A deep learning-based cable sheath micro-crack image recognition method

Abstract

The application discloses a kind of cable sheath micro crack image recognition methods based on deep learning, including the following steps: cable sheath image is collected and image pre-processing operation is executed;Input to improved MAE model, generate background reconstruction map and crack reconstruction map;Pixel-level residual fusion is carried out, and background mask map is generated;Pixel-level fusion is carried out to background mask map and pre-processing image, and background suppression map is generated;Micro crack identification operation is executed, and preliminary crack response heat map set is output by image feature extraction and crack area discrimination;Heat map cumulative analysis operation is executed, and multi-scale cumulative heat map is constructed;According to local response change rate, judge false response risk area;Response value retraction operation is carried out based on corresponding local area, and crack heat map after false response suppression is generated;High confidence crack area is extracted, and cable sheath micro crack recognition result is obtained.The application improves the precision and robustness of micro crack detection, reduces background interference and false detection risk.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method for recognizing microcracks in cable sheaths based on deep learning. Background Technology

[0002] With the increasing automation of power system operation and maintenance and the growing demand for safe and reliable operation of distribution networks, image-based detection and intelligent recognition technologies for micro-cracks in cable sheaths have attracted widespread attention. Existing cable sheath detection methods largely rely on manual inspection, single optical imaging, or traditional image processing algorithms for crack identification. However, these methods commonly suffer from the following problems in practical applications: Under conditions of varying lighting, complex background interference, and inconsistent shooting angles, the contrast of the cracked areas in the acquired cable sheath surface images is low, and edge features are not obvious. This leads to insufficient response of traditional threshold segmentation or edge detection methods to micro-cracks, resulting in missed detections and false detections. Furthermore, in the image preprocessing and feature extraction stages, existing methods often employ fixed filters or static feature templates, failing to adaptively adjust feature extraction based on different texture intensities and background complexity. The current approach suffers from several drawbacks. Firstly, in situations with high background texture interference, crack region features may be obscured or key information may be lost in low-contrast areas. Secondly, traditional detection frameworks based on manual features and shallow classifiers struggle to simultaneously address crack morphology diversity and pixel-level localization accuracy, especially when cracks are thin, discontinuous, or have weak reflections, resulting in poor stability of detection results. Thirdly, existing deep learning methods are still susceptible to spurious responses in complex backgrounds, lacking multi-scale response consistency analysis and jitter suppression mechanisms. This leads to discrete high-response points in crack heatmaps under high-noise conditions, affecting the reliability and interpretability of the final identification.

[0003] Therefore, how to provide a deep learning-based method for recognizing microcracks in cable sheaths is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based image recognition method for cable sheath microcracks. This invention fully utilizes an improved MAE model, a pixel-level residual fusion mechanism, a background suppression map generation strategy, a multi-scale cumulative heatmap analysis method, and a pseudo-response suppression mechanism. It describes in detail the technical process for achieving high-precision recognition of cable sheath microcracks under complex backgrounds and low-contrast conditions, and has the advantages of high recognition accuracy, strong anti-interference ability, and adaptability to different scene environments.

[0005] A method for identifying microcracks in cable sheaths based on deep learning according to an embodiment of the present invention includes the following steps:

[0006] Step 1: Acquire images of the cable sheath and perform image preprocessing operations, then output the preprocessed image;

[0007] Step 2: Input the preprocessed image into the improved MAE model to generate a background reconstruction map and a crack reconstruction map;

[0008] Step 3: Perform pixel-level residual fusion of the background reconstruction image and the crack reconstruction image to generate a background occlusion image;

[0009] Step 4: Perform pixel-level fusion of the background masking map and the preprocessed image to generate a background suppression map;

[0010] Step 5: Perform microcrack identification based on the background suppression map, and output a preliminary set of crack response heatmaps by extracting image features and identifying crack regions.

[0011] Step 6: Perform heatmap cumulative analysis on the crack response heatmap set to construct a multi-scale cumulative heatmap;

[0012] Step 7: Determine the risk area of ​​false response based on the local response change rate in the multi-scale cumulative heat map; if the local response change rate is greater than the preset jitter threshold, then perform a response value reduction operation based on the corresponding local area to generate a crack heat map after false response suppression.

[0013] Step 8: Extract high-confidence crack regions based on the crack heat map to obtain the microcrack identification results of the cable sheath.

[0014] Preferably, step one specifically comprises:

[0015] A color image of the cable sheath surface is acquired, and the color image is converted into a grayscale image to generate an initial grayscale image;

[0016] Perform contrast-limited adaptive histogram equalization on the initial grayscale image to enhance local image contrast and output a contrast-enhanced image;

[0017] High-pass filtering is performed on the contrast-enhanced image to improve the detail features of the microcrack edges in the image, generating a preprocessed image.

[0018] Preferably, step two specifically includes:

[0019] The preprocessed image is then input into the improved MAE model;

[0020] The improved MAE model includes a structure-guided dynamic mask generator, a structure-aware Transformer encoder, and a dual-channel decoder structure.

[0021] The structure guides the dynamic mask generator to receive the preprocessed image and generate a mask image based on the crack confidence of the image patch and the background texture intensity.

[0022] The structure-aware Transformer encoder receives image patches that are not masked by the mask image and extracts global semantic feature representations of the image;

[0023] The dual-path decoder structure includes a crack reconstruction branch and a background reconstruction branch. The crack reconstruction branch receives semantic feature representations and outputs a crack reconstruction map, while the background reconstruction branch receives semantic feature representations and outputs a background reconstruction map.

[0024] Preferably, the improved MAE model is as follows:

[0025] The improved MAE model consists of a structure-guided dynamic mask generator, a structure-aware Transformer encoder, and a dual-channel decoder structure connected in sequence.

