Local super-resolution reconstruction algorithm based on infrared image of transmission line insulator

Abstract

The local super-resolution reconstruction algorithm based on the transmission line insulator infrared image includes the following steps: Step 1, image recognition; Step 2, image segmentation; Step 3, super-resolution reconstruction; By using the improved adversarial network structure, unsupervised reconstruction is adopted, the RRDB network structure is adopted, when the network structure is deepened, the stability of training can be maintained, the calculation amount and the use of memory are reduced, and the reconstruction effect is improved, the SeNET attention mechanism is added, from the feature channel level, according to the importance, the useful features such as the outline and the details of the insulator are improved, and the features which are not useful for the program are suppressed. The resolution of the current transmission line insulator infrared image is low, and when the insulator detection is carried out, the workload is large and the problems such as false detection and missed detection occur.

Description

Technical Field

[0001] This invention relates to the field of image processing technology for power transmission lines, and specifically to a local super-resolution reconstruction algorithm based on infrared images of power transmission line insulators. Background Technology

[0002] Infrared imaging technology is an imaging technique that reflects the specific content and temperature field information of a scene image by observing the differences in thermal radiation between targets and background points. Infrared imaging technology has numerous applications in civilian, industrial, and military fields. Due to its advantages such as strong penetration, immunity to electromagnetic interference, sensitivity to thermal radiation, and night vision capabilities, in recent years, maintenance personnel have used infrared thermal imagers mounted on monitoring platforms such as inspection robots and drones to conduct intelligent inspections of electrical equipment, reducing the workload of collecting infrared images and improving inspection efficiency. However, with the increasing precision of electrical equipment and the lack of improvement in hardware such as infrared detectors, the resolution of infrared scanned images is relatively low. Therefore, high-resolution reconstruction of infrared images of power equipment has attracted relevant attention. This paper, "A Super-Resolution Fault Identification Method for Electrical Equipment Based on a Multi-Scale Collaborative Model," proposes a super-resolution fault identification method for electrical equipment based on infrared images using a multi-scale collaborative model. This paper first constructs a super-resolution reconstruction network for infrared images of electrical equipment based on a multi-scale cooperative model. This network, based on generative adversarial networks (GANs), improves the adaptability of the super-resolution reconstruction network to infrared images and optimizes image feature extraction by introducing a multi-scale cooperative model and a dual-channel structure. CN113920015A discloses an edge-preserving super-resolution reconstruction method for infrared images based on GANs, comprising a two-layer GAN that decomposes the image generation process into two stages. Patent CN202010360557 discloses a super-resolution reconstruction method for infrared images based on GANs, improving the existing SRGAN algorithm in two aspects: the generator network and the loss function. In improving the generator network structure, the generator network incorporates traditional bicubic interpolation. In improving the loss function, to achieve high objective evaluation metrics (peak signal-to-noise ratio and structural similarity) while maintaining good visual effects, a pixel-wise mean squared error loss is added to the generator network's loss function. Current patents focus on high-resolution reconstruction of the original object as a whole, which suffers from high memory consumption, long computation time, and the generation of images containing a lot of useless information. Furthermore, reconstruction of complex background images is prone to introducing interference and noise, affecting the detection and recognition of targets. This is especially true when inspecting transmission line insulators, where the large number of insulators and small infrared temperature variations make them prone to missed or false detections. Therefore, improving the resolution of infrared images is crucial to enhancing the efficiency and accuracy of transmission line insulator detection.

[0003] With the rapid development of computer technology, image reconstruction technology has become increasingly mature, and super-resolution image reconstruction has been applied in image compression, medical detection, remote sensing mapping, and public safety. Image super-resolution reconstruction technology can be categorized by reconstruction space into spatial domain, frequency domain, and hybrid spatial-frequency domain super-resolution reconstruction algorithms. Based on research methods, super-resolution reconstruction algorithms can be classified into interpolation-based algorithms, such as nearest neighbor interpolation and bicubic interpolation; constraint-based algorithms, such as iterative back-projection and convex set projection; and learning-based algorithms, such as SRCNN, FSRCNN, SRResNet, VDSR, and SRGAN.

