Transmission line inspection image super-resolution reconstruction method and system based on SA-GAN

Abstract

The application discloses a kind of transmission line inspection image super-resolution reconstruction method and system based on SA-GAN, the method includes: obtaining original inspection image and conducting degradation processing to construct data set;SA-GAN network containing generator and discriminator is constructed, wherein the generator is successively integrated input layer, residual layer, spatial adaptive multi-scale feature modulation module, up-sampling layer, global perception attention module and local-global fusion channel attention module;The network is trained using the data set, and a multi-scale attention feature matching loss function is used for optimization;Finally, the low-resolution inspection image is input into the trained model, and a high-quality super-resolution image is output.The application significantly improves the image detail recovery capability and visual authenticity through spatial adaptive multi-scale fusion and double attention mechanism, effectively solves the defect detection difficulty caused by blurred and low-resolution unmanned aerial vehicle inspection image, and provides reliable technical support for intelligent inspection of power equipment.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and intelligent inspection technology of power equipment, and particularly relates to a super-resolution reconstruction method and system for transmission line inspection images based on SA-GAN. Background Technology

[0002] Insulators are critical components in power transmission lines, and their integrity directly affects the safe and stable operation of the power grid. Traditional manual inspection methods are inefficient, risky, and prone to missing components. With the development of drone technology, drone inspection equipped with visible light or infrared imaging equipment has become the mainstream method. However, during the inspection process, due to factors such as drone flight vibration, changes in lighting conditions, and long shooting distance, the images of insulators and other equipment often suffer from problems such as blurriness, low resolution, and loss of detail. Low-quality images seriously affect the accuracy and reliability of subsequent deep learning-based target detection or defect recognition algorithms.

[0003] To improve image quality, super-resolution reconstruction technology has been introduced into this field. Existing techniques typically employ super-resolution models based on convolutional neural networks. However, when faced with complex power scenes (such as cluttered backgrounds, small target sizes, and weak textures), these methods often result in overly smoothed details, texture distortion, or the introduction of artifacts, making it difficult to meet the requirements of high-precision defect detection. While generative adversarial networks (GANs) can generate more realistic textures, standard GAN models still have shortcomings in reconstructing high-frequency details and maintaining global structural consistency, and their training is unstable, easily leading to model collapse.

[0004] Therefore, existing technologies lack a dedicated super-resolution reconstruction method that can accurately restore the fine textures and structural details of equipment such as insulators while maintaining high perception quality, taking into account the characteristics of transmission line inspection images. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a super-resolution reconstruction method and system for transmission line inspection images based on SA-GAN, which aims to significantly improve the visual quality and detail clarity of low-resolution inspection images, and provide a high-quality image foundation for subsequent intelligent defect detection.

[0006] In a first aspect, the present invention provides a super-resolution reconstruction method for transmission line inspection images based on SA-GAN, comprising:

[0007] The original images of power line inspections collected by drones are obtained, and bicubic interpolation is performed to reduce the resolution to obtain low-resolution images. A dataset containing low-resolution images and the original high-resolution images is constructed and divided into training set and validation set.

[0008] An SA-GAN super-resolution reconstruction network is constructed, and the network is iteratively trained and validated using the training set and validation set to obtain a super-resolution reconstruction model. The SA-GAN network includes a generator, which sequentially includes an input layer, a residual layer, a spatial adaptive multi-scale feature modulation module, an upsampling layer, a global perception attention mechanism module, a channel attention mechanism module, and an output layer.

[0009] The spatially adaptive multi-scale feature modulation module enhances the input features through channel grouping, multi-scale convolution processing, spatially adaptive scale weight calculation and weighted fusion;

[0010] The global perception attention mechanism module enhances the perception of the overall structure of the image through global channel attention and deep convolutional spatial attention mechanisms;

[0011] The channel attention mechanism module integrates local convolutional details and global channel attention weighted features to optimize feature representation;

[0012] The low-resolution transmission line inspection image to be reconstructed is input into the trained super-resolution reconstruction model, and the corresponding super-resolution image is output.

[0013] Secondly, the present invention provides a super-resolution reconstruction system for transmission line inspection images based on SA-GAN, comprising:

[0014] The acquisition module is configured to acquire the original images of power line inspection collected by the UAV, perform bicubic interpolation to reduce the resolution, obtain low-resolution images, and construct a dataset containing low-resolution images and the original high-resolution images, which is divided into training set and validation set.

[0015] The construction module is configured to build an SA-GAN super-resolution reconstruction network. The network is iteratively trained and validated using the training set and validation set to obtain a super-resolution reconstruction model. The SA-GAN network includes a generator, which sequentially includes an input layer, a residual layer, a spatial adaptive multi-scale feature modulation module, an upsampling layer, a global perception attention mechanism module, a channel attention mechanism module, and an output layer.