[0026] The structure-guided dynamic mask generator includes: receiving a preprocessed image and extracting local texture features from image patches, outputting a local feature map; receiving the local feature map, calculating a mask probability map based on the crack confidence score and background texture intensity score of each image patch, and setting a probability threshold to adjust the mask probability. Figure 2 Value-enhanced to generate a mask image, which identifies the image block regions in the image to be masked;

[0027] The structure-aware Transformer encoder includes: performing a linear transformation on the unmasked image patches of the mask image to generate an embedded feature vector sequence; a position encoding module to encode the position information of the image patches in the embedded feature vector sequence to generate a feature representation containing spatial position information; performing cross-position attention mechanism calculation on the embedded vector sequence to capture long-distance dependencies between image patches; receiving the feature representation output by the multi-head self-attention module, and connecting depthwise separable convolution, dilated convolution, and positional convolution in parallel in the feedforward path to extract the structural edge information of the image, and fusing the convolution outputs to form a global semantic feature vector sequence;

[0028] The dual-path decoder structure includes: a crack decoding branch, which receives a global semantic feature vector sequence, performs upsampling and deconvolution operations sequentially to restore the crack texture in the unmasked area of ​​the image, and outputs a crack reconstruction map; and a background decoding branch, which receives a global semantic feature vector sequence, performs the same deconvolution structure as the crack decoding branch, reconstructs the background information of the masked area, and outputs a background reconstruction map.

[0029] Preferably, step three specifically includes:

[0030] The difference between the background reconstruction image and the crack reconstruction image is calculated according to the pixel position correspondence to obtain the initial residual image;

[0031] Normalization is performed on the initial residual map to map the pixel values ​​to a set range, thereby generating a background residual map;

[0032] High response suppression and low response enhancement operations are performed on the background residual map. The pixel-level response amplitude is adjusted by setting a response threshold, and a response enhancement map is output.

[0033] A combination of smoothing and edge-preserving filtering is performed on the response enhancement map to remove local noise and abrupt changes in the residual map, generating a background masking map.

[0034] Preferably, step four specifically includes:

[0035] Perform pixel-by-pixel multiplication on the corresponding pixel values ​​in the background masking image and the preprocessed image to generate a pixel-level fused image;

[0036] Pixel value normalization processing is performed on the pixel-level fused image to adjust the overall brightness dynamic range of the image and generate a fused enhanced image;

[0037] The fused and enhanced image is processed with image smoothing and edge sharpening to reduce background noise and improve the contrast of crack edges, generating a background suppression map.

[0038] Preferably, step five specifically includes:

[0039] The background suppression map is input into the microcrack recognition neural network model, which includes a feature extraction and encoding module, a feature fusion module, and a crack region discrimination module.

[0040] The feature extraction and encoding module extracts image texture information under different receptive fields based on a multi-scale convolutional structure, and generates a multi-scale feature map.

[0041] The feature fusion module integrates structural features, texture features, and edge information from multi-scale feature maps and constructs a fused feature representation.

[0042] The crack region discrimination module performs pixel-level classification of possible crack regions based on the fused feature expression and outputs a preliminary crack response heatmap set.

[0043] Preferably, step six specifically includes:

[0044] The crack response heatmaps at different scales in the preliminary crack response heatmap set are superimposed according to their positional alignment to generate a cumulative response image;

[0045] The response cumulative image is subjected to scale weight allocation processing, and a weighted cumulative image is constructed by setting weight coefficients according to the corresponding scale level.

[0046] Normalization is performed on the weighted cumulative image to map the cumulative response value to a set range;

[0047] The normalized weighted cumulative image is output as a multi-scale cumulative heatmap.

[0048] Preferably, step seven specifically comprises:

[0049] Local sliding window traversal is performed on the multi-scale cumulative heatmap to extract the response value sequence of each local region at different scales;

[0050] Variance calculation is performed on the response value sequence to obtain the response change rate index map of the corresponding local region;

[0051] The response change rate of each pixel in the response change rate index graph is compared with a preset jitter threshold;

[0052] If the response change rate is greater than the preset jitter threshold, the corresponding pixel is marked as a pseudo-response risk pixel;

[0053] Local mean suppression and response amplitude compression are performed on the response values ​​of the pseudo-response risk pixel regions to generate a pseudo-response suppression heatmap with the response values ​​shrunk and corrected, and the crack heatmap is output.

[0054] Preferably, step eight specifically includes:

[0055] Perform a full image traversal on the crack heatmap and extract the set of high-response pixels whose response values ​​are greater than a preset threshold.

[0056] Perform connected component analysis on the set of high-response pixels and merge spatially adjacent pixels to form candidate crack regions;

[0057] For each candidate crack region, calculate the region area, aspect ratio, and mean response, and match and filter them with preset crack feature rules to eliminate false detection regions that do not conform to the geometric and response rules;

[0058] Output the filtered crack region mask image and crack boundary box set.

[0059] The beneficial effects of this invention are:

[0060] This invention addresses the problems of low contrast, background texture interference, and the easy obscuring of weak crack features in cable sheath microcrack images at the acquisition end by synergistically integrating an improved MAE model and a multi-stage response suppression mechanism. It employs a preprocessing strategy combining contrast-limited adaptive histogram equalization and high-pass filtering, constructing a dynamic mask image by combining crack confidence and background texture intensity. Furthermore, it introduces a structural feature extraction path that fuses depthwise separable convolution, dilated convolution, and positional convolution in the structure-aware Transformer encoder, achieving differentiated encoding of background and crack region features. In the image reconstruction stage, a dual-path decoder structure is constructed to generate background reconstruction and crack reconstruction images separately. The image shows a high-separation background masking map generated through pixel-level residual fusion, adaptive response amplitude adjustment, and noise smoothing filtering. This map is then fused with the preprocessed image at the pixel level to obtain a background suppression map. In the microcrack identification stage, structural features, texture features, and edge features are fused based on multi-scale convolution and attention mechanisms to output a high-confidence crack response heatmap set. Multi-scale cumulative heatmaps are constructed through scale-weighted superposition and normalization. Combined with local response change rate judgment and shrinkage correction to suppress pseudo-response regions, and finally, through connected component analysis and crack feature rule screening, a high-confidence crack region with continuous contour and sufficient background suppression is output, achieving high-precision detection and robust identification of cable sheath microcracks in complex backgrounds. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0062] Figure 1 This is a flowchart of a deep learning-based image recognition method for cable sheath microcracks proposed in this invention.