[0004] Current image super-resolution technology has achieved good results in the reconstruction of natural images, but there is still room for improvement in the reconstruction of pseudo-color images such as infrared images. Infrared images of power transmission lines have complex backgrounds and high levels of interference. When reconstructing electrical equipment in these images, the background is often enhanced simultaneously, resulting in a blurred difference between the background and the target equipment, making detection quite difficult. Therefore, this invention proposes improvements to current super-resolution reconstruction technology to achieve better results in infrared image reconstruction. Summary of the Invention

[0005] The technical problem this invention aims to solve is to provide a local super-resolution reconstruction algorithm based on infrared images of transmission line insulators. This algorithm addresses the issues of low resolution in current infrared images of transmission line insulators, which leads to a large workload and problems such as false detections and missed detections during insulator inspection. Because the background of infrared images of transmission lines is complex and its color significantly interferes with the foreground target, this invention comprehensively considers both local image reconstruction and attention mechanisms to optimize the image super-resolution reconstruction algorithm, thereby improving the resolution and detection accuracy of infrared images of transmission line insulators.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] The local super-resolution reconstruction algorithm based on infrared images of transmission line insulators includes the following steps:

[0008] Step 1, Image Recognition: First, label the insulators in the infrared image, train the image recognition algorithm, and then input the dataset into the image recognition algorithm to recognize the insulators in the infrared image of the transmission line. The recognized image is marked with a recognition box, and the recognized image is used as the input for the subsequent steps.

[0009] Step 2, Image Segmentation: The recognized image is fed into the image segmentation network. The network will take the insulator that has been located in the recognition box as the target and the rest of the image as the background, and perform image segmentation to separate the foreground target from the background. The final result is that except for the insulator in the box, which is still an infrared image, all other backgrounds are changed to black.

[0010] Step 3, Super-resolution reconstruction: The segmented image is fed into an adversarial network for training to obtain the reconstructed super-resolution infrared image of the insulator.

[0011] The adversarial network structure in Step 3 above is as follows:

[0012] Random noise is input into the generator G for network training, and then data is generated. The data generated by the generator G and the real data are fed into the discriminator D. The discriminator D determines whether the input data is real data or data generated by the generator G, and outputs either 0 (generated data from the generator G) or 1 (real data). This output is fed into the generator G, and the generator G learns again based on the feedback. This process is repeated, and the generator G and the discriminator D learn from each other and compete with each other until a Nash equilibrium is reached.

[0013] The Nash equilibrium formula achieved by the aforementioned adversarial network is:

[0014]

[0015] In the formula, p data(x) Based on real data distribution, p z(z) This refers to the data distribution generated by the generator; ultimately, the discriminator D cannot accurately determine whether the input data is real data or data generated by the generator G. That is, the discriminator D's accuracy in judging the input data is only 50%, and the generator G achieves the optimal effect.

[0016] Choosing MSE as the loss function, the calculation formula is as follows:

[0017]

[0018] In the formula, r represents the magnification factor, and W and H are the width and height of the image, respectively.

[0019] In Step 3 above, image super-resolution reconstruction can achieve a larger Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). However, a larger PSNR does not necessarily mean a better-looking reconstructed image. Therefore, an overall loss function L is adopted.

[0020] L = L per +λL adv +μL con (3)

[0021] In equation (3), L con Let L be the MSE loss function. per To perceive loss, L adv To counteract the loss, λ and μ are the corresponding weights. The perceptual loss is provided by a pre-trained VGG19 network, and its calculation formula is as follows:

[0022]

[0023] In the formula, W ij and H ij These are the width and height of the feature map obtained through j convolutions before the i-th max-pooling layer in the discrimination module, respectively; x and y correspond to the x and y coordinates of a current pixel in the feature map, respectively; φ ij It is the feature map obtained by j convolutions before the i-th max pooling layer in the discrimination module; HR It is a high-resolution image; I LR It is a low-resolution image; G is the generator network.

