Segmented cable damage detection method based on Retinex theory

Abstract

The invention provides a segmented cable damage detection method based on a Retinex theory. The segmented cable damage detection method comprises the following steps: S1, inputting a low-illumination cable image; s2, preprocessing the image data; s3, using a Retinex module to enhance the brightness feature and the contrast feature of the image; s4, extracting high-frequency features of the image by adopting a variational mode decomposition (VMD) module; s5, performing high-frequency modal fusion to generate an enhanced image with richer details; s6, adopting a void space pyramid pooling ASPP module to perform void convolution processing on the brightness features, the contrast features and the high-frequency features, and fusing the features on the basis of multi-scale feature extraction; and S7, a deep learning segmentation network of a SegNet architecture is adopted, and the segmentation effect is optimized through a coding-decoding structure of the deep learning segmentation network. According to the method, the problems of insufficient image brightness and low contrast in cable damage inspection under a low-light condition are solved, and details of cracks and damage can be accurately identified; the limitation of a traditional detection method is overcome, and the situations of false detection and missing detection are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a segmented cable damage detection method based on Retinex theory. Background Art

[0002] In all kinds of electrical equipment and machines in modern society, the transmission of electrical energy, information and control signals usually relies on polymer insulated coaxial cables. As a key component of power and communication networks, cables play a vital role in the normal operation of society. However, cables may develop minor defects during long-term use. If these defects are not discovered and handled in time, they may gradually evolve into serious problems and eventually lead to cable failure. Cable failure not only causes equipment downtime and production interruption, but also may cause safety accidents such as electrical fires. Therefore, early and accurate detection of minor defects in cables and monitoring the development trend of defects are crucial for the safe operation and maintenance of cables. During the manufacture, installation and operation of cables, small defects may occur due to mechanical stress, environmental factors (such as moisture, temperature changes) or electromagnetic interference. These defects may initially be only tiny cracks on the surface of the material, slight wear of the insulation layer or local deformation of the conductor. Over time, minor defects may gradually expand under the action of continuous power load and environmental stress. For example, tiny cracks may expand under mechanical vibration or thermal expansion effects, and damage to the insulation layer may cause moisture to penetrate, thereby accelerating material aging and further electrical insulation failure. When the defects evolve to a certain extent, the insulation performance of the cable will significantly decrease, which may cause short circuit, leakage and even cable burning. These faults will not only cause power system interruption, but also cause serious consequences such as equipment damage and casualties. Through sensitive defect detection methods, problems can be found at the stage of minor defects, so as to carry out preventive maintenance and avoid further development of faults. Preventive maintenance can not only extend the service life of the cable, but also reduce unexpected downtime and maintenance costs. Continuous monitoring of the development trend of cable defects can provide an effective means of evaluating the health status of the cable. Based on the trend of defect expansion, maintenance plans can be formulated in advance, resource allocation can be optimized, and the stability and reliability of the power system can be improved. Early diagnosis of cable defects plays an important role in preventing electrical fires and other safety accidents, and can effectively protect the safety of equipment, property and personnel. Therefore, early diagnosis and trend monitoring of cable damage are crucial to ensuring the safe operation of the power system. The development of sensitive and reliable cable defect detection technology, especially methods that can identify defects of different severity, is of great significance to the stable operation and safety of modern society.

[0003] At home and abroad, research on cable defect detection has made some progress. Li Cheng et al. (2020) used the penetration characteristics of X-rays to conduct research on internal cable defect detection. They scanned the cable with X-rays and displayed the collected images on the screen to detect potential defects inside the cable. Liu Rong (2015) studied a partial discharge detection system based on ultrasound for the identification of insulation aging defects inside cables. The system uses the ultrasonic signal generated by partial discharge of the cable, captures the signal through an ultrasonic sensor and extracts the discharge characteristics, and then classifies the defects and locates the faults.