[0063] Figure 2 This is a structural diagram of the improved MAE model proposed in this invention;

[0064] Figure 3 This is a data flow diagram of a deep learning-based image recognition method for cable sheath microcracks proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1-3 A deep learning-based image recognition method for microcracks in cable sheaths includes the following steps:

[0067] Step 1: Acquire image data of the cable sheath surface and perform image preprocessing operations to obtain a preprocessed image. Image preprocessing operations include image grayscale conversion, local contrast enhancement, and high-frequency edge detail enhancement;

[0068] Step 2: Input the preprocessed image into the improved MAE model to generate a background reconstruction map and a crack reconstruction map; separate the background and crack information through a structure-guided dynamic masking mechanism to enhance the structural consistency of the reconstruction expression;

[0069] Step 3: Perform pixel-level residual fusion on the background reconstruction image and the crack reconstruction image to generate a background occlusion image. Pixel-level residual fusion includes calculating the difference between corresponding pixel values, performing normalization and response adjustment operations to enhance the features of the microcrack region;

[0070] Step 4: The background masking map and the preprocessed image are fused pixel by pixel to generate a background suppression map; the fusion process adopts a pixel-by-pixel multiplication enhancement strategy, and combines image smoothing and edge sharpening mechanisms to reduce background interference and highlight crack structures;

[0071] Step 5: Perform microcrack identification based on the background suppression map, and output a preliminary set of crack response heatmaps by extracting image features and identifying crack regions.

[0072] Step 6: Perform heatmap cumulative analysis on the crack response heatmap set to construct a multi-scale cumulative heatmap; the heatmap cumulative analysis adopts a scale-weighted fusion and response normalization strategy to enhance the response consistency of continuous crack regions and suppress isolated response noise;

[0073] Step 7: Determine the risk area of ​​false response based on the local response change rate in the multi-scale cumulative heat map; if the local response change rate is greater than the preset jitter threshold, then perform a response value reduction operation based on the corresponding local area to generate a crack heat map after false response suppression.

[0074] Step 8: Extract high-confidence crack regions based on the crack heat map to obtain the microcrack identification results of the cable sheath.

[0075] This implementation achieves significant separation between background texture and microcrack signals in cable sheath images through an improved MAE model, avoiding misidentification caused by background interference. Pixel-level residual fusion and background suppression strategies effectively enhance the response of microcracks in the image, improving the identifiability of minor defects. Multi-scale heatmap cumulative analysis and local response fluctuation suppression mechanisms significantly reduce the false detection rate caused by spurious responses, enhancing the spatial consistency and structural integrity of crack identification results. The resulting high-confidence crack identification results possess higher accuracy and robustness, making them suitable for cable sheath defect detection tasks in complex backgrounds.

[0076] In this embodiment, step one specifically includes:

[0077] Color images of the cable sheath surface are acquired using an industrial camera with linear gain adjustment. The acquisition process is conducted under a stable light source to suppress interference in highlight areas and avoid overexposure.

[0078] The acquired color image is converted into a grayscale image using a weighted average method. Grayscale values ​​are generated by weighting and fusing the red, green, and blue channels. The weighting coefficients are obtained by fitting a typical cable sample dataset to generate an initial grayscale image. The initial grayscale image retains the main brightness information in the color image and removes color interference.

[0079] The initial grayscale image is subjected to contrast-limited adaptive histogram equalization. Histogram equalization is performed on each local image block. The cropping limit value is obtained by fitting the grayscale distribution histogram of the crack region, and the contrast-enhanced image is output.

[0080] The contrast-enhanced image is subjected to high-pass filtering and combined with an edge response threshold adjustment strategy. The filtered output image is then cropped by a response threshold, which is obtained by fitting a large amount of crack sample data to generate a preprocessed image. This process enhances the fine-grained edge feature information in the image and improves the local response significance of the microcrack contour.

[0081] In this embodiment, step two specifically includes:

[0082] The preprocessed image is divided into several image blocks according to a fixed resolution to obtain an image block set; the image block set is then input into the improved MAE model.

[0083] The improved MAE model consists of a structure-guided dynamic mask generator, a structure-aware Transformer encoder, and a dual-channel decoder connected in sequence. The overall parameters of the model are obtained through training and optimization using a microcrack image recognition task.

[0084] The structure-guided dynamic mask generator includes a shallow convolution module and a mask probability scoring module. The shallow convolution module uses ReLU as the activation function to extract local texture response features of image blocks. The mask probability scoring module scores mask weights based on the intensity of shallow texture response and the crack confidence of the corresponding pixels in the preprocessed image. The mask weights are obtained by fitting through the difference analysis between cracked and non-cracked regions and output a mask image. The higher the value in the mask image, the easier it is for the image block to be masked.

[0085] The structure-aware Transformer encoder includes an image patch embedding module, a position encoding module, a multi-head self-attention module, and a structure-aware feedforward network module. The image patch embedding module uses a linear transformation to map unmasked image patches into a sequence of embedding vectors. The position encoding module uses two-dimensional sinusoidal encoding to represent the spatial position information of image patches in the original image. The multi-head self-attention module extracts context-related features at different scales, and the number of attention heads is obtained through experimental fitting. The structure-aware feedforward network module combines depthwise separable convolution, dilated convolution, and positional convolution in a concatenated manner to introduce spatial local structure perception capability.