[0016] The spatially adaptive multi-scale feature modulation module enhances the input features through channel grouping, multi-scale convolution processing, spatially adaptive scale weight calculation and weighted fusion;

[0017] The global perception attention mechanism module enhances the perception of the overall structure of the image through global channel attention and deep convolutional spatial attention mechanisms;

[0018] The channel attention mechanism module integrates local convolutional details and global channel attention weighted features to optimize feature representation;

[0019] The output module is configured to input the low-resolution transmission line inspection image to be reconstructed into the trained super-resolution reconstruction model and output the corresponding super-resolution image.

[0020] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the SA-GAN-based super-resolution reconstruction method for transmission line inspection images according to any embodiment of the present invention.

[0021] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the super-resolution reconstruction method for transmission line inspection images based on SA-GAN according to any embodiment of the present invention.

[0022] The method in this application, by introducing a spatially adaptive multi-scale feature modulation module, achieves a pixel-level, content-aware multi-scale feature fusion mechanism. This allows the network to dynamically adjust the fusion weights of features at different scales based on the complexity of local image regions (such as edges, textures, or flat backgrounds), thereby specifically enhancing the recovery ability of key high-frequency details such as fine cracks and dirt textures in insulators. Secondly, the collaborative design of the global awareness attention mechanism module and the channel attention mechanism module strengthens feature expression from the channel and spatial dimensions, and the local detail and global semantic levels, respectively, effectively improving the overall structural consistency and local realism of the reconstructed image, and suppressing background noise and artifacts. Furthermore, it is specifically designed for the network... The multi-scale attention feature matching loss function designed for network training significantly improves the perceptual quality of generated images by forcing the generator to learn the feature importance distribution of real images while imitating feature content. This makes the reconstructed results visually more natural and the details more believable. Finally, the complete technical solution formed by this method can reconstruct low-resolution images caused by motion blur, insufficient lighting, or long shooting distance in UAV inspections into clear, detailed, high-resolution images. This provides reliable, high-quality input for subsequent advanced vision tasks such as automatic insulator defect detection and component identification, comprehensively improving the accuracy and reliability of intelligent power inspection while maintaining high processing efficiency. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart of a super-resolution reconstruction method for transmission line inspection images based on SA-GAN provided in an embodiment of the present invention;

[0025] Figure 2 This is a structural block diagram of a super-resolution reconstruction system for transmission line inspection images based on SA-GAN, provided in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1 The diagram shows a flowchart of a super-resolution reconstruction method for transmission line inspection images based on SA-GAN, as described in this application.

[0029] like Figure 1 As shown, the super-resolution reconstruction method for transmission line inspection images based on SA-GAN specifically includes the following steps:

[0030] Step S101: Obtain the original images of the power transmission line inspection collected by the UAV, perform bicubic interpolation to reduce the resolution, obtain low-resolution images, and construct a dataset containing low-resolution images and the original high-resolution images, which is divided into training set and validation set.

[0031] In this step, during the data preparation phase, a large number of clear images of transmission line insulators and other equipment taken by drones are collected as the original high-resolution image set. To simulate actual low-quality images, a bicubic interpolation algorithm is used to downsample each original image to 1 / 4 of its original size, resulting in corresponding low-resolution images, forming an image pair dataset. This dataset is then randomly divided into training and validation sets at an 8:2 ratio.

[0032] Step S102: Construct an SA-GAN super-resolution reconstruction network. Iteratively train and validate the network using the training set and validation set to obtain a super-resolution reconstruction model. The SA-GAN network includes a generator, which sequentially includes an input layer, a residual layer, a spatial adaptive multi-scale feature modulation module, an upsampling layer, a global perception attention mechanism module, a channel attention mechanism module, and an output layer.

[0033] In this step, the spatially adaptive multi-scale feature modulation module enhances the input features through channel grouping, multi-scale convolution processing, spatially adaptive scale weight calculation, and weighted fusion.