[0024] In equation (3) above, L adv The binary cross-entropy loss (BCEloss) function is used. In the training of the adversarial network, the output of the discriminator D is a binary classification of 0 or 1. A weighted BCEloss is used to constrain the overall loss function, and its calculation formula is as follows:

[0025] L adv =-(Dlog(D(G(z)))+(1-D)log(1-D(G(z)))) (5)

[0026] In the formula, D represents the label information of the real data; D(G(z)) represents the data information generated by the generator.

[0027] The generator G described above embeds an SENet network. During training, the SENet network is used to reconstruct the details and edges of the insulator. The SENet network structure is as follows:

[0028] Given X as the input to feature channel C1, it passes through multiple convolutions and transformations to enter feature channel C2. First, a Squeeze operation is performed, compressing the features along the spatial dimension. Each two-dimensional feature channel is compressed into a one-dimensional real number. This real number possesses a global receptive field to some extent, and the output dimension matches the number of input feature channels. This real number also represents the global distribution of the response across the feature channels, allowing layers closer to the input to also obtain a global receptive field. The calculation formula is as follows:

[0029]

[0030] In the formula, H and W are the height and width of the image, respectively. The formula for calculating · in formula (6) is as follows:

[0031]

[0032] In the formula, Vc represents the c-th convolution kernel; * represents the convolution operation; It is a two-dimensional spatial kernel, a single channel of Vc, which acts on the corresponding channel of X;

[0033] Then, an excitation operation is performed to generate weights for each feature channel using parameter w, where parameter w is used to learn and explicitly model the correlation of feature channels.

[0034] Finally, a reweighting operation is used, treating the output weights of the excitation as the importance of each feature channel after feature selection. Then, the channels are gradually weighted onto the previous features through multiplication, completing the recalibration of the original features in the channel dimension. The calculation formula is as follows:

[0035] F ex (·,W)=σ(g(·,W))=σ(W2δ(W1·)) (8)

[0036] In the formula, δ represents the ReLU function; σ represents the Sigmoid function; r is the reduction ratio; c is the number of image channels.

[0037] The SENet network described above is combined with the Dense Residual Block RDB network to form the SERRDB network. The structure of the Dense Residual Block RDB network is as follows:

[0038] The Dense Residual Block RDB network is composed of Residual Blocks and Dense Blocks. It extracts local features through densely connected convolutional layers, directly connects all layers of the current RDB from the front-end RDB, and then uses the local features of the RDB to adaptively learn features from the front and current local features.

[0039] The SERRDB network is composed of an RDB network and a SENet network. The RRDB network is formed by adding residual edges to three cascaded RDBs, and then adding the SENet network to the RDB network structure to form the SERDB network structure. The three cascaded RDB networks combined with the SERDB network form the SERRDB network.

[0040] Following Step 3 above, to evaluate the quality of the reconstructed image SR, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) are used as the evaluation criteria for image quality. The calculation formulas are as follows:

[0041]

[0042]

[0043] SSIM measures image similarity from three aspects: brightness, contrast, and structure. The values ​​range from 0 to 1, with higher values ​​indicating higher similarity. The specific formulas are as follows:

[0044]

[0045] In the formula, μ X and μ Y Let σ represent the mean values ​​of the low-resolution image X and the high-resolution image Y, respectively; X and σ Y Then, σ represents the variance of X and Y, respectively; XY C1 and C2 are the covariances of X and Y, respectively; C1 and C2 are constants.

[0046] The present invention provides a local super-resolution reconstruction algorithm based on infrared images of transmission line insulators, which has the following beneficial effects:

[0047] 1. By adopting a generative adversarial network structure, it achieves better reconstruction results compared to traditional reconstruction methods, and it is an unsupervised reconstruction.

[0048] 2. The RRDB network structure is adopted, which can maintain training stability, reduce computation and memory usage, and improve reconstruction results when deepening the network structure.

[0049] 3. Incorporate the SENet attention mechanism to enhance useful features (the outline and details of the insulator) based on their importance at the feature channel level, while suppressing features that are not very useful to this program.

