Insulator surface defect detection method, system, storage medium and device

Abstract

The present invention proposes an insulator surface defect detection method, system, storage medium and equipment, and relates to the technical field of power equipment surface defect detection. The method comprises: obtaining a visible light image and an infrared image of an insulator, and preprocessing the images respectively; performing image registration and image fusion on the preprocessed visible light image and infrared image to obtain a fused image; inputting the fused image into a trained insulator surface defect detection model for detection to obtain an insulator surface defect detection result; the present invention fuses the infrared image and the visible light image to complement the advantages of the two images, and uses the advantages of the infrared image to compensate for the disadvantages of the visible light, so that the detection work can be carried out around the clock and is not easily disturbed by the external environment. In addition, by improving the detection model, the accuracy of the detection result is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of surface defect detection of power equipment, and in particular relates to a method, system, storage medium and equipment for detecting surface defects of insulators. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The insulating string in the high-voltage line is composed of insulators stacked together. Because of its good insulation performance, it is used in high-voltage overhead transmission and distribution lines to play a role in large-scale and efficient transmission of electricity. Insulators are exposed to the external environment for a long time and are easily affected by rain, snow, haze, sunlight, strong electric fields, and strong mechanical tension, which can cause defects such as damage, self-explosion, and dirtiness in the insulators. Under the influence of the harsh external environment, some small defects, such as rust, cracks, no glaze, spots, etc., will quickly grow larger, seriously shortening the life of ceramic insulators and thus reducing the efficiency of high-voltage lines. Therefore, there are high requirements for the surface integrity of ceramic insulators, and the requirements for their quality inspection are becoming more and more stringent.

[0004] The defect diagnosis methods based on image recognition mainly include defect detection based on infrared images and visible light images. Chinese invention patent CN116183621A discloses a porcelain insulator quality detection method and system, which obtains the visible light image of the insulator and inputs the visible light image into the detection model for detection; the visible light image imaging can collect images with more high-frequency components, and the more high-frequency information, the more detailed information of the image, and the higher the clarity. However, it is susceptible to environmental interference, has weak fog penetration ability, and when visibility is low, the collected image will be unclear and the image effect will be poor, so it cannot work all day. Chinese invention patent CN111652857A discloses an insulator defect infrared detection method, which obtains the infrared image of the insulator and inputs the visible light image into the detection model for detection; the imaging principle of the infrared image is to capture the infrared light waves generated on the object through an infrared thermal imager, and generate an infrared image in the thermal imager, which can reflect the temperature distribution of the object. This imaging method can work all day, has strong fog penetration ability and strong resistance to external environmental interference, and can work well even at night. However, the pixels, contrast and clarity of infrared imaging are relatively low, and the image detail information is not comprehensive enough, which can easily be regarded as suspicious defects by the convolutional neural network (CNN), causing great interference to the model's judgment, causing the model to easily misjudge difficult-to-classify samples; although the commonly used target detection SSD algorithm greatly reduces the parameter size of the model and improves the model's computational efficiency, the accuracy of the model's recognition results is insufficient. Summary of the invention

[0005] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a method, system, storage medium and device for detecting surface defects of insulators, which improves the accuracy of the detection results by fusing infrared images and visible light images and improving the detection model.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0007] In a first aspect, the present invention provides a method for detecting surface defects of an insulator, comprising the following steps:

[0008] Obtaining a visible light image and an infrared image of the insulator, and preprocessing the visible light image and the infrared image respectively;

[0009] The pre-processed visible light image and infrared image are firstly registered and then fused to obtain a fused image;

[0010] The fused image is input into the trained insulator surface defect detection model for detection to obtain the insulator surface defect detection result;

[0011] Among them, the acquisition process of the fused image is: first, the aligned visible light image and infrared image are decomposed into a low-rank part, a visually salient part and a sparse noise part using potential low-rank representation respectively; then the low-rank part of the visible light image and the low-rank part of the infrared image are fused to obtain a low-rank part fused image; then the visually salient part of the visible light image and the visually salient part of the infrared image are fused to obtain a visually salient part fused image; finally, the sparse noise part is directly discarded, and the fused image is reconstructed using the low-rank part fused image and the visually salient part fused image.

[0012] Furthermore, the image preprocessing process includes image denoising and image resolution enhancement, wherein the image denoising adopts a three-dimensional matching algorithm for denoising, and the image resolution enhancement adopts an improved automatic color balancing algorithm for enhancement.