[0086] The dual-path decoder structure includes a crack reconstruction branch and a background reconstruction branch. The crack reconstruction branch receives the semantic feature vector output by the structure-aware Transformer encoder. The decoding network structure consists of four deconvolutional layers, each containing BatchNorm and ReLU activation functions, and outputs a crack reconstruction map. The background reconstruction branch structure is symmetrically set with the crack reconstruction branch and outputs a background reconstruction map. The two branches respectively reconstruct the significant crack region and the continuous background region in the preprocessed image.

[0087] Guided by deep mask learning and structure-aware representation, this implementation effectively separates crack structures from background texture features in images, providing clear and fully represented multi-channel image inputs for subsequent residual fusion and response extraction, thereby improving the model's ability to identify microcrack information under complex background interference.

[0088] In this embodiment, the improved MAE model is specifically as follows:

[0089] The improved MAE model consists of a structure-guided dynamic mask generator, a structure-aware Transformer encoder, and a dual-channel decoder connected in sequence; the input is a preprocessed image, which is divided into image blocks, and the specific proportions are obtained by fitting a validation set.

[0090] The structure-guided dynamic mask generator includes a shallow convolution module and a mask probability scoring module. The shallow convolution module consists of convolution, batch normalization, and ReLU, outputting a local feature map. The mask probability scoring module generates a crack confidence map and a background texture intensity map on the local feature map, respectively. The weight coefficients of the two are obtained by fitting the training data and fused to obtain the mask probability map. The mask probability map is binarized to generate a mask map through a probability threshold τ, which is obtained by fitting the validation set performance. The mask map corrects isolated mask blocks and fracture boundaries through opening and closing operations, and the overall mask ratio is limited to the target range. The limiting parameters are obtained by fitting through grid search.

[0091] The structure-aware Transformer encoder comprises an image patch embedding module, a position encoding module, a multi-head self-attention module, and a structure-aware feedforward network module. The image patch embedding module linearly maps image patches not obscured by the mask image into a sequence of embedded feature vectors of dimension d, where d is obtained by fitting the model capacity and memory constraints. The position encoding module injects spatial position information into the embedded feature vector sequence using two-dimensional sinusoidal position encoding. The number of heads h and layers L in the multi-head self-attention module are obtained by jointly fitting the training and validation sets, and outputs context-sensitive feature representations. The structure-aware feedforward network module connects depthwise separable convolutions, dilated convolutions, and positional convolutions in parallel in the feedforward path, and performs weighted fusion of the parallel branches through a gated fusion unit. The fusion weights are obtained by fitting the model during training. Residual connections and layer normalization are set at both ends of the module, and the output is a sequence of global semantic feature vectors.

[0092] The dual-path decoder structure includes a crack decoding branch and a background decoding branch. The global semantic feature vector sequence is rearranged into feature maps through shape restoration mapping and then input into the two branches respectively. The crack decoding branch consists of four layers of transposed convolution and upsampling concatenation, with each layer followed by batch normalization and ReLU, and the number of channels decreases step by step, outputting a crack reconstruction map. The background decoding branch structure is symmetrically set with the crack decoding branch, outputting a background reconstruction map. The upsampling ratio, transposed convolution stride, and padding parameters of the two branches are obtained by fitting a joint index of reconstruction accuracy and reconstruction smoothness.

[0093] The training configuration consists of two stages: pre-training and fine-tuning. The pre-training stage aims to maximize background reconstruction accuracy and structure preservation, with the background branch loss weights and crack branch loss weights obtained through grid search fitting. The fine-tuning stage is performed on the cable sheath dataset, with mask ratio scheduling progressively from high to medium masks, and the scheduling curve obtained through validation set performance fitting. The optimizer is AdamW, and the base learning rate, weight decay, and number of training epochs are obtained through learning rate search and early stopping strategy fitting. Data augmentation includes brightness perturbation, contrast perturbation, and affine micro-rotation, with the augmentation range obtained through training stability and validation set metrics fitting.

[0094] The inputs for the inference stage are a preprocessed image and a mask image obtained by a structure-guided dynamic mask generator; the structure-aware Transformer encoder outputs a global semantic feature vector sequence; and the dual-channel decoder structure outputs a background reconstruction image and a crack reconstruction image, respectively.

[0095] This implementation method, through an improved MAE model construction method, obtains structurally separated outputs of background reconstruction map and crack reconstruction map under complex background conditions, improves the separability of background texture and microcrack signal, reduces the probability of pseudo response caused by background texture, enhances the visibility and continuity of microcrack details, and improves the detection rate and stability of subsequent pixel-level residual fusion and microcrack identification.

[0096] In this embodiment, step three specifically includes:

[0097] The background reconstruction image and the crack reconstruction image are compared by pixel-by-pixel subtraction to generate an initial residual image. Positive values ​​in the initial residual image indicate that the pixel response of the crack reconstruction image is higher than that of the background reconstruction image, negative values ​​indicate that the pixel response of the background reconstruction image is higher than that of the crack reconstruction image, and zero values ​​indicate that the responses of the two images are the same.

[0098] Normalization is performed on the initial residual map, and the pixel values ​​are linearly mapped to the interval between zero and one according to the minimum and maximum values. The upper and lower limits of the normalization interval are obtained by fitting the pixel value distribution of the reconstructed image in the crack sample set, and the background residual map is output. In the background residual map, the closer the value of a pixel is to one, the more significant the crack response is, and the closer the value of a pixel is to zero, the more significant the background response is.