[0050] 4. An image segmentation algorithm is added to separate the background from the foreground target (insulator), which not only reduces the workload of super-resolution reconstruction, but also avoids the influence of complex background on infrared images. Attached Figure Description

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0052] Figure 1 This is the image reconstruction process of the present invention;

[0053] Figure 2 This is a diagram of the adversarial network structure of the present invention;

[0054] Figure 3 Here is a diagram of the SENet network structure;

[0055] Figure 4Here is a diagram of the RDB network structure;

[0056] Figure 5 Here is a diagram of the SERRDB network structure;

[0057] Figure 6 This is a comparison chart of the actual effects in the embodiments. Detailed Implementation

[0058] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0059] like Figure 1 As shown, the local super-resolution reconstruction algorithm based on infrared images of transmission line insulators includes the following steps:

[0060] Step 1, Image Recognition: First, label the insulators in the infrared image, train the image recognition algorithm, and then input the dataset into the image recognition algorithm to recognize the insulators in the infrared image of the transmission line. The recognized image is marked with a recognition box, and the recognized image is used as the input for the subsequent steps.

[0061] Step 2, Image Segmentation: The recognized image is fed into the image segmentation network. The network will take the insulator that has been located in the recognition box as the target and the rest of the image as the background, and perform image segmentation to separate the foreground target from the background. The final result is that except for the insulator in the box, which is still an infrared image, all other backgrounds are changed to black.

[0062] Step 3, Super-resolution reconstruction: The segmented image is fed into an adversarial network for training to obtain the reconstructed super-resolution infrared image of the insulator.

[0063] like Figure 2 As shown in the diagram, the adversarial network structure in Step 3 above is as follows:

[0064] Random noise is input into the generator G for network training, and then data is generated. The data generated by the generator G and the real data are fed into the discriminator D. The discriminator D determines whether the input data is real data or data generated by the generator G, and outputs either 0 (generated data from the generator G) or 1 (real data). This output is fed into the generator G, and the generator G learns again based on the feedback. This process is repeated, and the generator G and the discriminator D learn from each other and compete with each other until a Nash equilibrium is reached.

[0065] The Nash equilibrium formula achieved by the aforementioned adversarial network is:

[0066]

[0067] In the formula, pdata(x) Based on real data distribution, p z(z) This refers to the data distribution generated by the generator; ultimately, the discriminator D cannot accurately determine whether the input data is real data or data generated by the generator G. That is, the discriminator D's accuracy in judging the input data is only 50%, and the generator G achieves the optimal effect.

[0068] Choosing MSE as the loss function, the calculation formula is as follows:

[0069]

[0070] In the formula, r represents the magnification factor, and W and H are the width and height of the image, respectively.

[0071] In Step 3 above, image super-resolution reconstruction can achieve a larger Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). However, a larger PSNR does not necessarily mean a better-looking reconstructed image. Therefore, an overall loss function L is adopted.

[0072] L = L per +λL adv +μL con (3)

[0073] In equation (3), L con Let L be the MSE loss function. per To perceive loss, L adv To counteract the loss, λ and μ are the corresponding weights. The perceptual loss is provided by a pre-trained VGG19 network, and its calculation formula is as follows:

[0074]

[0075] In the formula, Wij and Hij are the width and height of the feature map obtained by j convolutions before the i-th max pooling layer in the discrimination module, respectively; x and y correspond to the horizontal and vertical coordinates of a certain pixel in the feature map, respectively; φ ij It is the feature map obtained by j convolutions before the i-th max pooling layer in the discrimination module; HR It is a high-resolution image; I LR It is a low-resolution image; G is the generator network.

[0076] In equation (3) above, L adv The binary cross-entropy loss (BCEloss) function is used. In the training of the adversarial network, the output of the discriminator D is a binary classification of 0 or 1. A weighted BCEloss is used to constrain the overall loss function, and its calculation formula is as follows:

[0077] L adv=-(Dlog(D(G(z)))+(1-D)log(1-D(G(z)))) (5)

[0078] In the formula, D represents the label information of the real data; D(G(z)) represents the data information generated by the generator.