[0013] Furthermore, the improved automatic color balancing algorithm firstly determines the brightness of the original image, and then performs the automatic algorithm balancing operation after the brightness thereof meets a certain range.

[0014] Furthermore, the visible light image and the infrared image are registered using a feature point matching method, and the feature points are extracted using an accelerated robust feature algorithm.

[0015] Furthermore, the low-rank part adopts a feature extraction algorithm based on the VGG network combined with a multi-layer fusion strategy to obtain a low-rank partial fusion image, specifically including: using the VGG network to extract features; obtaining a feature map based on the l1 norm and regional average operator; obtaining weight values ​​through soft maximum processing; reconstructing the low-rank partial image, and using the maximum selection strategy to obtain a low-rank partial fusion image.

[0016] Furthermore, the visually salient parts are fused using a regional energy-based fusion method to obtain a visually salient part fused image, specifically including: first determining the region size, then calculating the energy of the area around each point, and finally fusing the visually salient parts using a fusion rule with the largest regional energy.

[0017] Furthermore, the insulator surface defect detection model adopts an improved YOLOX-s target detection model, in which multi-scale depth-separable convolution is introduced in the spatial pyramid pooling at the tail end of the backbone network, and the SENet attention mechanism is added to the head prediction network part.

[0018] A second aspect of the present invention provides an insulator surface defect detection system, comprising:

[0019] The image acquisition and preprocessing module is configured to: acquire a visible light image and an infrared image of the insulator, and preprocess the visible light image and the infrared image respectively;

[0020] The image fusion module is configured to: first perform image registration and then perform image fusion on the preprocessed visible light image and infrared image to obtain a fused image;

[0021] The detection module is configured to: input the fused image into a trained insulator surface defect detection model for detection, and obtain an insulator surface defect detection result;

[0022] Among them, the acquisition process of the fused image is: first, the aligned visible light image and infrared image are decomposed into a low-rank part, a visually salient part and a sparse noise part using potential low-rank representation respectively; then the low-rank part of the visible light image and the low-rank part of the infrared image are fused to obtain a low-rank part fused image; then the visually salient part of the visible light image and the visually salient part of the infrared image are fused to obtain a visually salient part fused image; finally, the sparse noise part is directly discarded, and the fused image is reconstructed using the low-rank part fused image and the visually salient part fused image.

[0023] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the insulator surface defect detection method as described in the first aspect of the present invention.

[0024] A fourth aspect of the present invention provides an electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method described in the first aspect of the present invention are implemented.

[0025] One or more of the above technical solutions have the following beneficial effects:

[0026] The insulator surface defect detection method provided by the present invention fuses visible light images and infrared images, uses the fused image as the detection image, complements the advantages of the two images, uses the advantages of infrared images to make up for the disadvantages of visible light, enables detection work to be carried out around the clock, is not easily affected by external environment, and increases the ability of the detection system to resist external interference.

[0027] The insulator surface defect detection method provided by the present invention first pre-processes the infrared image and the visible light image, and then uses an accelerated robust feature algorithm to match the feature points of the two images, so that the two images can be aligned in different dimensions, and then determine the mapping relationship between the two, laying a foundation for image fusion; wherein, an improved automatic color balancing algorithm is used for image enhancement during the pre-processing process, which not only well preserves the structure of the original image, but also significantly improves the contrast and saturation of the image.

[0028] In the insulator surface defect detection method provided by the present invention, in the process of image fusion of visible light images and infrared images, the external image adopts potential low-rank representation to decompose the visible light image and the infrared image into a low-rank part, a visually significant part and a sparse noise part, the low-rank part adopts a feature extraction algorithm based on a VGG network combined with a multi-layer fusion strategy to obtain a low-rank part fusion image; the visually significant part is fused using a regional energy-based fusion method to achieve local fusion, the sparse noise part is directly discarded, and the fused low-rank part and the visually significant part are used to reconstruct the fused image to achieve global fusion, thereby improving the image fusion effect through step-by-step fusion.

[0029] The insulator surface defect detection method provided by the present invention improves the YOLOX-s target detection algorithm and reduces the information weight in important channels by introducing the classic attention mechanism SENet, thereby effectively reducing the influence of interference information, thereby improving the detection accuracy of insulator defects and solving the problem of complex background interference; by inserting a depthwise separable convolution into the spatial pyramid pooling module at the tail end of the backbone network, multi-scale information containing contextual features is fully extracted, thereby solving the problem of information loss of small-scale defects of insulators.