[0004] In terms of cable surface defect detection, Zhang Jun (2017) used a complementary metal oxide semiconductor (CMOS) industrial camera for image acquisition, combined with a fixed threshold image segmentation method, and analyzed the grayscale mean vertical projection curve of the image and the variance mean of the quadratic fitting curve to identify surface defects. Qiao Xiangyang et al. (2020) designed a detection system based on machine vision technology for surface defects such as scratches and small holes that may appear in the cable production process. The line grayscale mean method was used to extract the cable area, and an adaptive threshold segmentation method was used for defect detection.

[0005] In foreign countries, B Li et al. (2017) developed a machine vision-based detection device to solve the problem of power outages caused by icing on the cable surface in cold weather. The device uses the Canny edge detection operator to process the collected images and uses a support vector machine for classification to detect icing on the cable surface. These studies show that the use of advanced image processing and machine vision technology can significantly improve the efficiency and accuracy of cable defect detection, whether it is internal defect or surface defect detection. Summary of the invention

[0006] The present invention discloses a segmented cable damage detection method based on Retinex theory, which solves the problems of insufficient image brightness and low contrast in cable damage inspection under low light conditions, and can accurately identify cracks and damage details; in addition, it overcomes the limitations of traditional detection methods and effectively reduces the occurrence of false detection and missed detection.

[0007] The specific plan is as follows:

[0008] A method for detecting damage of a split cable based on Retinex theory, characterized by comprising the following steps:

[0009] S1, low light cable image input;

[0010] S2, image data preprocessing;

[0011] S3, using the Retinex module to enhance the brightness and contrast features of the image;

[0012] S4, using variational mode decomposition VMD module to extract high-frequency features of the image;

[0013] S5, high-frequency modality fusion, generating enhanced images with richer details;

[0014] S6, using the atrous spatial pyramid pooling ASPP module, performs atrous convolution processing on brightness features, contrast features and high-frequency features, and fuses these features based on multi-scale feature extraction;

[0015] S7, a deep learning segmentation network using the SegNet architecture, optimizes the segmentation effect through its encoding-decoding structure.

[0016] In step S1, the input low-light cable images are taken using an iPad Pro to provide sufficient resolution and clarity to truly reflect the actual situation of the cable. These images contain the original images of the cable, which may have low contrast and high noise because they are taken under low-light conditions. In the monitoring of thermal power plants, hyperspectral imaging can accurately identify different materials and states, such as exhaust from cooling towers and other industrial emissions. These emissions often have unique spectral features, such as spectral absorption peaks of specific chemical components.

[0017] The step S2 specifically includes, firstly, adjusting the size of the input image to 400×400 pixels to unify the image size for subsequent processing and analysis; then, denoising the image, because images under low light conditions are usually accompanied by higher noise levels, which will affect the accuracy of detection. Denoising is a key step, which improves image quality by removing random noise in the image and reducing unnecessary interference. In this step, advanced denoising algorithms are used, including non-local means (NLM) denoising and bilateral filters. These algorithms can effectively retain the detailed information of the image while reducing noise, thereby improving the clarity of the image. Through this preprocessing step, it can be ensured that the input image used in the subsequent processing stage is of higher quality, which helps to improve the accuracy of cable damage detection.

[0018] The step S3 specifically includes:

[0019] S3.1 Image Decomposition

[0020] First, the Retinex module decomposes the input preprocessed image. This process is based on the classic Retinex theory and is divided into reflected component and illumination component.

[0021] Reflection component: refers to the part of the image formed by the light reflected by the object itself, which contains the details and texture of the object; the reflection component is independent of lighting and therefore reflects the inherent characteristics of the object, such as the texture and crack details of the cable surface;

[0022] Illumination component: refers to the part of the image brightness change caused by changes in ambient lighting conditions; the illumination component affects the brightness and contrast of the entire image, such as shadows and dark areas caused by insufficient light.

[0023] S3.2 Image Enhancement

[0024] After image decomposition, the Retinex module enhances the image by processing the reflection and illumination components separately.

[0025] S3.3 Image reconstruction

[0026] After enhancing the reflection and illumination components, the Retinex module recombines the two components to generate two different sets of images: a contrast-enhanced image and a brightness-enhanced image.

[0027] Contrast-enhanced images: These detail images are combined with the adjusted illumination image by retaining the enhanced reflectance component to produce an image with sharp edges and details even in low-light conditions. The high contrast of this image helps accurately locate and identify cracks and breaks in the cable.