[0079] In the generator G mentioned above, an SENet network is embedded. The SENet network is used to reconstruct the details and edges of the insulator during the training of the generator G.

[0080] The SENet network aims to learn from the interdependencies between feature channels of an image. It obtains the weight of each feature channel through a self-learning process, and then uses these weights to enhance useful features and suppress features that are less important to the current target.

[0081] like Figure 3 As shown, the SENet network structure is as follows:

[0082] Given an input X with C1 feature channels, after passing it through a series of convolutions and general transformations, the number of feature channels becomes C2. First, a Squeeze operation is performed on the input, compressing the features along the spatial dimension. Each two-dimensional feature channel is compressed into a one-dimensional real number. This real number possesses a global receptive field to some extent, and the output dimension matches the number of feature channels in the input. This real number also represents the global distribution of the response across the feature channels, allowing layers closer to the input to also obtain a global receptive field. The calculation formula is as follows:

[0083]

[0084] In the formula, H and W are the height and width of the image, respectively. The formula for calculating · in formula (6) is as follows:

[0085]

[0086] In the formula, Vc represents the c-th convolution kernel; * represents the convolution operation; It is a two-dimensional spatial kernel, a single channel of Vc, which acts on the corresponding channel of X;

[0087] Then, the Excitation operation is performed, which is a mechanism similar to a gate in a loop. Weights are generated for each feature channel through the parameter w, where the parameter w is used to learn to explicitly model the correlation of feature channels.

[0088] Finally, a reweighting operation is used, treating the output weights of the excitation as the importance of each feature channel after feature selection. Then, the channels are gradually weighted onto the previous features through multiplication, completing the recalibration of the original features in the channel dimension. The calculation formula is as follows:

[0089] F ex (·,W)=σ(g(·,W))=σ(W2δ(W1·)) (8)

[0090] In the formula, δ represents the ReLU function; σ represents the Sigmoid function; r is the reduction ratio; c is the number of image channels.

[0091] The SENet network described above is combined with the Dense Residual Block RDB network to form the SERRDB network, as follows: Figure 4 As shown, the dense residual block RDB network structure is as follows:

[0092] The Dense Residual Block RDB network is composed of Residual Blocks and Dense Blocks. It extracts local features through densely connected convolutional layers, directly connects all layers of the current RDB from the front-end RDB, and then uses the local features of the RDB to adaptively learn features from the front and current local features, making the training more stable and achieving better results.

[0093] Compared to traditional Dense Block, this network removes the Batch Normalization (BN) layer. BN normalizes features using the batch mean and variance. However, in deep networks and GAN training, BN often introduces artifacts, affecting image generation quality. Therefore, removing the BN layer not only improves the network's generalization ability but also reduces computational cost and memory usage.

[0094] like Figure 5 As shown, the SERRDB network is composed of an RDB network and an SENet network. The SERRDB network is formed by adding residual edges to three cascaded RDBs and then incorporating the SENet network into the RDB network structure. This network not only has the advantages of RDB such as generalization ability and low computational cost, but also combines the attention mechanism of SENet. While deepening the network structure, it can not only stabilize training, but also better extract and reconstruct the main features of the image during training.

[0095] The three cascaded RDB networks combine with the SERDB network to form a SERRDB network;

[0096] SERRDB combines multiple dense residual blocks with SENet to deepen the network and improve reconstruction results. By incorporating the SERRDB network for image segmentation into the code, the interference of the image background on the main target can be minimized during super-resolution reconstruction of infrared images. After generating a high-resolution image, there will be a clear difference between the target and the background, so as not to affect the detection by the inspector.