[0099] High response suppression and low response enhancement operations are performed on the background residual map. High response suppression compresses the amplitude of pixel values ​​exceeding the threshold by setting a high threshold, while low response enhancement amplifies the amplitude of pixel values ​​below the threshold by setting a low threshold. Both the high and low thresholds are obtained by fitting a joint index of crack detection rate and false detection rate, and the response enhancement map is output. The response enhancement map reduces the interference of high response regions in the background while retaining the effective crack response.

[0100] A combination of smoothing and edge-preserving filtering is performed on the response enhancement image. The smoothing kernel type is Gaussian kernel, and the standard deviation is obtained by fitting the signal-to-noise ratio of the filtered image. The edge-preserving filter adopts a bilateral filtering structure. The filtering parameters in the spatial domain and pixel value domain are obtained by fitting the comprehensive score of crack edge sharpness and background smoothness to generate a background masking image.

[0101] This implementation significantly suppresses background interference areas visually by using a background masking map while preserving local detail features of the crack area. This provides high-quality pixel-level feature input for subsequent background suppression fusion and microcrack identification, improving the signal-to-noise ratio and segmentability of microcrack features.

[0102] In this embodiment, step four specifically includes:

[0103] The background masking image and the preprocessed image are multiplied point-by-point at pixel positions. The result of the multiplication operation retains the high-response pixels in the crack region and attenuates the pixel values ​​in the background region, outputting a pixel-level fused image. The pixel value distribution range in the pixel-level fused image is directly related to the weight distribution of the background masking image. The weight distribution is obtained by fitting the response separation degree between the background region and the crack region.

[0104] Pixel value normalization is performed on the pixel-level fused image. The pixel values ​​are linearly mapped to the global minimum and maximum values, with a mapping range of zero to one. The upper and lower limits of normalization are obtained by fitting the brightness histogram distribution of the fused image on the training set, and a fused enhanced image is generated. The fused enhanced image maintains the brightness advantage of the crack area and reduces the brightness fluctuation of the background area within the global brightness dynamic range.

[0105] The fused and enhanced image is subjected to a combination of image smoothing and edge sharpening. The image smoothing uses a Gaussian filter structure, and the standard deviation of the filter kernel is obtained by fitting the structural similarity index of the fused and enhanced image with the joint index of noise level. The edge sharpening uses an unsharpened mask method, and the sharpening intensity coefficient and radius parameter are obtained by fitting the comprehensive score of crack edge sharpness and background smoothness. The smoothing and sharpening are performed in the order of smoothing first and then sharpening, generating a background suppression map.

[0106] This implementation method enhances the contrast and continuity of crack edges by reducing stray noise in the background region through background suppression maps, providing input images with high signal-to-noise ratio and significant edge features for subsequent microcrack identification, thereby improving the accuracy and stability of crack detection.

[0107] In this embodiment, step five specifically includes:

[0108] The background suppression map is input into the DCNN model. Before the input image is fed into the model, the size is normalized and the pixel value is standardized. The target resolution of the size normalization is obtained by fitting the balance between detection accuracy and computational cost. The pixel value standardization range is set to zero to one. The standardization parameters are obtained by fitting the statistical distribution of the training set.

[0109] The feature extraction and encoding module extracts image texture information under different receptive fields based on a multi-scale convolutional structure. The combination of convolutional kernel sizes includes a set of small receptive field convolutional kernels, a set of medium receptive field convolutional kernels, and a set of large receptive field convolutional kernels. The outputs of the three sets of convolutional kernels are concatenated in the channel dimension to form a multi-scale feature map. The convolutional kernel size configuration and channel number allocation ratio are obtained by fitting a joint optimization objective of crack detection rate and false detection rate. The multi-scale feature map maintains the consistency of the fusion between low-level texture information and high-level semantic information through residual connections in the downsampling path.

[0110] The feature fusion module receives multiple outputs from multi-scale feature maps. First, it uses a channel attention mechanism to weight and adjust different feature channels. The attention weights are obtained by fitting the response intensity distribution of the multi-scale features to the crack region. Then, a spatial attention mechanism is executed on the channel-weighted feature map to capture the spatial aggregation features of the crack region and the suppression features of the background region. The fusion result is then subjected to parallel convolution operations of structural feature branches, texture feature branches, and edge feature branches to extract the corresponding structural feature maps, texture feature maps, and edge feature maps. These are then weighted and fused in the feature fusion unit. The fusion weights are obtained by fitting the joint loss function during model training to generate a fused feature representation.

[0111] The crack region discrimination module receives the fused feature expression and uses a pixel-by-pixel classification method to distinguish cracked pixels from non-cracked pixels. The classifier is a pixel-level Softmax structure with convolutional layers and fully connected layers. The classification threshold is obtained by ROC curve analysis and fitting. The output classification result represents the confidence of crack existence in the form of a probability heatmap, forming a preliminary crack response heatmap set.

[0112] This implementation method uses a preliminary crack response heatmap set to show that the crack region has a high response and continuous profile, while the background region has a low response and stable suppression characteristics. This provides high-quality response input for subsequent multi-scale cumulative analysis, significantly improving the recall and accuracy of microcrack detection.

[0113] In this embodiment, step six specifically includes:

[0114] The crack response heatmaps at different scales in the preliminary crack response heatmap set are aligned and superimposed according to pixel position. During the alignment process, subpixel interpolation is used to correct the edge offset caused by scale changes. The interpolation kernel parameters are obtained by fitting the crack contour alignment accuracy. The heatmaps at each scale are stacked by channel to form a scale response tensor.

[0115] The scale response tensor is subjected to scale weighting processing. The weighting coefficients are set according to the crack detection rate and false detection rate of each scale on the training set. The weighting coefficients are obtained by fitting a multi-objective optimization function. The weighted response maps of each scale are summed in the channel dimension to generate a weighted cumulative image.