[0028] Brightness-enhanced images: Ensure cable visibility in the overall image by focusing on the enhanced lighting component. This process produces an evenly-lit image, reducing the occlusion effect in low-light areas and ensuring that the cable surface is clearly visible.

[0029] After being processed by the Retinex module, the two sets of enhanced images finally obtained provide favorable conditions for subsequent damage detection, so that cable damage and cracks can be accurately detected under various lighting conditions.

[0030] The step S3.2 specifically includes the following steps:

[0031] 3.2.1. Reflection component enhancement

[0032] To generate contrast-enhanced images, the Retinex module enhances the reflected component, including enhancing details and edge information to make cracks and damaged areas in the cable more obvious. By adjusting the reflected component, the contrast between different areas in the image is increased, making the detailed features of the cable clearer.

[0033] Local contrast enhancement: Laplace enhancement, a local contrast adjustment technique, is used to enhance small-scale details and edges in the image, which helps to highlight cracks and defects on the cable surface;

[0034] High-pass filtering: The high-pass filter is used to enhance the high-frequency components in the image and further highlight the details; this processing can remove low-frequency information in smooth areas and focus on preserving and enhancing the contrast of crack edges.

[0035] 3.2.2 Lighting Component Adjustment

[0036] To generate a brightness-enhanced image, the Retinex module adjusts the illumination components to make the overall brightness of the image more uniform; thereby reducing the effect of shadows in the image, increasing the brightness of images taken in low-light conditions, and making the entire image more visible.

[0037] Global illumination equalization: Use Adaptive Histogram Equalization technology to smooth the illumination components, thereby reducing the extremely dark and extremely bright areas in the image and making the overall brightness distribution more uniform;

[0038] Local brightness enhancement: By adjusting the local brightness, the shadow area is appropriately enhanced, thereby reducing the visibility problem caused by low light; ensuring that the details of the cable are still visible under different lighting conditions.

[0039] In step S4, the variational mode decomposition (VMD) module decomposes the image into multiple modes, each mode contains information of different frequencies, so as to distinguish high-frequency and low-frequency information in the image; through iterative optimization, the image is decomposed into a number of intrinsic mode functions (IMFs), and these IMFs represent different frequency components in the image respectively; each IMF corresponds to a characteristic mode in the image and contains signal information in a specific frequency range; through mode decomposition, the image is decomposed into multiple IMFs, and each IMF is a new image; some of these IMFs contain lower frequency information, including the overall outline and background of the image; some contain high frequency information, including details and edges in the image; high-frequency modes containing cable damage information are selected from multiple IMFs, and these high-frequency modes contain details of the cable damage area, such as the edges of cracks, small breaks, etc.

[0040] In step S5, the enhanced image more clearly displays the edge and small damaged area of ​​the cable, providing clear visual information, so as to improve the accuracy and reliability of detection in the subsequent damage detection process; a weight is assigned to each high-frequency mode, and these weights are determined based on the frequency characteristics of the mode or by analyzing the mode; the details and edge information in the high-frequency mode are more significant, and these modes are given higher weights; and multiple high-frequency mode images are superimposed using the following formula:

[0041] Enhanced Image=w1×High-Frequency Mode1+w2×High-Frequency Mode2

[0042] Among them, ω1 and ω2 are the weight coefficients of each high-frequency mode, High-Frequency Mode1 and High-FrequencyMode2 are high-frequency modes;

[0043] Through the above-mentioned weighted superposition and fusion strategy, an enhanced high-frequency image is generated. Compared with the original image, this image significantly improves the contrast and clarity of details, especially at the edges and damaged areas of the cable; small cracks and damaged areas become more prominent, making them easier to identify visually.

[0044] In step S6, the atrous spatial pyramid pooling ASPP module performs multi-scale feature extraction for each type of feature through a series of atrous convolution layers (for example, convolutions with expansion rates of 1, 6, and 12); all feature maps processed by atrous convolution are spliced, and then feature compression and fusion are performed through a 1x1 convolution layer to generate the final fused feature map. The feature map fused by the ASPP module contains high-frequency information, brightness information, and contrast information, which are effectively integrated at multiple scales. This fused feature can better identify the damaged area of ​​the cable in a low-light environment, ensure the clarity of edges and details, and thus provide more reliable input for subsequent segmentation networks (such as SegNet).