[0030] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0032] Figure 1 The present invention is a flow chart of the insulator surface defect detection method. DETAILED DESCRIPTION

[0033] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0034] Embodiment 1

[0035] This embodiment discloses a method for detecting surface defects of an insulator. Figure 1 As shown, the following steps are included:

[0036] Step 1: Obtain a visible light image and an infrared image of the insulator, and preprocess the visible light image and the infrared image respectively.

[0037] In this embodiment, the image preprocessing process includes image denoising and image resolution enhancement, wherein the image denoising adopts a three-dimensional matching algorithm for denoising, and the three-dimensional matching algorithm has a good denoising effect on both Gaussian mixture noise and salt and pepper noise.

[0038] In this embodiment, the image resolution enhancement is performed using an improved automatic color balancing algorithm. The automatic color balancing algorithm can make the image more consistent with human eye perception. By calculating the difference between the relative brightness and darkness of the target point and the surrounding pixels, the image is corrected and the final pixel value is obtained. Although the traditional automatic color balancing algorithm has made great improvements in image dehazing, its slow processing speed also greatly limits its application. To this end, this embodiment proposes to improve the automatic color balancing algorithm. The improved original first performs brightness judgment on the original image to make its brightness meet a certain range, and then performs the automatic algorithm balancing operation. The specific process is as follows:

[0039] (1) Image input, obtain the RMS of image pixels, return perceived brightness, and perform white balance processing;

[0040] (2) Determine whether the mean brightness of the image is less than or equal to 140. If so, perform logarithmic transformation on the image to increase the brightness of the dark image. If not, perform pyramid transformation, color / spatial domain adjustment, dynamic expansion, and filtering / pyramid restoration on the image features.

[0041] (3) Output the processed image.

[0042] By adding the logarithmic transformation of the image to the original algorithm, the brightness of the image is greatly improved, and then the image color temperature is balanced through white balance. Before automatic color balancing, the image brightness is judged first. When the image brightness is dark, the logarithmic transformation is introduced to improve the image brightness. When the image brightness meets the requirements, the next step is directly performed. After the improved automatic color balancing process, the proportion of the three primary colors in the histogram is more balanced and the color contrast is more obvious.

[0043] Step 2: Perform image registration and then image fusion on the preprocessed visible light image and infrared image to obtain a fused image.

[0044] Among them, the acquisition process of the fused image is: first, the aligned visible light image and infrared image are decomposed into a low-rank part, a visually salient part and a sparse noise part using potential low-rank representation respectively; then the low-rank part of the visible light image and the low-rank part of the infrared image are fused to obtain a low-rank part fused image; then the visually salient part of the visible light image and the visually salient part of the infrared image are fused to obtain a visually salient part fused image; finally, the sparse noise part is directly discarded, and the fused image is reconstructed using the low-rank part fused image and the visually salient part fused image.

[0045] In this embodiment, the visible light image and the infrared image are registered using the existing feature point matching method, and the feature points are extracted using an accelerated robust feature algorithm, so that the two images can be aligned in different dimensions, and then the mapping relationship between the two is determined, laying the foundation for image fusion.

[0046] The accelerated robust feature algorithm is an algorithm that approximates the Gaussian second-order derivative filter through a box filter. The return image of the Hessian matrix is ​​calculated by taking the value of the box filter. The expression of the Hessian matrix is:

[0047]

[0048] In the formula, σ represents the scale of the image, L xx (x,σ) represents the convolution value of the image at the pixel point (x,y) with the second-order partial derivative of the Gaussian function of σ, L xy (x,σ),L yy (x,σ) can be obtained in the same way.

[0049] As an unsupervised feature extraction algorithm, latent low-rank representation can completely extract the global and local structural information of the image, and can better extract the salient target features of the source image compared to low-rank decomposition. In addition, the source image does not change the rank structure of the data matrix after being decomposed by the latent low-rank representation. After decomposition by the latent low-rank representation, the low-rank part (global structure), visually salient part (local structure) and sparse noise part of the source image can be obtained.

[0050] Furthermore, the low-rank part adopts a feature extraction algorithm based on the VGG network combined with a multi-layer fusion strategy to obtain a low-rank partial fusion image, specifically including:

[0051] (1) Using VGG network to extract features:

[0052] The low-rank parts {IR -lrr , VIS -lrr} are respectively input into the trained VGG-19 neural network, and the output is in, Represents the feature map of the kth low-rank part extracted by the jth convolution group; in this embodiment, the value of k is 2, representing infrared and visible light images respectively; j∈{1, 2, 3, 4}, representing 4 convolution groups, namely relu1-1, relu2-1, rulu3-1, relu4-1; n is the total number of channels of the jth layer output feature map, which is determined by the number of convolution kernels, n∈{1, 2...N}, N=64*2 i-1 ;make Represents the output feature map At the value at (x,y), we can see is an N-dimensional vector.