[0116] The weighted cumulative image is normalized to map the cumulative response value to a set range of zero to one. The normalization mapping function parameters are obtained by fitting the crack response value distribution in the training set. During the normalization process, response suppression is performed on isolated high-response pixels and response enhancement is performed on continuous low-response crack regions. The suppression coefficient and enhancement coefficient are obtained by fitting the optimal F1 score.

[0117] The output is a normalized and response-adjusted weighted cumulative image as a multi-scale cumulative heatmap.

[0118] This implementation method can maintain the continuity and consistency of the crack region response at different spatial resolutions by using multi-scale cumulative heatmaps, while effectively reducing the number of isolated pseudo-response points caused by single-scale detection, thereby improving the accuracy and stability of subsequent pseudo-response suppression and high-confidence crack extraction steps.

[0119] In this embodiment, step seven specifically includes:

[0120] The multi-scale cumulative heatmap is traversed in the spatial domain using a local sliding window of fixed size. The sliding step size and window size are obtained by fitting the joint optimization objective of crack continuity preservation rate and pseudo response suppression rate. Within each sliding window, the response values ​​of the corresponding positions of the window region at all scales are extracted in scale order to construct a response value sequence matrix.

[0121] Variance calculation is performed on the response value sequence matrix in the scale dimension to generate a local response change rate index map. The variance calculation method is obtained by statistical modeling and fitting to adapt the crack response to the change characteristics at multiple scales.

[0122] The response change rate of each pixel in the local response change rate index map is compared pixel by pixel with a preset jitter threshold. The jitter threshold is obtained by fitting an objective function that minimizes the weighted sum of the false detection rate and the false negative rate on the training set. If the response change rate of a pixel is greater than the preset jitter threshold, the pixel is marked as a pseudo response risk pixel in the labeling matrix.

[0123] Within the pixel regions marked as having a risk of spurious responses, local mean suppression is performed on the response values. The mean kernel size is obtained by fitting the balance between response smoothness and detail preservation. Based on the mean suppression results, response amplitude compression is further performed, with the compression ratio obtained by maximizing the F1 score. The suppressed and compressed pixel values ​​are then updated in the heatmap to generate a spurious response suppression heatmap that has been shrunk back.

[0124] The pseudo-response suppression heatmap, after being shrunk and corrected, is output as the crack heatmap.

[0125] In this embodiment, the pixel value of the crack heatmap in the pseudo-response region is significantly reduced, while maintaining the continuity and response intensity of the real crack region, thereby improving the accuracy and stability of subsequent high-confidence crack extraction.

[0126] In this embodiment, step eight specifically includes:

[0127] A full image traversal is performed on the crack heatmap, and the response value is compared pixel by pixel with a preset confidence threshold. The confidence threshold is obtained by fitting an objective function that maximizes the weighted sum of detection rate and precision on the training set. Pixels with response values ​​greater than the confidence threshold are marked as high-response pixels, and a set of high-response pixels is constructed.

[0128] Connectivity analysis is performed on the set of high-response pixels. An eight-neighbor pixel connection rule is used in the connectivity determination. The connection rule parameters are obtained by joint optimization fitting of the connectivity rate of the real crack region and the separation rate of background noise. The sets of pixels that are determined to be adjacent are merged into candidate crack regions.

[0129] For each candidate crack region, the region area, aspect ratio, and mean response are calculated. The threshold ranges for region area and aspect ratio are obtained by statistical fitting of crack morphological features, and the threshold range for mean response is obtained by fitting the distribution of crack and non-crack responses. The calculation results are matched and filtered with preset crack feature rules to remove candidate regions that do not conform to the geometric and response rules.

[0130] The selected and retained candidate crack regions are converted into crack region mask images, and their bounding rectangles are extracted to form a set of crack boundary boxes.

[0131] The crack region mask image output by this embodiment has high confidence and spatial continuity. The crack boundary box set can accurately reflect the location and shape characteristics of the crack, thereby improving the visualization effect and detection reliability of microcrack identification in cable sheaths.

[0132] Example 1:

[0133] To verify the feasibility of this invention in practice, it was applied to the distribution operation and maintenance center of a power grid company, covering three scenarios: suburban interconnection cable channels, cable interlayers within stations, and outgoing lines of ring main units. The on-site imaging terminal was fixed on two types of carriers: an inspection robot and a portable gimbal. The acquired images underwent basic format verification and secure encrypted transmission via an edge computing gateway, and were processed at the central inference node according to the process described in claims 1 to 10. The scenarios presented interference factors such as highly reflective sheaths, oil and dust accumulation, complex background textures, and overlapping cables. Microcracks generally exhibited characteristics such as being thin, discontinuous, and having low grayscale contrast. Traditional methods suffer from high false negative rates and heavy manual review burdens. This embodiment focuses on verifying the ability to detect microcracks, suppress false positives, and improve the efficiency of the operation and maintenance process under complex backgrounds and low contrast conditions.

[0134] In this scenario, after the image enters the process, a preprocessed image is generated according to step one. Then, according to step two, the improved MAE model is used to construct a background reconstruction map and a crack reconstruction map. In step three, pixel-level residual fusion is used to obtain a background masking map, which, together with the preprocessed image, generates a background suppression map in step four. The background suppression map completes feature extraction, feature fusion, and crack region discrimination in the DCNN model, outputting a preliminary set of crack response heatmaps. Subsequently, according to step six, scale alignment, weight allocation, and normalization are performed to obtain a multi-scale cumulative heatmap. In step seven, a shrinkage correction is performed based on the local response change rate, and a pseudo-response suppression heatmap is output as the crack heatmap. In step eight, connected component aggregation and rule filtering are completed, and finally, a crack region mask map and a set of crack bounding boxes are output and synchronized to the defect work order module of the operation and maintenance platform for review and defect assignment by the scheduler. The entire process runs in a way that combines edge-side batch caching and center-side high-concurrency inference, balancing latency and throughput.