[0097] Following Step 3 above, to evaluate the quality of the reconstructed image SR, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) are used as the evaluation criteria for image quality. The calculation formulas are as follows:

[0098]

[0099]

[0100] SSIM measures image similarity from three aspects: brightness, contrast, and structure. The values ​​range from 0 to 1, with higher values ​​indicating higher similarity. The specific formulas are as follows:

[0101]

[0102] In the formula, μ X and μ Y Let σ represent the mean values ​​of the low-resolution image X and the high-resolution image Y, respectively; X and σ Y Then, σ represents the variance of X and Y, respectively; XY C1 and C2 are the covariances of X and Y, respectively; C1 and C2 are constants.

[0103] Example:

[0104] The system was Windows 10, the graphics card was RTX 3090, and the deep learning frameworks were PyTorch 1.10.0, Python 3.8, and CUDA 11.3. The comparison algorithms were bicubic interpolation (BICUBIC) and SRGAN (Super Resolution Generative Adversarial Networks). The peak signal-to-noise ratio and structural similarity values ​​obtained by SRGAN and the test images of this invention are shown in the table below:

[0105] PSNR SSIM SRGAN 24.2698 0.7821 This invention 24.2715 0.7922

[0106] Reconstruction results as follows Figure 6 As shown, (a) is the original low-resolution image LR; (b) is the reconstruction result of the BICUBIC algorithm; (c) is the reconstruction result of the SRGAN algorithm; and (d) is the reconstruction result of the algorithm of this invention. The PSNR and SSIM results of this invention on the test image are superior to those of the SRGAN algorithm.

[0107] In this invention, the number of transmission line insulators is huge and the infrared image resolution is low, resulting in a large workload, low efficiency, and easy false detection and missed detection during inspection. Therefore, it is proposed to perform super-resolution reconstruction of the infrared images of transmission line insulators to reduce the workload of inspection.

[0108] The SERRDB network structure was proposed. SERRDB combines the residual learning of RRDB with the feature extraction method of network depth and SENet attention mechanism, which improves the image reconstruction effect and has a significant improvement in super-resolution reconstruction of infrared images.

[0109] An image segmentation algorithm is introduced into image reconstruction to separate the background and foreground targets in the infrared image. Only the insulator part of the transmission line is reconstructed in super-resolution, which reduces background interference and improves the detection efficiency and accuracy of the inspectors.

