feature maps and LBP feature maps of the same scale sub-images are adjusted to the same size, and then weighted and added pixel by pixel according to the formula: fusion feature map = 0.6×HOG feature map + 0.4×LBP feature map, to output a multi-channel fusion feature map.

[0057] Finally, the gradient energy variance of the fused feature map of each sub-image is calculated, the channel weights are assigned according to the asymmetric weighting function, the weighted multi-channel feature map is flattened into a one-dimensional tensor, and the final asymmetric feature vector is generated.

[0058] The above steps improve the detection rate of high-frequency defects, and the fusion of HOG and LBP overcomes the limitations of single features. For example, it can effectively reduce the false alarm rate under oil pollution interference.

[0059] In this embodiment of the invention, the multimodal image data is preprocessed to obtain a standardized image, including:

[0060] The multimodal image data is subjected to noise reduction processing using a median filtering algorithm to generate a denoised image;

[0061] The denoised image is enhanced by adaptive histogram equalization to generate an enhanced image.

[0062] The enhanced image is normalized by bicubic interpolation algorithm to generate a normalized image.

[0063] The normalized image is generated by mapping pixel values ​​to the normalized image using a grayscale normalization formula.

[0064] Specifically, firstly, a 5×5 pixel window is used to traverse the image. The window size is odd to ensure the existence of a center point. The gray values ​​of the 25 pixels in the window are sorted in ascending order, and the 13th bit (median) is used to replace the center pixel value. A mirror filling strategy is used for the edge pixels of the image to avoid information loss. The denoised image is output to eliminate isolated noise caused by dust and electromagnetic interference in the industrial environment, while preserving the sharpness of defect edges.

[0065] Next, the image is divided into 8×8 tiles, each with a size of 128×128 pixels. The histogram of each tile is calculated, a contrast limit threshold is set, and the excess is evenly distributed to each gray level. The cumulative distribution function (CDF) of each tile is calculated, and a gray-level mapping table is generated according to the formula: mapping value = (current CDF value - minimum CDF value) / (total number of pixels - minimum CDF value) × 255. The pixel values ​​of the boundary areas of the tiles are smoothly transitioned, and the enhanced image is output to solve the local detail blur caused by uneven reflection on the steel plate surface.

[0066] Based on the target resolution of 1024×1024, calculate the width / height scaling factor (if the original image size is 2048×1536, then the scaling factor W=0.5, H=0.67). For the target pixel (x,y), select 16 pixels in a 4×4 neighborhood. Calculate the distance weight (a=-0.5, d is the relative distance) using a cubic function weight = f(d) = [a×d³ - 5a×d² + 8a×d - 4a] and sum the neighborhood pixel values ​​to generate a normalized image, eliminating scale differences caused by multi-camera acquisition.

[0067] Finally, the image mean μ and standard deviation σ are extracted, and the standardized pixel value is calculated using the formula: Standardized pixel value = (Original pixel value - μ) / σ × New standard deviation + New mean. Values ​​exceeding [0, 255] are truncated to the boundary, and the standardized image is output, so that images acquired under different lighting conditions have a consistent grayscale distribution.

[0068] The above steps, through the coordinated efforts of noise suppression, contrast enhancement, resolution unification, and grayscale normalization, provide anti-interference capabilities, optimize the accuracy of weak defect detection, and enhance cross-device compatibility.

[0069] In this embodiment of the invention, multi-scale pyramid decomposition is performed on the standardized image to generate a target image, including:

[0070] The normalized image is subjected to three-layer downsampling processing based on Gaussian kernel to generate the low-frequency sub-image;

[0071] The high-frequency sub-image is generated by extracting high-frequency residuals from the low-frequency sub-image using the Laplacian operator.

[0072] The low-frequency sub-image is used to extract edge features through a bandpass filtering algorithm to generate the mid-frequency sub-image;

[0073] The target image is generated by stacking the low-frequency sub-image, the mid-frequency sub-image, and the high-frequency sub-image in a pyramid structure.

[0074] Specifically, a 5×5 Gaussian kernel (σ=1.6) is used to convolve and smooth the normalized image, and interlaced sampling with a stride of 2 is performed (keeping odd rows / columns) to generate a 1 / 4 size image. Gaussian convolution with a stride of 2 is repeatedly performed on the output, reducing the size of the second layer to 1 / 16 and the third layer to 1 / 64. The output of the third layer is a low-frequency sub-image with the lowest resolution (e.g., 1024×1024→128×128), preserving the plate-shaped macro structure but filtering out high-frequency details.