[0135] During the three-month pilot program, image samples were collected from three types of tasks: routine inspections, specialized encrypted inspections, and fault debriefings. These samples were acquired across multiple weather conditions, time periods, media types, and angles. The operations and maintenance team conducted sampling verification and closed-loop confirmation of the model output on the platform according to unified annotation standards, and the verification results were included in monthly evaluations. This embodiment statistically analyzes indicators such as microcrack detection rate, false detection rate, F1 score, average processing latency, manual review time, number of triggered warnings, and number of finally confirmed defects. Combined with comparisons with three baseline methods, the beneficial effects of this invention are demonstrated.

[0136] Table 1. Operational data of the cable sheath microcrack identification method during a three-month pilot program.

[0137] ;

[0138] As shown in Table 1, the pilot sample size in the first month was 12,000 images. The microcrack detection rate reached 90.2%, the false positive rate was 9.8%, the F1 score was 89.5%, the average processing time was 0.80 seconds per image, the manual review time was 164 hours, 226 warnings were triggered, 181 valid microcracks were confirmed, and 8 unplanned power outage work orders were reduced. In the second month, the sample size increased to 13,400 images, the microcrack detection rate improved to 91.7%, the false positive rate decreased to 8.6%, the F1 score improved to 90.7%, and the average processing time slightly decreased. The processing time was reduced to 0.79 seconds per image, manual review time decreased to 148 hours, 238 warnings were triggered, 194 valid microcracks were confirmed, and 9 unplanned power outage work orders were reduced. In the third month, the sample size continued to increase to 14,100 images, the microcrack detection rate further improved to 93.1%, the false positive rate decreased to 7.9%, the F1 score reached 92.1%, the average processing time decreased to 0.77 seconds per image, manual review time decreased to 130 hours, 245 warnings were triggered, 203 valid microcracks were confirmed, and 10 unplanned power outage work orders were reduced. The data shows that the detection rate of the method of this invention showed a continuous upward trend within three months, the false positive rate decreased month by month, the workload of manual review was significantly reduced, and the number of unplanned power outage work orders was reduced, proving the efficiency and reliability of the method in actual cable operation and maintenance scenarios.

[0139] Table 2. Method Comparison Evaluation and False Alarm Structure Analysis

[0140] ;

[0141] Table 2 compares the methods of threshold + edge detection, U-Net segmentation (excluding background occlusion), FPN-ResNet detection, and the method of this invention. It is evident that the method of this invention achieves a microcrack detection rate of 91.7%–93.1%, a significant improvement compared to 67.9% for threshold + edge detection, 82.6% for U-Net segmentation, and 78.3% for FPN-ResNet. Regarding the false detection rate, the method of this invention remains within the range of 9.8%–7.9%, lower than 26.5% for threshold + edge detection, 17.4% for U-Net segmentation, and 14.8% for FPN-ResNet. The F1 score of the method of this invention reaches 89.5%, slightly higher than the 0.17s / image of threshold + edge detection, but still faster than the 0.98s / image of U-Net segmentation, and within an acceptable range compared to the 0.64s / image of FPN-ResNet detection. In the analysis of false alarm sources, the main false alarms of the method of the present invention come from the extremely fine fiber scratches near the multi-scale cumulative edge. However, this type of false alarm has shown a downward trend, indicating that the multi-scale cumulative heat map and background masking mechanism of the method have a good effect on suppressing noise.

[0142] At the operation and maintenance process level, this method synchronizes the crack area mask image and crack boundary box set to the work order system, automatically filling in fields such as line, location, confidence level, and suggested handling level. Dispatchers can quickly locate the site by opening the preview. Records that pass the review are written back to the training set, forming a continuous optimization loop of small steps and rapid progress. Three months of practice have shown that this method solves the key problems of high missed detections and false alarms in traditional methods under low contrast and complex backgrounds. Relying on background masking encoding and spurious response suppression, it significantly reduces the cost of manual review while maintaining high detection rate, and the actual reduction in unplanned power outage work orders reflects the improvement in equipment reliability and power supply continuity.

[0143] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for recognizing microcracks in cable sheaths based on deep learning, characterized in that, Includes the following steps: Step 1: Acquire images of the cable sheath and perform image preprocessing operations, then output the preprocessed image; Step 2: Input the preprocessed image into the improved MAE model to generate a background reconstruction map and a crack reconstruction map; Step 2 specifically involves: The preprocessed image is then input into the improved MAE model; The improved MAE model includes a structure-guided dynamic mask generator, a structure-aware Transformer encoder, and a dual-channel decoder structure. The structure guides the dynamic mask generator to receive the preprocessed image and generate a mask image based on the crack confidence of the image patch and the background texture intensity. The structure-aware Transformer encoder receives image patches that are not masked by the mask image and extracts global semantic feature representations of the image; The dual-path decoder structure includes a crack reconstruction branch and a background reconstruction branch. The crack reconstruction branch receives semantic feature representation and outputs a crack reconstruction map, while the background reconstruction branch receives semantic feature representation and outputs a background reconstruction map. Step 3: Perform pixel-level residual fusion of the background reconstruction image and the crack reconstruction image to generate a background occlusion image; Step 3 specifically involves: The difference between the background reconstruction image and the crack reconstruction image is calculated according to the pixel position correspondence to obtain the initial residual image; Normalization is performed on the initial residual map to map the pixel values ​​to a set range, thereby generating a background residual map; High response suppression and low response enhancement operations are performed on the background residual map. The pixel-level response amplitude is adjusted by setting a response threshold, and a response enhancement map is output. Perform a combination of smoothing and edge-preserving filtering on the response enhancement map to remove local noise and abrupt changes in the residual map and generate a background masking map; Step 4: Perform pixel-level fusion of the background masking map and the preprocessed image to generate a background suppression map; Step 5: Perform microcrack identification based on the background suppression map, and output a preliminary crack response heatmap set through image feature extraction and crack region discrimination; the preliminary crack response heatmap set includes crack response heatmaps at different scales; Step Six: Perform cumulative heatmap analysis on the crack response heatmap set to construct a multi-scale cumulative heatmap; Step Six specifically involves: The crack response heatmaps at different scales in the preliminary crack response heatmap set are superimposed according to their positional alignment to generate a cumulative response image; The response cumulative image is subjected to scale weight allocation processing, and a weighted cumulative image is constructed by setting weight coefficients according to the corresponding scale level. Normalization is performed on the weighted cumulative image to map the cumulative response value to a set range; The normalized weighted cumulative image is output as a multi-scale cumulative heatmap; Step 7: Determine the risk area of ​​false response based on the local response change rate in the multi-scale cumulative heat map; if the local response change rate is greater than the preset jitter threshold, then perform a response value reduction operation based on the corresponding local area to generate a crack heat map after false response suppression. Step 8: Extract high-confidence crack regions based on the crack heat map to obtain the microcrack identification results of the cable sheath.