[0045] In step S7, the encoder part is the same as the first 13 convolutional layers of the VGG16 network, consisting of several convolutional layers, batch normalization layers, and maximum pooling layers; after the maximum pooling layer, the pooling indices are saved for the deconvolution operation of the decoder part;

[0046] The decoder part inverts the encoder layer by layer and uses the maximum pooling index corresponding to the encoder for upsampling; through the deconvolution layer (upsampling), the spatial resolution of the feature map is restored, and the pooling index is used to ensure that the position information of the feature map is preserved; each decoder adds the feature map restored by deconvolution to the feature map of the corresponding encoder module element by element (skip connection) to fuse multi-scale information; the classifier part finally performs classification through a convolution layer and outputs the category probability of each pixel.

[0047] The beneficial effects of the present invention are:

[0048] 1. The present invention effectively enhances the details and edge information of the cable damage area by combining Retinex theory and variational mode decomposition (VMD) theory, thereby significantly improving the accuracy of cable damage detection. Under low light conditions, traditional methods often have difficulty accurately identifying the damage area, while the present invention can capture these details more clearly and reduce false detection and missed detection.

[0049] 2. Using the Atrous Spatial Pyramid Pooling (ASPP) module, the fusion of information of different scales is successfully realized. This fusion strategy can not only effectively capture various detailed features of the cable damage area, but also enhance the global and local perception capabilities of the model, providing technical support for improving detection accuracy.

[0050] 3. Through verification and experiments, the present invention demonstrates the ability to effectively detect cable damage under low light conditions. This technical achievement can not only be applied to cable health monitoring, but can also be extended to other similar industrial application scenarios, such as pipeline crack detection, structural defect monitoring, etc., with broad application prospects and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Flow chart of the present invention, wherein A represents the ASPP module.

[0052] Figure 2 It is a high-frequency information fusion diagram.

[0053] Figure 3 Prediction diagram for segmentation recognition. DETAILED DESCRIPTION

[0054] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0055] like Figure 1 As shown, the present invention provides a split cable damage detection method based on Retinex theory, which mainly consists of the following parts:

[0056] 1) Application of Retinex theory: Retinex theory is used to enhance the brightness and contrast of images under low light conditions. Through the steps of image decomposition, enhancement and reconstruction, Retinex theory can effectively highlight the details and edge information in the image, making the damaged area of ​​the cable more obvious and easier to detect in low light environments.

[0057] 2) Introduction of variational mode decomposition (VMD) technology: VMD technology is introduced in the image processing process to extract high-frequency information. VMD can separate the high-frequency components of the image, thereby enhancing the detailed features of cracks and damaged areas and effectively reducing the interference of low-frequency noise. The extraction and utilization of high-frequency information is the key to maintaining high detection accuracy of the model under low-light conditions.

[0058] 3) Multi-source feature fusion: By fusing the image features enhanced by Retinex with the high-frequency features extracted by VMD, the advantages of both methods can be fully utilized. The fused feature map contains enhanced brightness, contrast, and highlighted detail information, which provides richer features for the subsequent segmentation network.

[0059] 4) Deep learning segmentation network integrating ASPP and SegNet: A deep learning segmentation network integrating the Atrous Spatial Pyramid Pooling (ASPP) module and the SegNet architecture is designed. The ASPP module can effectively capture features of different scales through multi-scale atrous convolution operations, enhancing the model's ability to detect cracks and damaged areas. The SegNet architecture further optimizes the segmentation effect through its encoding-decoding structure, allowing the model to maintain high-precision segmentation when dealing with complex backgrounds and diverse damage forms.