[0034] The specific steps of the spatial adaptive multi-scale feature modulation module in processing the input feature map include:

[0035] The input feature map with C channels is evenly divided into N groups;

[0036] Perform independent 3x3 convolution operations on each group to generate N feature maps of different scales, as shown in the expression:

[0037] ,

[0038] In the formula, For the first Output feature maps of each scale group For batch indexing, The input channel index variable cycles through 0 to 15. This represents the offset of the convolution kernel in the x-direction. This represents the offset of the convolution kernel in the y-direction. For the first Convolutional kernel weights for each scale group For the first The output feature map of each scale group after passing through 16 residual blocks For the first Convolutional bias for each scale group;

[0039] The adaptive selection weights at N scales for each spatial location are calculated using a 1×1 convolutional layer and a softmax function, expressed as follows:

[0040] ,

[0041] ,

[0042] In the formula, For normalized scale weights, To calculate the exponent for the unnormalized scale weight score, The weight scale index is (0,1,2,3). For unnormalized scale weight scores, This is the input channel for calculating the scale selector weights. The 1×1 convolution weights are for the scale selector. This is the output feature map after 16 residual blocks. For the bias term of the scale selector;

[0043] The feature maps of N different scale groups are weighted and summed according to the adaptive selection of weights to obtain the fused feature map, expressed as:

[0044] ,

[0045] In the formula, The fused feature map To normalize the weights, For the first Output feature maps of each scale group;

[0046] Calculate the channel attention weight map of the fused feature map and multiply it element-wise with the fused feature map to obtain the final output.

[0047] The global perception attention mechanism module enhances the perception of the overall structure of the image through global channel attention and deep convolutional spatial attention mechanisms.

[0048] The processing of the global awareness attention mechanism module includes:

[0049] Global average pooling is performed on the input feature map, and global channel attention weights are generated through a first fully connected layer and a second fully connected layer. These weights are then multiplied by the input feature map to obtain the first target feature map. The expression for the first fully connected layer is:

[0050] ,

[0051] In the formula, This is the output of the first fully connected layer. It is the ReLU activation function. This is the input channel index for the first fully connected layer. This is the weight matrix of the first fully connected layer. As input to the global average pooling result, This is the bias term for the first fully connected layer;

[0052] The expression for the second fully connected layer is:

[0053] ,

[0054] In the formula, For global channel attention weights, It is the Sigmoid activation function. This is the input channel index for the second fully connected layer. This is the weight matrix for the second fully connected layer. This is the output of the first fully connected layer. For the bias term of the second fully connected layer, This is the weight matrix for the second fully connected layer;

[0055] The expression for the first target feature map is:

[0056] ,

[0057] In the formula, The feature map after applying channel attention. The original input feature map, It is the element-wise multiplication operator. For global channel attention weights;

[0058] Perform a depthwise convolution on the obtained first target feature map, generate a spatial attention weight map using the Sigmoid function, multiply it again by the first target feature map, and output the result, expressed as:

[0059] ,

[0060] ,

[0061] In the formula, This is the final output of the GPA module. The feature map after channel attention processing; Spatial attention weights, For the Sigmoid function, This represents the offset of the convolution kernel in the x-direction. This represents the offset of the convolution kernel in the y-direction. The weights of the depthwise convolution. This is the bias term for depthwise convolution;

[0062] The channel attention mechanism module integrates local convolutional details and global channel attention weighted features to optimize feature representation.

[0063] The input feature map is subjected to a 3×3 convolution to extract local features, expressed as follows:

[0064] ,

[0065] In the formula, For local convolutional feature output, For the global fusion attention mechanism channel index, This represents the offset of the convolution kernel in the x-direction. This represents the offset of the convolution kernel in the y-direction. This is the output feature map of the GPA module. This is the bias term for local convolution;

[0066] Global average pooling is performed on the input feature map, and global channel attention weights are generated through two fully connected layers. These weights are then multiplied by the input feature map to obtain the globally weighted features. The expression for the global channel attention weights is as follows:

[0067] ,

[0068] ,

[0069] In the formula, For LGCAM channel attention weights, For the Sigmoid function, This is the input channel index for the second fully connected layer of LGCAM. For the weights of the second fully connected layer, For the bias of the second fully connected layer, For the output of the first fully connected layer, It is the ReLU activation function. This is the input channel index for the first fully connected layer in LGCAM. For the weights of the first fully connected layer, This is the global pooling result of LGCAM. Bias for the first fully connected layer;

[0070] The local features are added to the global weighted features to obtain the final output, expressed as:

[0071] ,

[0072] In the formula, This is the final output of the LGCAM module. Preserving spatial details of local features for local convolutional features. Output features for the GPA module. Features that are globally semantically weighted for global channel attention weights.