[0053] (2) Obtain feature maps based on L1 norm and regional average operator:

[0054] The L1 norm can achieve the sparsity of feature parameters and can compress the output feature map from multiple channels to a single channel. The L1 norm of can be used as an activity-level measure of the feature map of the low-rank part of the image, so the initial feature map can be obtained The formula is as follows:

[0055]

[0056]

[0057] In order to increase the robustness of the algorithm to misregistration, a region-based average operator is used to obtain the final feature map The formula is as follows:

[0058]

[0059]

[0060] In the formula, θ determines the size of the region. In order to achieve a better fusion effect, θ is set to 1 in this embodiment.

[0061] (3) Soft maximum processing to obtain weight value:

[0062] The final feature map After being processed by the soft maximum operator, the initial normalized adaptive weight value is obtained The formula is as follows:

[0063]

[0064]

[0065] The pooling layer in the VGG-19 network is a downsampling process that can reduce the feature dimension of the image and reduce the number of data parameters. Each time the convolutional neural network performs a pooling operation, the output feature map is 1 / r of the input image size, where r is the step size (VGG step size is 2). Therefore, in different convolutional layers of the VGG-19 network, the size of the feature map is 1 / 2 of the source image. i-1 , i∈{1,2,3,4}. In order to make the fused image consistent with the source image size, the upsampling operator is used to adjust the initial weight value to the original image size, and the final weight value is

[0066] (4) Low-rank partial image reconstruction: a maximum selection strategy is used to obtain a low-rank partial fusion image.

[0067] After the final weight values ​​are obtained, the weight maps of the low-rank parts of the four pairs of images are obtained. Next, the low-rank part images are reconstructed to obtain four pairs of low-rank part fusion images. The formula is as follows:

[0068]

[0069] The maximum selection strategy is adopted for the four low-rank partial fusion images at each position to obtain the final low-rank partial fusion image {F lrr}, the formula is as follows:

[0070]

[0071] Furthermore, the visually significant parts are fused using a fusion method based on regional energy to obtain a fused image of the visually significant parts, which specifically includes: first determining the size of the region, then calculating the energy of the area around each point, and finally fusing the visually significant parts using a fusion rule that takes the largest regional energy:

[0072]

[0073] Where Q represents the range of the window, E(m,n) is the regional energy, and d(m+x,n+y) represents the subband coefficient of the source image at (m+x,n+y);

[0074] Finally, the visually significant parts are fused by taking the largest fusion rule of regional energy:

[0075]

[0076] Where, d F (m,n),d VIS (m,n),d IR (i, j) represent the sparse matrices of the fused image, visible light image and infrared image respectively, E VIS (m,n),E IR (m,n) represent the regional energy in the visible light image and infrared image respectively.

[0077] After obtaining the fused low-rank part and the visually salient part, the two components are superimposed and restored to obtain the final fused image.

[0078] Step three, input the fused image into the trained insulator surface defect detection model for detection to obtain the insulator surface defect detection result; the insulator surface defect detection model adopts an improved YOLOX-s target detection model, and the network structure of the improved YOLOX-s target detection model includes an input end, a backbone network, a feature fusion network and a head prediction network. The improved YOLOX-s target detection model introduces multi-scale depth-separable convolution in the spatial pyramid pooling at the tail end of the backbone network, and adds the SENet attention mechanism to the head prediction network part.

[0079] Specifically, (1) the SPP structure in the backbone network Dark5 is optimized by multi-scale deep separable convolution, and a multi-scale context-sensitive module is constructed. The insulator feature map output by the Dark2, Dark3, and Dark4 layers is used as input and divided into two paths. One path is subjected to three maximum pooling operations with convolution kernel sizes of 5×5, 9×9, and 13×13, which increases the receptive field and enables the network to better understand the overall characteristics of the insulator; the other path is subjected to three deep separable convolution operations with convolution kernel sizes of 3×3, 5×5, and 9×9, which is used to extract more fine-grained local features of the insulator fault area. The two paths of information are fused through the Concat operation, and the output contains a feature map containing fused multi-scale information, which can maintain the resolution while increasing the receptive field, thereby more comprehensively capturing the characteristics of the target in insulator defect detection.