[0075] Next, the low-frequency sub-image of the third layer is upsampled to the size of the second layer using bicubic interpolation. The high-frequency residual is calculated by subtracting the high-frequency sub-image pixel by pixel according to the formula: high-frequency residual = second-layer downsampled image - upsampled image. This process highlights the details that were not preserved in the low-frequency sub-image. The high-frequency sub-image is generated after the absolute value of the residual is truncated (0~255), which includes 0.1mm level defects such as microcracks and chipped edges.

[0076] Then, the second-layer downsampled image is Fourier transformed using a frequency band filter, and then multiplied by the frequency domain distance of the frequency band filter and inversely transformed back to the spatial domain. The filtered result generates a mid-frequency sub-image, which retains edge features with a width of 1-5mm such as single-sided waves and sickle bends.

[0077] Finally, the high-frequency sub-image (original image size) is downsampled to 128×128 using bicubic interpolation, and the mid-frequency sub-image is downsampled to the same size. The low-frequency sub-image (128×128), mid-frequency sub-image (128×128), and high-frequency sub-image (128×128) are stitched together according to the channel dimension to generate a three-channel target image. Channel 1 is the low-frequency structure, channel 2 is the mid-frequency edge, and channel 3 is the high-frequency texture.

[0078] The above steps, through three-layer Gaussian downsampling, Laplace residual extraction, bandpass filtering separation, and pyramid stacking, achieve multi-scale defect decoupling, enhance noise resistance, and optimize computational efficiency.

[0079] In this embodiment of the invention, the HOG feature map and the LBP feature map are fused pixel by pixel to generate a fused feature map, including:

[0080] The HOG feature map and the LBP feature map are subjected to Min-Max normalization to obtain normalized HOG feature map and normalized LBP feature map.

[0081] The normalized HOG feature map and the normalized LBP feature map are matrix concatenated to generate a multi-channel fusion matrix;

[0082] The fusion matrix is ​​characterized using convolution kernels of a preset size to generate an attention feature map;

[0083] Multi-scale feature extraction is performed on the attention feature map using spatial pyramid pooling to generate a fused feature map.

[0084] Specifically, extract the maximum value (max) of all pixel values ​​in the HOG feature map. HOG and minimum value min HOG Similarly, the max of the LBP feature map is calculated. LBP and min LBP The normalized value is calculated using the formula: Normalized value = (Original pixel value - Minimum value) / (Maximum value - Minimum value). The feature values ​​are scaled pixel by pixel and mapped to the [0,1] interval to generate normalized HOG feature maps and normalized LBP feature maps, eliminating dimensional differences and ensuring numerical consistency in subsequent fusion.

[0085] Next, the HOG feature map and the LBP feature map are concatenated along the channel dimension to generate a multi-channel fusion matrix. Bilinear interpolation is used to ensure that the spatial dimensions of the two feature maps are consistent. Channels 1 to 324 of the fusion matrix are HOG features, and channels 325 to 383 are LBP features, forming a tensor that is complementary to texture and shape.

[0086] Then, a fusion matrix is ​​scanned using a 3×3 convolution kernel (stride 1, padding 1), and a weighted sum of features within the local receptive field is calculated. The ReLU function output = max(0, convolution result) is used to enhance the nonlinear expressive power. Next, an attention feature map is generated, with high-response regions identifying defect-sensitive areas (such as crack edges).

[0087] Finally, the feature map is divided into three grid layers: 1×1, 2×2, and 4×4 (e.g., if the input size is 128×128, the grid sizes are 128×128, 64×64, and 32×32 respectively). Max pooling is performed on each grid cell to extract the most salient features of that region (e.g., 1×1 layer outputs 1 feature value, 2×2 layer outputs 4, and 4×4 layer outputs 16). All grid pooling results are concatenated to generate a fixed-length fused feature map (e.g., 1+4+16=21 dimensions), thus adapting to any input size.

[0088] The above steps, through normalization to eliminate dimensional differences, dual feature channel complementarity, convolutional attention focusing, and multi-scale pooling adaptation, can improve the detection rate of composite defects and achieve dynamic size compatibility.

[0089] In this embodiment of the invention, the feature channels in the fused feature map are dynamically weighted based on an asymmetric weight allocation function to generate the asymmetric feature vector, including:

[0090] Calculate the variance of the high-frequency sub-image feature channels and use the variance as the weighting benchmark value;

[0091] Based on the weight benchmark value and the preset exponential decay function, the weights of the mid-frequency sub-image feature channel and the weights of the low-frequency sub-image feature channel are attenuated to obtain the attenuated weight values.

[0092] The attenuated weight values ​​are nonlinearly mapped using the sigmoid function to obtain the mapped weight matrix;

[0093] The mapped weight matrix is ​​subjected to a Hadamard product operation with the fused feature map to generate the asymmetric feature vector.