[0060] In this embodiment, the present invention specifically includes the following steps:

[0061] S1, low light cable image input;

[0062] S2, image data preprocessing;

[0063] S3, using the Retinex module to enhance the brightness and contrast features of the image;

[0064] S4, using variational mode decomposition VMD module to extract high-frequency features of the image;

[0065] S5, high-frequency modality fusion, generating enhanced images with richer details;

[0066] S6, using the atrous spatial pyramid pooling ASPP module, performs atrous convolution processing on brightness features, contrast features and high-frequency features, and fuses these features based on multi-scale feature extraction;

[0067] S7, a deep learning segmentation network using the SegNet architecture, optimizes the segmentation effect through its encoding-decoding structure.

[0068] In this embodiment, in step S1, the input low-light cable images are taken using an iPad Pro to provide sufficient resolution and clarity to truly reflect the actual situation of the cable. These images contain the original images of the cable, which may have low contrast and high noise because they are taken under low-light conditions. In the monitoring of thermal power plants, hyperspectral imaging can accurately identify different materials and states, such as exhaust from cooling towers and other industrial emissions. These emissions often have unique spectral features, such as spectral absorption peaks of specific chemical components.

[0069] In this embodiment, step S2 specifically includes, firstly, adjusting the size of the input image to 400×400 pixels to unify the image size for subsequent processing and analysis; then, denoising the image, because images under low light conditions are usually accompanied by higher noise levels, which will affect the accuracy of detection. Denoising is a key step, which improves image quality by removing random noise in the image and reducing unnecessary interference. In this step, advanced denoising algorithms are used, including non-local means (NLM) denoising and bilateral filters. These algorithms can effectively retain the detail information of the image while reducing noise, thereby improving the clarity of the image. Through this preprocessing step, it can be ensured that the input image used in the subsequent processing stage is of higher quality, which helps to improve the accuracy of cable damage detection.

[0070] In this embodiment, the step S3 specifically includes:

[0071] S3.1 Image Decomposition

[0072] First, the Retinex module decomposes the input preprocessed image. This process is based on the classic Retinex theory and is divided into reflected component and illumination component.

[0073] Reflection component: refers to the part of the image formed by the light reflected by the object itself, which contains the details and texture of the object; the reflection component is independent of lighting and therefore reflects the inherent characteristics of the object, such as the texture and crack details of the cable surface;

[0074] Illumination component: refers to the part of the image brightness change caused by changes in ambient lighting conditions; the illumination component affects the brightness and contrast of the entire image, such as shadows and dark areas caused by insufficient light.

[0075] S3.2 Image Enhancement

[0076] After image decomposition, the Retinex module enhances the image by processing the reflection and illumination components separately.

[0077] S3.3 Image reconstruction

[0078] After enhancing the reflection and illumination components, the Retinex module recombines the two components to generate two different sets of images: a contrast-enhanced image and a brightness-enhanced image.

[0079] Contrast-enhanced images: These detail images are combined with the adjusted illumination image by retaining the enhanced reflectance component to produce an image with sharp edges and details even in low-light conditions. The high contrast of this image helps accurately locate and identify cracks and breaks in the cable.

[0080] Brightness-enhanced images: Ensure cable visibility in the overall image by focusing on the enhanced lighting component. This process produces an evenly-lit image, reducing the occlusion effect in low-light areas and ensuring that the cable surface is clearly visible.

[0081] After being processed by the Retinex module, the two sets of enhanced images finally obtained provide favorable conditions for subsequent damage detection, so that cable damage and cracks can be accurately detected under various lighting conditions.

[0082] In this embodiment, the step S3.2 specifically includes the following steps:

[0083] 3.2.1. Reflection component enhancement

[0084] To generate contrast-enhanced images, the Retinex module enhances the reflected component, including enhancing details and edge information to make cracks and damaged areas in the cable more obvious. By adjusting the reflected component, the contrast between different areas in the image is increased, making the detailed features of the cable clearer.

[0085] Local contrast enhancement: Laplace enhancement, a local contrast adjustment technique, is used to enhance small-scale details and edges in the image, which helps to highlight cracks and defects on the cable surface;

[0086] High-pass filtering: The high-pass filter is used to enhance the high-frequency components in the image and further highlight the details; this processing can remove low-frequency information in smooth areas and focus on preserving and enhancing the contrast of crack edges.