[0080] (2) The addition of SENet to the head prediction network of YOLOX-S is mainly based on the following two considerations: first, adding SENet to the head prediction network can improve the feature extraction capability of the head prediction network; second, adding SENet to the head prediction network will not bring a huge amount of parameters to the network.

[0081] The attention mechanism can increase the weight of information in relatively important channels. Compared with not in channel attention, the input feature map U has a total of C channels, where each channel size is H×W, and the weight of the kth channel in the feature map U is Z c The calculation formula is as follows:

[0082]

[0083] In the formula, Z c Represents the weight of each channel, so the feature map U is obtained by F sq (u c ) will get a one-dimensional array 1×1×C, u c (i, j) represents the point with horizontal and vertical coordinates i and j respectively on the cth channel in the feature map U.

[0084] To use F sq (u c ) The aggregated information after the operation models the dependency between C channels. Next, the second operation is performed, using the sigmoid activation function to achieve this goal. The dependency formula between the channel weights is as follows:

[0085] S=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0086] Among them, σ and δ represent the activation functions of sigmoid and ReLU respectively.

[0087] Finally, each channel of the input feature map U is weighted, so the feature map U' weighted by the SENet attention mechanism is obtained. The specific formula is as follows:

[0088]

[0089] Among them, U c represents the feature map before the attention mechanism, and U' represents the feature map after the attention mechanism.

[0090] In view of the small scale of insulator defect targets and the easy disappearance of features, which leads to missed detection, false detection and inaccurate positioning under complex backgrounds, the baseline network YOLOX-s is improved. First, the CA coordinate attention module is inserted to improve the backbone network's ability to extract features of insulator defect areas under the interference of complex backgrounds of transmission lines; the SENet attention mechanism is added to the head prediction network of YOLOX-S. This embodiment introduces the classic attention mechanism SENet, and the information weight in important channels is reduced, which can effectively reduce the impact of interference information, thereby improving the accuracy of insulator surface defect detection.

[0091] In this embodiment, the insulator surface defect detection results include insulator self-explosion, insulator segment cracks, and insulator surface contamination.

[0092] Embodiment 2

[0093] This embodiment discloses an insulator surface defect detection system, including:

[0094] The image acquisition and preprocessing module is configured to: acquire a visible light image and an infrared image of the insulator, and preprocess the visible light image and the infrared image respectively;

[0095] The image fusion module is configured to: first perform image registration and then perform image fusion on the preprocessed visible light image and infrared image to obtain a fused image;

[0096] The detection module is configured to: input the fused image into a trained insulator surface defect detection model for detection, and obtain an insulator surface defect detection result;

[0097] Among them, the acquisition process of the fused image is: first, the aligned visible light image and infrared image are decomposed into a low-rank part, a visually salient part and a sparse noise part using potential low-rank representation respectively; then the low-rank part of the visible light image and the low-rank part of the infrared image are fused to obtain a low-rank part fused image; then the visually salient part of the visible light image and the visually salient part of the infrared image are fused to obtain a visually salient part fused image; finally, the sparse noise part is directly discarded, and the fused image is reconstructed using the low-rank part fused image and the visually salient part fused image.

[0098] Embodiment 3

[0099] The purpose of this embodiment is to provide a computer-readable storage medium.

[0100] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the insulator surface defect detection method as described in Example 1 of the present disclosure.

[0101] Embodiment 4

[0102] The purpose of this embodiment is to provide an electronic device.

[0103] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the insulator surface defect detection method as described in Embodiment 1 of the present disclosure are implemented.