[0094] Specifically, feature channels corresponding to high-frequency sub-images are separated from the fused feature map. The variance of all pixel values ​​in each channel is calculated using the formula: Variance = Sum of squares of all pixel gradient magnitudes minus the mean ÷ Total number of pixels. The arithmetic mean of the variances of all high-frequency channels (e.g., mean = 105.3) is taken as the weight benchmark value for subsequent weight allocation, which reflects the significance level of high-frequency details.

[0095] Next, the initial weight of the mid-frequency sub-image feature channel is set to 0.5, and the initial weight of the low-frequency sub-image is set to 0.3. The weight is then attenuated according to the formula: initial weight × e^(-attenuation coefficient × weight baseline value) (attenuation coefficient = 0.01). For example, the mid-frequency weight is attenuated to 0.5 × e^(-0.01 × 105.3) = 0.18, and the low-frequency weight is attenuated to 0.3 × e^(-0.01 × 105.3) = 0.11. This process generates attenuated weight values ​​to ensure that the weight of the high-frequency channel is significantly higher than that of the mid- and low-frequency channels.

[0096] Then, the attenuated mid-frequency weight value (0.18) and low-frequency weight value (0.11) are arranged in channel order and calculated according to the formula: Mapping value = 1 ÷ (1 + e^(-attenuation weight)). For example, the mid-frequency weight is mapped to 1 / (1+e^(-0.18))=0.54, the low-frequency weight is mapped to 1 / (1+e^(-0.11))=0.52, and the high-frequency channel weight is fixed at 1.0. These are then concatenated with the mapped mid- and low-frequency weights to form the mapped weight matrix.

[0097] Finally, the mapped weight matrix is ​​adjusted to the same spatial size as the fused feature map through bilinear interpolation. The weighted feature is calculated element by element according to the formula: Weighted feature = Pixel value of fused feature map × Channel value of corresponding weight matrix. The weighted three-dimensional tensor is flattened into a one-dimensional asymmetric feature vector (e.g., 128×128×855 → 1,396,480 dimensions).

[0098] The above steps, through the synergistic effect of four steps—variance-driven weighting benchmark, exponentially decaying weighting, Sigmoid nonlinear enhancement, and Hadamard product dynamic weighting—improve the high-frequency defect detection rate, enhance resistance to background interference, and optimize feature discriminability.

[0099] In this embodiment of the invention, based on the heat map of the defect region, and combined with a geometric constraint optimization algorithm, the defect boundary is segmented to determine the morphological parameters and spatial location of the asymmetric defect, including:

[0100] Based on the level set algorithm, the initial contour of the heat map of the defect region is evolved to obtain the initial contour curve.

[0101] Based on the plate-shaped prior geometric model, an energy functional constraint equation is constructed for the initial contour curve;

[0102] The minimization problem of the energy functional constraint equation is solved using the finite difference method to obtain the optimized boundary curve.

[0103] Based on the curvature extremum point calculation and bounding box parameter extraction of the optimized boundary curve, the defect morphology and spatial location are quantified to generate the morphological parameters and spatial location.

[0104] Specifically, pixel regions in the defect heatmap above a threshold of 0.7 are used as initial contours. A signed distance function φ (φ>0 indicates outside the contour, φ<0 indicates inside the contour) is embedded to generate an implicit representation of the initial contour curve. This is combined with edge gradient force (extracted by the Canny operator) and region fitting force (local gray-level mean difference) according to the partial differential equation. (α·curvature - β·image gradient) Iteratively updates φ (α=0.1 is the curvature weight, β=0.9 is the gradient weight), updating only within the zero level set ±2 pixels of φ. After 10-15 iterations, it outputs a smooth and closed initial contour curve, which can handle topological changes such as cracks and bifurcations.

[0105] Next, we define the energy terms, which include the bending potential energy term and the geometric matching term.

[0106] Bending potential energy term: Based on Kirchhoff's thin plate theory, according to the formula Penalize non-smooth boundaries (w is the contour normal displacement);

[0107] Geometric matching term: Calculate the Hausdorff distance d_H = max{sup inf ||xy||} between the contour point and the preset plate shape template (such as a sine wave shape). The constraint conforms to the deformation law of the rolling process. The total energy functional E_total = γ·E_bend + λ·d_H (γ=0.3 is the smoothing weight, λ=0.7 is the geometric matching weight), forming an energy functional constraint equation with prior knowledge.