[0073] It should be noted that the generator loss function of the SA-GAN network is a multi-scale attention feature matching loss, expressed as:

[0074] ,

[0075] ,

[0076] ,

[0077] In the formula, For multi-scale attention feature matching loss function, This represents the total number of network depth levels. This represents the total number of spatial scale levels. For L1 norm operators, For multi-scale attention weight functions, For multi-scale feature extraction functions, It is the element-wise multiplication operator. The super-resolution output image generated by the generator. For a true high-resolution reference image, To separately and Perform feature extraction. For network depth level indexing, For spatial scale level indexing, It is a 3x3 two-dimensional convolution. The channel grouping function uniformly divides the channel dimensions of the input features into S groups, and then takes the s-th group. For residual block functions, It is the product of the feature map height and the feature map width. For height indexing, Indexed in the width direction. Represented as a slice of the feature map. The Softmax normalization function transforms the input vector into a probability distribution. This is the transposed feature map of the original feature map. Step S103: Input the low-resolution transmission line inspection image to be reconstructed into the trained super-resolution reconstruction model, and output the corresponding super-resolution image.

[0078] In one specific embodiment, the generator of the SA-GAN network is the core. The input low-resolution image (e.g., 640x640) first passes through a 9x9 convolutional layer (input layer), increasing the number of channels to 64, and uses the Swish activation function. Subsequently, the feature map enters a residual layer consisting of 16 stacked Enhanced ResBlocks. Each Enhanced ResBlock contains two 3x3 convolutional layers, with a Channel Attention Module (CAM) and a Spatial Attention Module (SAM) embedded in between, and residual connections are made through a learnable scaling parameter. This effectively promotes gradient flow and enhances feature selection capabilities.

[0079] After deep feature extraction via the residual layer, the feature map enters the crucial spatial adaptive multi-scale feature modulation module. This module first divides the input feature map (C=64 channels) into four groups (N=4) along the channel dimension. Each group is processed by an independent 3x3 convolutional layer to extract features at different scales. Then, a 1x1 convolutional layer is applied to the original input features, calculating four unnormalized scores for each spatial location (i,j). Next, a Softmax function normalizes these scores along the scale dimension, obtaining four adaptive weights corresponding to that location, with a sum of 1. These weights are then used to perform pixel-level weighted summation of the feature maps across the four scale groups, yielding a fused feature. Finally, channel attention modulation is applied to this fused feature to generate the final output of the spatial adaptive multi-scale feature modulation module. This process achieves content-aware feature fusion. For complex textures and edge regions of insulator strings, the network may assign higher weights to mid-to-high frequency scale features to recover details; for flat background regions such as the sky, it may rely more on low-frequency contextual information.

[0080] Next, the feature map is upsampled through two convolutional and pixel-shuffle operations to gradually enlarge its spatial size by a factor of 4 to the target high-resolution size. The upsampled feature map is then refined by a global awareness attention mechanism module and a channel attention mechanism module.

[0081] The global awareness attention mechanism module first performs global average pooling on the input feature map to obtain a global descriptor. Then, it generates channel attention weights through two fully connected layers (with ReLU activation in between) and multiplies them with the input features to recalibrate the channel dimension. Next, it performs a depthwise convolution (3x3 kernel, number of groups equal to number of channels) on the channel-enhanced features, and then generates a spatial attention map through the sigmoid function, which is multiplied again with the feature map to achieve spatial dimension focusing. This helps the network better grasp the overall shape and position of targets such as insulators during reconstruction.

[0082] The channel attention mechanism module processes two branches in parallel. The local branch directly extracts fine spatial details through a 3x3 convolutional layer. The global branch first performs global average pooling on the input, then generates channel attention weights through two fully connected layers (containing ReLU and Sigmoid), multiplying these weights with the input features to obtain semantically weighted global features. Finally, the spatial details are added to the global features. This ensures that the reconstructed image retains both clear local textures (such as cracks in the insulator skirt) and maintains consistent tone and lighting with the overall scene.

[0083] The generator ends with a 9x9 convolutional output layer that reduces the number of channels from 64 to 3 (RGB) and uses the Tanh activation function to map pixel values ​​to the [-1, 1] range to obtain the final super-resolution image.

[0084] The discriminator adopts the classic PatchGAN structure to determine the authenticity of local patches in the input image. Its output feature map will be used to calculate the adversarial loss and the multi-scale attention feature matching loss proposed in this invention.

[0085] During the training phase, the generator and discriminator are optimized alternately. The generator's total loss typically includes: adversarial loss (to encourage generated images to fool the discriminator), content loss (such as L1 or perceptual loss), and multi-scale attention feature matching loss.

[0086] Multi-scale attention feature matching loss enables the generator to learn the importance distribution of different features in real images while learning the distribution of image content, thereby generating visually more natural and more believable images.

[0087] During training, the Adam optimizer can be used, with an initial learning rate of 1e-4 and a batch size determined by the available GPU memory. Training should last approximately 200 epochs until the model converges. The best-performing model weights from the validation set should then be saved.