[0104] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0105] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0106] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for detecting surface defects of an insulator, characterized in that: The following steps are involved: Obtaining a visible light image and an infrared image of the insulator, and preprocessing the visible light image and the infrared image respectively; The pre-processed visible light image and infrared image are firstly registered and then fused to obtain a fused image; The fused image is input into the trained insulator surface defect detection model for detection to obtain the insulator surface defect detection result; The image preprocessing process includes image denoising and image resolution enhancement, wherein the image denoising adopts a three-dimensional matching algorithm for denoising, and the image resolution enhancement adopts an improved automatic color balancing algorithm for enhancement; The improved automatic color balancing algorithm first determines the brightness of the original image, and then performs the automatic algorithm balancing operation after the brightness is within a certain range. The specific process is as follows: Image input, get the image pixel root mean square, return the perceived brightness, and perform white balance processing; Determine whether the mean brightness of the image is less than or equal to 140. If so, perform logarithmic transformation on the image to increase the brightness of the dark image. If not, perform pyramid transformation, color / spatial domain adjustment, dynamic expansion, filtering / pyramid restoration on the image features. Output the processed image; The acquisition process of the fused image is as follows: firstly, the registered visible light image and the infrared image are respectively decomposed into a low-rank part, a visually significant part and a sparse noise part by using a potential low-rank representation; then, the low-rank part of the visible light image and the low-rank part of the infrared image are fused to obtain a low-rank part fused image; then, the visually significant part of the visible light image and the visually significant part of the infrared image are fused to obtain a visually significant part fused image; finally, the sparse noise part is directly discarded, and the fused image is reconstructed by using the low-rank part fused image and the visually significant part fused image; The low-rank part uses a feature extraction algorithm based on the VGG network combined with a multi-layer fusion strategy to obtain a low-rank partial fusion image, specifically including: using the VGG network to extract features; obtaining feature maps based on the l1 norm and regional average operator; obtaining weight values ​​through soft maximum processing; reconstructing the low-rank partial image, and using the maximum selection strategy to obtain a low-rank partial fusion image; The insulator surface defect detection model adopts an improved YOLOX-s target detection model. In the improved YOLOX-s target detection model, multi-scale depth-separable convolution is introduced in the spatial pyramid pooling at the tail end of the backbone network, and the SENet attention mechanism is added to the head prediction network part.

2. The insulator surface defect detection method according to claim 1, characterized in that: The visible light image and infrared image are registered using the feature point matching method, and the feature points are extracted using the accelerated robust feature algorithm.

3. The insulator surface defect detection method according to claim 1, characterized in that: The visually salient parts are fused using a regional energy-based fusion method to obtain a visually salient part fusion image, which specifically includes: first determining the region size, then calculating the energy of the area around each point, and finally fusing the visually salient parts using a fusion rule that takes the largest regional energy.

4. An insulator surface defect detection system, characterized in that: include: The image acquisition and preprocessing module is configured to: acquire a visible light image and an infrared image of the insulator, and preprocess the visible light image and the infrared image respectively; The image fusion module is configured to: first perform image registration and then perform image fusion on the preprocessed visible light image and infrared image to obtain a fused image; The detection module is configured to: input the fused image into a trained insulator surface defect detection model for detection, and obtain an insulator surface defect detection result; The image preprocessing process includes image denoising and image resolution enhancement, wherein the image denoising adopts a three-dimensional matching algorithm for denoising, and the image resolution enhancement adopts an improved automatic color balancing algorithm for enhancement; The improved automatic color balancing algorithm first determines the brightness of the original image, and then performs the automatic algorithm balancing operation after the brightness is within a certain range. The specific process is as follows: Image input, get the image pixel root mean square, return the perceived brightness, and perform white balance processing; Determine whether the mean brightness of the image is less than or equal to 140. If so, perform logarithmic transformation on the image to increase the brightness of the dark image. If not, perform pyramid transformation, color / spatial domain adjustment, dynamic expansion, filtering / pyramid restoration on the image features. Output the processed image; The acquisition process of the fused image is as follows: firstly, the registered visible light image and the infrared image are respectively decomposed into a low-rank part, a visually significant part and a sparse noise part by using a potential low-rank representation; then, the low-rank part of the visible light image and the low-rank part of the infrared image are fused to obtain a low-rank part fused image; then, the visually significant part of the visible light image and the visually significant part of the infrared image are fused to obtain a visually significant part fused image; finally, the sparse noise part is directly discarded, and the fused image is reconstructed by using the low-rank part fused image and the visually significant part fused image; The low-rank part uses a feature extraction algorithm based on the VGG network combined with a multi-layer fusion strategy to obtain a low-rank partial fusion image, specifically including: using the VGG network to extract features; obtaining feature maps based on the l1 norm and regional average operator; obtaining weight values ​​through soft maximum processing; reconstructing the low-rank partial image, and using the maximum selection strategy to obtain a low-rank partial fusion image; The insulator surface defect detection model adopts an improved YOLOX-s target detection model. In the improved YOLOX-s target detection model, multi-scale depth-separable convolution is introduced in the spatial pyramid pooling at the tail end of the backbone network, and the SENet attention mechanism is added to the head prediction network part.

5. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the insulator surface defect detection method as described in any one of claims 1 to 3 are implemented.

6. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the insulator surface defect detection method according to any one of claims 1 to 3 are implemented.

Images