[0087] 3.2.2 Lighting Component Adjustment

[0088] To generate a brightness-enhanced image, the Retinex module adjusts the illumination components to make the overall brightness of the image more uniform; thereby reducing the effect of shadows in the image, increasing the brightness of images taken in low-light conditions, and making the entire image more visible.

[0089] Global illumination equalization: Use Adaptive Histogram Equalization technology to smooth the illumination components, thereby reducing the extremely dark and extremely bright areas in the image and making the overall brightness distribution more uniform;

[0090] Local brightness enhancement: By adjusting the local brightness, the shadow area is appropriately enhanced, thereby reducing the visibility problem caused by low light; ensuring that the details of the cable are still visible under different lighting conditions.

[0091] In this embodiment, in step S4, the variational mode decomposition (VMD) module decomposes the image into multiple modes, each mode contains information of different frequencies, so as to distinguish high-frequency and low-frequency information in the image; through iterative optimization, the image is decomposed into a number of intrinsic mode functions (IMFs), and these IMFs respectively represent different frequency components in the image; each IMF corresponds to a characteristic mode in the image and contains signal information in a specific frequency range; through mode decomposition, the image is decomposed into multiple IMFs, and each IMF is a new image; some of these IMFs contain lower frequency information, including the overall outline and background of the image; some contain high frequency information, including details and edges in the image;

[0092] High-frequency modes containing cable damage information are selected from multiple IMFs. These high-frequency modes contain details of the cable damage area, such as the edges of cracks, small breaks, etc.

[0093] In this embodiment, in step S5, the enhanced image more clearly displays the edge and small damaged area of ​​the cable, providing clear visual information, so as to improve the accuracy and reliability of detection in the subsequent damage detection process; a weight is assigned to each high-frequency mode, and these weights are determined based on the frequency characteristics of the mode or by analyzing the mode; the details and edge information in the high-frequency mode are more significant, and these modes are given higher weights; and multiple high-frequency mode images are superimposed using the following formula:

[0094] Enhanced Image=w1×High-Frequency Mode1+w2×High-Frequency Mode2

[0095] Among them, ω1 and ω2 are the weight coefficients of each high-frequency mode, High-Frequency Mode1 and High-FrequencyMode2 are high-frequency modes;

[0096] Through the above-mentioned weighted superposition and fusion strategy, an enhanced high-frequency image is generated. Compared with the original image, this image significantly improves the contrast and clarity of details, especially at the edges and damaged areas of the cable; small cracks and damaged areas become more prominent, making them easier to identify visually.

[0097] In this embodiment, in step S6, the atrous spatial pyramid pooling ASPP module performs multi-scale feature extraction on each type of feature through a series of atrous convolution layers (for example, convolutions with expansion rates of 1, 6, and 12); all feature maps processed by atrous convolution are spliced, and then feature compression and fusion are performed through a 1x1 convolution layer to generate the final fused feature map. The feature map fused by the ASPP module contains high-frequency information, brightness information, and contrast information, which are effectively integrated at multiple scales. This fused feature can better identify the damaged area of ​​the cable in a low-light environment, ensure the clarity of edges and details, and thus provide more reliable input for subsequent segmentation networks (such as SegNet).

[0098] In this embodiment, in step S7, the encoder part is the same as the first 13 convolutional layers of the VGG16 network, consisting of several convolutional layers, batch normalization layers, and maximum pooling layers; after the maximum pooling layer, the pooling indices are saved for the deconvolution operation of the decoder part;

[0099] The decoder part inverts the encoder layer by layer and uses the maximum pooling index corresponding to the encoder for upsampling; through the deconvolution layer (upsampling), the spatial resolution of the feature map is restored, and the pooling index is used to ensure that the position information of the feature map is preserved; each decoder adds the feature map restored by deconvolution to the feature map of the corresponding encoder module element by element (skip connection) to fuse multi-scale information; the classifier part finally performs classification through a convolution layer and outputs the category probability of each pixel.