[0108] Then, the profile curve is discretized into 200 nodes, with a spatial step size Δx = 0.1 mm and a time step size Δt = 0.01 s. A difference grid is designed, where the curvature term adopts central difference. The gradient term uses forward difference ∂ Through explicit Euler method Update nodes when the rate of energy change is <10⁻ 6 The iteration may terminate after 300 iterations, outputting the optimized boundary curve.

[0109] Finally, sampling points are taken along the curve every 0.1 mm, and the curvature is calculated using the formula κ = |x'y'' - y'x''| / (x'² + y'²)^{3 / 2} to identify the curvature. The extreme points are used as defect corner points. The minimum bounding rectangle is constructed with the extreme points of curvature as vertices. The major axis L (maximum side length), minor axis W (minimum side length), and tilt angle θ (angle between the major axis and the horizontal) are extracted. The centroid coordinates of the bounding box (x_c, y_c) = (Σx_i / 4, Σy_i / 4) are calculated. Combined with the image pixel resolution (0.1mm / px), the coordinates are mapped to the steel plate coordinate system to generate morphological parameters (L, W, θ) and spatial position data (x_c, y_c).

[0110] The above steps improve the accuracy of weak boundary segmentation and enhance the decoupling capability of composite defects through four collaborative steps: level set adaptive evolution, prior energy constraint, differential efficient solution, and curvature-bounding box quantization.

[0111] Example 2: This embodiment of the invention also provides an intelligent identification system for asymmetric plate defects. This intelligent identification system for asymmetric plate defects is used to execute the intelligent identification method for asymmetric plate defects provided in the above-described embodiments of the invention. The following is a detailed description of the intelligent identification system for asymmetric plate defects provided in this embodiment of the invention.

[0112] like Figure 2 As shown, Figure 2 This is a schematic diagram of the intelligent identification system for the aforementioned asymmetric plate defects. The system includes:

[0113] The preprocessing unit 10 is used to acquire multimodal image data of the surface of the plate to be detected, and to preprocess the multimodal image data to obtain a standardized image;

[0114] Extraction unit 20 is used to extract features from the standardized image based on an asymmetric feature enhancement algorithm to obtain an asymmetric feature vector, wherein the asymmetric feature vector includes: spatial gradient distribution, local texture entropy, and curvature change;

[0115] The generation unit 30 is used to input the asymmetric feature vector into a pre-trained asymmetric defect recognition model to generate preliminary defect classification results and defect region heatmaps.

[0116] The determining unit 40 is used to segment the defect boundary based on the defect region heat map and a geometric constraint optimization algorithm, and to determine the morphological parameters and spatial location of the asymmetric defect.

[0117] The correction unit 50 is used to correct the preliminary defect classification result based on the morphological parameters and the spatial position using a dynamic threshold adjustment algorithm to obtain the identification result, wherein the identification result includes: asymmetric defect type, level and location information.

[0118] In embodiment three, this invention also provides a computing device for executing a program of the method described in embodiment one above, such as... Figure 3 As shown, the computing device may include a storage component 41 and a processing component 42;

[0119] The storage component 41 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 42.

[0120] The processing component 42 may include one or more processors to execute computer instructions to complete all or part of the steps in the method of Embodiment 1. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0121] Storage component 41 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0122] Display component 43 may be an electroluminescent (EL) element, a liquid crystal display or a microdisplay with a similar structure, or a retina-direct display or a similar laser scanning display.

[0123] Of course, computing devices may also include other components, such as input / output interfaces and communication components.

[0124] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0125] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0126] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0127] Example 4: This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The method of the illustrated embodiment.

[0128] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0129] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0130] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligent identification of asymmetric plate defects, characterized in that, include: Multimodal image data of the surface of the plate to be inspected is acquired, and the multimodal image data is preprocessed to obtain a standardized image; The standardized image is subjected to feature extraction based on an asymmetric feature enhancement algorithm to obtain an asymmetric feature vector, wherein the asymmetric feature vector includes: spatial gradient distribution, local texture entropy and curvature change; The asymmetric feature vector is input into a pre-trained asymmetric defect recognition model to generate preliminary defect classification results and defect region heatmaps. Based on the heat map of the defect region, the defect boundary is segmented using a geometric constraint optimization algorithm to determine the morphological parameters and spatial location of the asymmetric defect. Based on the morphological parameters and the spatial location, the preliminary defect classification result is corrected by a dynamic threshold adjustment algorithm to obtain the identification result, wherein the identification result includes: asymmetric defect type, level and location information; The standardized image is subjected to feature extraction based on an asymmetric feature enhancement algorithm to obtain an asymmetric feature vector, including: The standardized image is decomposed into a multi-scale pyramid to generate a target image, wherein the target image includes: a low-frequency sub-image, a mid-frequency sub-image, and a high-frequency sub-image; For each sub-image, calculate the histogram of oriented gradients and the local binary pattern to generate HOG feature maps and LBP feature maps; The HOG feature map and the LBP feature map are fused pixel by pixel to generate a fused feature map; The asymmetric feature vector is generated by dynamically weighting the feature channels in the fused feature map based on an asymmetric weight allocation function, wherein the weight of the high-frequency sub-image is higher than the weight of the low-frequency sub-image. The HOG feature map and the LBP feature map are fused pixel by pixel to generate a fused feature map, including: The HOG feature map and the LBP feature map are subjected to Min-Max normalization to obtain normalized HOG feature map and normalized LBP feature map. The normalized HOG feature map and the normalized LBP feature map are matrix concatenated to generate a multi-channel fusion matrix; The fusion matrix is ​​characterized using convolution kernels of a preset size to generate an attention feature map; Multi-scale feature extraction is performed on the attention feature map using spatial pyramid pooling to generate a fused feature map.