2. The method for recognizing microcracks in cable sheaths based on deep learning according to claim 1, characterized in that, Step one specifically involves: A color image of the cable sheath surface is acquired, and the color image is converted into a grayscale image to generate an initial grayscale image; Perform contrast-limited adaptive histogram equalization on the initial grayscale image to enhance local image contrast and output a contrast-enhanced image; High-pass filtering is performed on the contrast-enhanced image to improve the detail features of the microcrack edges in the image, generating a preprocessed image.

3. The method for recognizing microcracks in cable sheaths based on deep learning according to claim 2, characterized in that, The improved MAE model is specifically as follows: The improved MAE model consists of a structure-guided dynamic mask generator, a structure-aware Transformer encoder, and a dual-channel decoder structure connected in sequence. The structure-guided dynamic mask generator includes: receiving a preprocessed image and extracting local texture features of image blocks, and outputting a local feature map; receiving the local feature map, calculating a mask probability map based on the crack confidence score and background texture intensity score of each image block, binarizing the mask probability map by setting a probability threshold, and generating a mask map, wherein the mask map identifies the image block region to be masked in the image; The structure-aware Transformer encoder includes: performing a linear transformation on the unmasked image patches of the mask image to generate an embedded feature vector sequence; a position encoding module to encode the position information of the image patches in the embedded feature vector sequence to generate a feature representation containing spatial position information; performing cross-position attention mechanism calculation on the embedded vector sequence to capture long-distance dependencies between image patches; receiving the feature representation output by the multi-head self-attention module, and connecting depthwise separable convolution, dilated convolution, and positional convolution in parallel in the feedforward path to extract the structural edge information of the image, and fusing the convolution outputs to form a global semantic feature vector sequence; The dual-path decoder structure includes: a crack decoding branch, which receives a global semantic feature vector sequence, performs upsampling and deconvolution operations sequentially to restore the crack texture in the unmasked area of ​​the image, and outputs a crack reconstruction map; and a background decoding branch, which receives a global semantic feature vector sequence, performs the same deconvolution structure as the crack decoding branch, reconstructs the background information of the masked area, and outputs a background reconstruction map.

4. The method for recognizing microcracks in cable sheaths based on deep learning according to claim 3, characterized in that, Step four specifically involves: Perform pixel-by-pixel multiplication on the corresponding pixel values ​​in the background masking image and the preprocessed image to generate a pixel-level fused image; Pixel value normalization processing is performed on the pixel-level fused image to adjust the overall brightness dynamic range of the image and generate a fused enhanced image; The fused and enhanced image is processed with image smoothing and edge sharpening to reduce background noise and improve the contrast of crack edges, thereby generating a background suppression map.

5. The method for recognizing microcracks in cable sheaths based on deep learning according to claim 4, characterized in that, Step five specifically involves: The background suppression map is input into the microcrack recognition neural network model, which includes a feature extraction and encoding module, a feature fusion module, and a crack region discrimination module. The feature extraction and encoding module extracts image texture information under different receptive fields based on a multi-scale convolutional structure, and generates a multi-scale feature map. The feature fusion module integrates structural features, texture features, and edge information from multi-scale feature maps and constructs a fused feature representation. The crack region discrimination module performs pixel-level classification of possible crack regions based on the fused feature expression and outputs a preliminary crack response heatmap set.

6. The method for recognizing microcracks in cable sheaths based on deep learning according to claim 5, characterized in that, Step seven specifically involves: Local sliding window traversal is performed on the multi-scale cumulative heatmap to extract the response value sequence of each local region at different scales; Variance calculation is performed on the response value sequence to obtain the response change rate index map of the corresponding local region; The response change rate of each pixel in the response change rate index graph is compared with a preset jitter threshold; If the response change rate is greater than the preset jitter threshold, the corresponding pixel is marked as a pseudo-response risk pixel; Local mean suppression and response amplitude compression are performed on the response values ​​of the pseudo-response risk pixel regions to generate a pseudo-response suppression heatmap with the response values ​​shrunk and corrected, and the crack heatmap is output.

7. The method for recognizing microcracks in cable sheaths based on deep learning according to claim 6, characterized in that, Step eight specifically involves: Perform a full image traversal on the crack heatmap and extract the set of high-response pixels whose response values ​​are greater than a preset threshold. Perform connected component analysis on the set of high-response pixels and merge spatially adjacent pixels to form candidate crack regions; For each candidate crack region, calculate the region area, aspect ratio, and mean response, and match and filter them with preset crack feature rules to eliminate false detection regions that do not conform to the geometric and response rules; Output the filtered crack region mask image and crack boundary box set.

Images