[0100] In this embodiment, the experimental results are as follows:

[0101] Image enhancement effects:

[0102] Contrast enhancement: By applying the Retinex theoretical module, the experimental results show that the overall brightness and contrast of the image are significantly improved. Details that were difficult to identify in the original low-light image become clearly visible in the enhanced image. This enhancement effect is particularly evident in the edges and cracks of the cable. By comparing the images before and after processing, it can be seen that the enhanced image is visually sharper and can better highlight the damaged areas of the cable.

[0103] High-frequency detail preservation: VMD theory further separates and strengthens the high-frequency information in the image. Experimental results show that the fusion of high-frequency modalities effectively preserves the details of cracks and small damaged areas, allowing these details to be accurately detected even under low-light conditions. These enhanced high-frequency images provide a more favorable information basis for the subsequent segmentation network.

[0104] VMD fusion information: Performance of ASPP+SegNet network: Using the enhanced image input ASPP (Atrous Spatial Pyramid Pooling) + SegNet segmentation network, the experimental results show that the network can effectively segment the damaged area in the cable. The ASPP module enhances the model's perception of damaged areas at different scales through multi-scale feature fusion. SegNet uses the encoder-decoder structure and combines the feature information output by ASPP to accurately segment the damaged area of ​​the cable.

[0105] Segmentation accuracy: In the experiment, we used mean intersection over union (mIOU) as the evaluation metric. The results show that the mIOU value of the proposed model is significantly higher than that of the traditional method under low-light conditions. This shows that the detection method combining Retinex theory, VMD theory and ASPP+SegNet architecture performs well in the cable damage detection task and can maintain high-precision detection capabilities under various low-light conditions.

[0106] Quantitative Analysis: Classification Accuracy: The proposed model also outperforms the traditional methods in the classification accuracy of cable damage detection. Through quantitative evaluation, SegNet has a mean intersection over union (mIOU) of 63.85%. The method of this application reaches 74.83%.

[0107] The technical means disclosed in the scheme of the present invention are not limited to the technical means disclosed in the above-mentioned implementation mode, but also include technical schemes composed of any combination of the above-mentioned technical features. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.

Claims

1. A segmented cable damage detection method based on Retinex theory, characterized in that: The following steps are involved: S1, low-light cable image input; S2, image data preprocessing; S3, using the Retinex module to enhance the brightness and contrast features of the image; S4, using variational mode decomposition VMD module to extract high-frequency features of the image; S5, high-frequency modality fusion, generating enhanced images with richer details; S6, using the atrous spatial pyramid pooling ASPP module, performs atrous convolution processing on brightness features, contrast features and high-frequency features, and fuses these features based on multi-scale feature extraction; S7, a deep learning segmentation network using the SegNet architecture, optimizes the segmentation effect through its encoding-decoding structure.

2. According to claim 1, a split cable damage detection method based on Retinex theory is characterized in that: In step S1, the input low-light cable image is taken using an iPad Pro to provide sufficient resolution and clarity to truly reflect the actual condition of the cable.

3. According to claim 2, a split cable damage detection method based on Retinex theory is characterized in that: The step S2 specifically includes, firstly, adjusting the size of the input image to 400×400 pixels; then, performing denoising processing on the image, including non-local mean denoising and bilateral filter.

4. According to claim 3, a split cable damage detection method based on Retinex theory is characterized in that: The step S3 specifically includes: S3.1, Image decomposition: First, the Retinex module decomposes the input preprocessed image, which is based on the classic Retinex theory and is divided into reflection component and illumination component; S3.2, image enhancement; After completing the image decomposition, the Retinex module enhances the image by processing the reflection and illumination components separately; S3.3, Image reconstruction; After the reflection and illumination components are enhanced, the Retinex module recombines the two components to generate two different sets of images: a contrast-enhanced image and a brightness-enhanced image; where Contrast-enhanced images: These detail images are combined with the adjusted illumination image to produce an image with sharp edges and details in low-light conditions by preserving the enhanced reflectance components. Brightness-enhanced images: Ensure cable visibility within the overall image by focusing on the enhanced lighting component; After being processed by the Retinex module, the two sets of enhanced images finally obtained provide favorable conditions for subsequent damage detection, so that cable damage and cracks can be accurately detected under various lighting conditions.