2. The method according to claim 1, characterized in that, The multimodal image data is preprocessed to obtain a standardized image, including: The multimodal image data is subjected to noise reduction processing using a median filtering algorithm to generate a denoised image; The denoised image is enhanced by adaptive histogram equalization to generate an enhanced image. The enhanced image is normalized by bicubic interpolation algorithm to generate a normalized image. The normalized image is generated by mapping pixel values ​​to the normalized image using a grayscale normalization formula.

3. The method according to claim 1, characterized in that, The standardized image is subjected to multi-scale pyramid decomposition to generate the target image, including: The normalized image is subjected to three-layer downsampling processing based on Gaussian kernel to generate the low-frequency sub-image; The high-frequency sub-image is generated by extracting high-frequency residuals from the low-frequency sub-image using the Laplacian operator. The low-frequency sub-image is used to extract edge features through a bandpass filtering algorithm to generate the mid-frequency sub-image; The target image is generated by stacking the low-frequency sub-image, the mid-frequency sub-image, and the high-frequency sub-image in a pyramid structure.

4. The method according to claim 1, characterized in that, The asymmetric feature vector is generated by dynamically weighting the feature channels in the fused feature map based on an asymmetric weight allocation function, including: Calculate the variance of the high-frequency sub-image feature channels and use the variance as the weighting benchmark value; Based on the weight benchmark value and the preset exponential decay function, the weights of the mid-frequency sub-image feature channel and the weights of the low-frequency sub-image feature channel are attenuated to obtain the attenuated weight values. The attenuated weight values ​​are nonlinearly mapped using the sigmoid function to obtain the mapped weight matrix; The mapped weight matrix is ​​subjected to a Hadamard product operation with the fused feature map to generate the asymmetric feature vector.

5. The method according to claim 1, characterized in that, Based on the heat map of the defect region, and combined with a geometric constraint optimization algorithm, the defect boundary is segmented to determine the morphological parameters and spatial location of the asymmetric defect, including: Based on the level set algorithm, the initial contour of the heat map of the defect region is evolved to obtain the initial contour curve. Based on the plate-shaped prior geometric model, an energy functional constraint equation is constructed for the initial contour curve; The minimization problem of the energy functional constraint equation is solved using the finite difference method to obtain the optimized boundary curve. Based on the curvature extremum point calculation and bounding box parameter extraction of the optimized boundary curve, the defect morphology and spatial location are quantified to generate the morphological parameters and spatial location.

6. An intelligent identification system for asymmetric plate defects, characterized in that, The intelligent identification method for asymmetric plate defects as described in claim 1 includes: The preprocessing unit is used to acquire multimodal image data of the surface of the plate to be detected, and to preprocess the multimodal image data to obtain a standardized image; The extraction unit is used to extract features from the standardized image based on the asymmetric feature enhancement algorithm to obtain an asymmetric feature vector, wherein the asymmetric feature vector includes: spatial gradient distribution, local texture entropy, and curvature change; The generation unit is used to input the asymmetric feature vector into a pre-trained asymmetric defect recognition model to generate preliminary defect classification results and defect region heatmaps. The determining unit is used to segment the defect boundary based on the defect region heat map and a geometric constraint optimization algorithm, and to determine the morphological parameters and spatial location of the asymmetric defect. The correction unit is used to correct the preliminary defect classification result based on the morphological parameters and the spatial position using a dynamic threshold adjustment algorithm to obtain the identification result, wherein the identification result includes: asymmetric defect type, level and location information.

7. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement the method as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements the method as described in any one of claims 1 to 5.

Images