Claims

1. A local super-resolution reconstruction algorithm based on infrared images of insulators of a power transmission line, characterized in that, The method comprises the following steps: Step 1, image recognition: first, label the insulators in the infrared image, train the image recognition algorithm, input the data set into the image recognition algorithm after training, recognize the insulators in the transmission line infrared image, identify the insulators in the image using an identification box, and input the recognized image into the subsequent step; Step 2, image segmentation: the image after recognition is sent to the image segmentation network, the network takes the insulator located in the identification box as the target, and the other part of the image as the background, and separates the foreground target and the background, finally, the effect is that the insulator in the box is still the infrared image, and the other background is black; Step 3, super-resolution reconstruction: the segmented image is sent to the adversarial network for training to obtain the reconstructed insulator super-resolution infrared image; The adversarial network structure in Step 3 is: Random noise enters the input end of the generator G for network training, then generates data, and the data generated by the generator G and the real data are sent into the discriminator D to judge whether the input data is real data or data generated by the generator G, and give an output of 0, that is, the generated data from the generator G or 1, that is, the real data, the output is sent into the generator G, and the generator G learns again according to the feedback obtained, so as to repeatedly learn and mutually oppose each other, and finally reach a Nash equilibrium; The Nash equilibrium formula reached by the adversarial network is: ；（1） In the formula, from a real data distribution, is a data distribution generated by the generator; finally, the discriminator D cannot accurately determine whether the input data is real data or data generated by the generator G, that is, the accuracy of the judgment of the discriminator D on the input data is only 50%, and the generator G reaches the optimal effect; MSE is selected as the loss function, and the calculation formula is: ；（2） In the formula, r represents the magnification, W and H are the width and height of the image respectively; In Step 3, the image super-resolution reconstruction can obtain a larger peak signal-to-noise ratio PSNR and structural similarity SSIM, but a larger peak signal-to-noise ratio PSNR does not mean a better reconstructed image in the sense, therefore, the overall loss function L is adopted: ；（3） In formula (3), Lcon is a MSE loss function, Lper is a perceptual loss, L adv is an adversarial loss, λ and μ are corresponding weights, the perceptual loss is provided by a pre-trained VGG19 network, and a calculation formula thereof is as follows: ；（4） In the formula, Wij and Hij are the width and height of the feature map obtained by j times of convolution before the ith max-pooling layer in the discrimination module; and x and y correspond to the horizontal coordinate and vertical coordinate of a certain pixel point in the feature map, respectively. is the feature map obtained by j times of convolution before the ith max-pooling layer in the discrimination module, I HR is a high-resolution image; I LR is a low-resolution image; G is a generator network; In the generator G, the SE Net network is embedded, which makes the generator G reconstruct the details and edge parts of the insulator during training, and the SE Net network structure is: The input end of the feature channel C1 is given X, which enters the feature channel C2 through multi-layer convolution and transformation, first performs the Squeeze operation, performs feature compression along the spatial dimension, compresses each two-dimensional feature channel into a one-dimensional real number, the real number has a global receptive field to a certain extent, and the output dimension and the input feature channel number are matched, the real number also represents the global distribution of the response on the feature channel, so that the layers close to the input can also obtain the global receptive field, and the calculation formula is as follows: ； （6） where H and W are the height and width of the image, respectively, and the formula for • in equation (6) is as follows: ；（7） In the formula, represents the cth convolution kernel; represents a convolution operation; is a two-dimensional spatial kernel, and is a single channel of X corresponding channel; Then, the Excitation operation is performed, and the parameter w is used to generate the weight for each feature channel, wherein the parameter w is used to learn and model the correlation of the feature channels; Finally, the output weight of the Excitation is regarded as the importance of each feature channel after feature selection, and then the channels are gradually weighted to the previous features through multiplication to complete the reweighting of the original features in the channel dimension, and the calculation formula is as follows: ；（8） wherein, delta denotes a ReLU function; σ denotes a Sigmoid function; , r is a reduction ratio; c is the number of image channels; The SENet network is combined with the dense residual block RDB network into the SERRDB network, and the structure of the dense residual block RDB network is as follows: The dense residual block RDB network is combined into RDB by the residual block Residual Block and the dense block Dense Block, local features are extracted through dense connected convolution layers, all layers of the front RDB are directly connected to the current RDB, and then the local features of the RDB are used to adaptively learn features from the front and current local features; The SERRDB network is combined by the RDB network and the SENet network, wherein the RRDB network is added with a residual side on the basis of cascading three RDBs, then the SENet network is added into the RDB network structure to form the SERDB network structure, and the three cascaded RDB networks combine the SERDB network to form the SERRDB network.

2. The partial super-resolution reconstruction algorithm based on the infrared images of power line insulators according to claim 1, characterized in that, In the formula (3), L The binary cross-entropy loss function BCEloss is adopted by the adv. In the training of the adversarial network, the output of the discriminator D is a binary classification 0 or 1, and the overall loss function is constrained by adopting a certain weight of the BCEloss. The calculation formula is as follows: ；（5） In the formula, D label information representing real data; D (G(z)) represents data information generated by the generator.

3. The partial super-resolution reconstruction algorithm based on the infrared images of power line insulators according to claim 2, characterized in that, After Step3, in order to evaluate the quality of the reconstructed image SR, the peak signal-to-noise ratio PSNR and the structural similarity SSIM are used as the evaluation standard of the image quality, and the calculation formula is as follows: ；（9） ； （10） The SSIM measures the image similarity from three aspects of brightness, contrast and structure, and the value range is 0~1, the larger the value, the higher the similarity, and the specific formula is as follows: ；（11） wherein and denote the mean of the low resolution image X and the high resolution image Y respectively; and denote the variance of the low resolution image X and the high resolution image Y respectively; is the covariance of the low resolution image X and the high resolution image Y respectively; C 1 and C 2 are constants.

Images