5. The method for detecting damage of split cables based on Retinex theory according to claim 4 is characterized in that: The step S3.2 specifically includes the following steps: 3.2.

1. Enhancement of reflection component: In order to generate contrast-enhanced images, the Retinex module enhances the reflection component, including enhancing details and edge information to make cracks and damaged areas in the cable more obvious; by adjusting the reflection component, the contrast between different areas in the image is improved, making the detailed features of the cable clearer; Local contrast enhancement: Laplace enhancement, a local contrast adjustment technique, is used to enhance small-scale details and edges in the image, which helps to highlight cracks and defects on the cable surface; High-pass filtering: The high-pass filter is used to enhance the high-frequency components in the image and further highlight the details; thereby removing the low-frequency information in the smooth area and focusing on preserving and enhancing the contrast of the crack edges; 3.2.

2. Adjustment of lighting components: To generate an image with enhanced brightness, the Retinex module adjusts the lighting components to make the overall brightness of the image more uniform, thereby reducing the shadow effect in the image, increasing the brightness of images taken under low light conditions, and making the entire image more visible. Global illumination equalization: Uses adaptive histogram equalization technology to smooth the illumination components; thereby reducing extremely dark and extremely bright areas in the image and making the overall brightness distribution more uniform; Local brightness enhancement: By adjusting the local brightness, the shadow area is appropriately enhanced, thereby reducing the visibility problem caused by low light; ensuring that the details of the cable are still visible under different lighting conditions.

6. A method for detecting damage of split cables based on Retinex theory according to claim 5, characterized in that: In step S4, the variational mode decomposition (VMD) module decomposes the image into multiple modes, each mode containing information of different frequencies, thereby distinguishing high-frequency and low-frequency information in the image; decomposing the image into a number of intrinsic mode functions (IMFs), some of which contain lower-frequency information, including the overall outline and background of the image; and some contain high-frequency information, including details and edges in the image; and selecting a high-frequency mode containing cable damage information from multiple IMFs, wherein the high-frequency mode contains details of the cable damage area.

7. A split cable damage detection method based on Retinex theory according to claim 6, characterized in that: In step S5, the enhanced image more clearly shows the edge and small damaged area of ​​the cable; a weight is assigned to each high-frequency mode, and these weights are determined based on the frequency characteristics of the mode or by analyzing the mode; the details and edge information in the high-frequency mode are more significant, and these modes are given higher weights; and multiple high-frequency mode images are superimposed using the following formula: Enhanced Image=w1×High-Frequency Mode1+w2×High-Frequency Mode2 Among them, ω1 and ω2 are the weight coefficients of each high-frequency mode, High-Frequency Mode1 and High-FrequencyMode2 are high-frequency modes; Through the above-mentioned weighted superposition and fusion strategy, an enhanced high-frequency image is generated. Compared with the original image, this image significantly improves the contrast and clarity of details, especially at the edges and damaged areas of the cable; small cracks and damaged areas become more prominent, making them easier to identify visually.

8. The method for detecting damage of split cables based on Retinex theory according to claim 7, characterized in that: In step S6, the atrous spatial pyramid pooling ASPP module extracts multi-scale features for each type of feature through a series of atrous convolution layers; all feature maps processed by atrous convolution are concatenated, and then feature compression and fusion are performed through a 1x1 convolution layer to generate a final fused feature map.

9. A method for detecting damage of split cables based on Retinex theory according to claim 8, characterized in that: In step S7, the encoder part is the same as the first 13 convolutional layers of the VGG16 network, consisting of several convolutional layers, batch normalization layers and maximum pooling layers; after the maximum pooling layer, the pooling index is saved for the deconvolution operation of the decoder part; the decoder part inverts the encoder layer by layer and uses the maximum pooling index corresponding to the encoder for upsampling; The spatial resolution of the feature map is restored through the deconvolution layer, and the pooling index is used to ensure that the position information of the feature map is preserved; each decoder adds the feature map restored by deconvolution to the feature map of the corresponding encoder module element by element to fuse multi-scale information; The classifier part finally performs classification through a convolutional layer and outputs the category probability of each pixel.

Images