[0088] In the application phase, simply load the trained generator and input the low-resolution inspection image to be processed, and the reconstructed high-resolution image can be quickly output. The processing speed can meet the requirements of real-time or near-real-time inspection.

[0089] In summary, the method of this application, by introducing a spatially adaptive multi-scale feature modulation module, achieves a pixel-level, content-aware multi-scale feature fusion mechanism. This enables the network to dynamically adjust the fusion weights of features at different scales based on the complexity of local image regions (such as edges, textures, or flat backgrounds), thereby specifically enhancing the recovery capability of key high-frequency details such as insulator microcracks and dirt textures. Secondly, the collaborative design of the global awareness attention mechanism module and the channel attention mechanism module strengthens feature expression from the channel and spatial dimensions, and the local detail and global semantic levels, respectively, effectively improving the overall structural consistency and local realism of the reconstructed image, and suppressing background noise and artifacts. Furthermore, it is specifically designed for... The multi-scale attention feature matching loss function designed for network training significantly improves the perceptual quality of generated images by forcing the generator to learn the feature importance distribution of real images while imitating feature content. This makes the reconstruction results more visually natural and the details more believable. Finally, the complete technical solution formed by this method can reconstruct low-resolution images caused by motion blur, insufficient lighting, or long shooting distance in UAV inspections into clear, detailed, high-resolution images. This provides reliable, high-quality input for subsequent advanced vision tasks such as automatic insulator defect detection and component identification, and comprehensively improves the accuracy and reliability of intelligent power inspection while maintaining high processing efficiency.

[0090] Please see Figure 2 The diagram shows a structural block diagram of a super-resolution reconstruction system for transmission line inspection images based on SA-GAN, as described in this application.

[0091] like Figure 2 As shown, the transmission line inspection image super-resolution reconstruction system 200 includes an acquisition module 210, a construction module 220, and an output module 230.

[0092] The acquisition module 210 is configured to acquire the original images of power line inspection collected by the UAV, perform bicubic interpolation to reduce the resolution, obtain low-resolution images, and construct a dataset containing low-resolution images and original high-resolution images, which is divided into training set and validation set.

[0093] Module 220 is configured to construct an SA-GAN super-resolution reconstruction network. The network is iteratively trained and validated using the training set and validation set to obtain a super-resolution reconstruction model. The SA-GAN network includes a generator, which sequentially includes an input layer, a residual layer, a spatial adaptive multi-scale feature modulation module, an upsampling layer, a global perception attention mechanism module, a channel attention mechanism module, and an output layer.

[0094] The spatially adaptive multi-scale feature modulation module enhances the input features through channel grouping, multi-scale convolution processing, spatially adaptive scale weight calculation and weighted fusion;

[0095] The global perception attention mechanism module enhances the perception of the overall structure of the image through global channel attention and deep convolutional spatial attention mechanisms;

[0096] The channel attention mechanism module integrates local convolutional details and global channel attention weighted features to optimize feature representation;

[0097] The output module 230 is configured to input the low-resolution transmission line inspection image to be reconstructed into the trained super-resolution reconstruction model and output the corresponding super-resolution image.

[0098] It should be understood that Figure 2 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 2 The various modules in the document will not be described in detail here.

[0099] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the super-resolution reconstruction method for transmission line inspection images based on SA-GAN in any of the above method embodiments.

[0100] In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows:

[0101] The original images of power line inspections collected by drones are obtained, and bicubic interpolation is performed to reduce the resolution to obtain low-resolution images. A dataset containing low-resolution images and the original high-resolution images is constructed and divided into training set and validation set.

[0102] An SA-GAN super-resolution reconstruction network is constructed, and the network is iteratively trained and validated using the training set and validation set to obtain a super-resolution reconstruction model. The SA-GAN network includes a generator, which sequentially includes an input layer, a residual layer, a spatial adaptive multi-scale feature modulation module, an upsampling layer, a global perception attention mechanism module, a channel attention mechanism module, and an output layer.

[0103] The spatially adaptive multi-scale feature modulation module enhances the input features through channel grouping, multi-scale convolution processing, spatially adaptive scale weight calculation and weighted fusion;

[0104] The global perception attention mechanism module enhances the perception of the overall structure of the image through global channel attention and deep convolutional spatial attention mechanisms;

[0105] The channel attention mechanism module integrates local convolutional details and global channel attention weighted features to optimize feature representation;

[0106] The low-resolution transmission line inspection image to be reconstructed is input into the trained super-resolution reconstruction model, and the corresponding super-resolution image is output.

[0107] Computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of the SA-GAN-based super-resolution reconstruction system for transmission line inspection images, etc. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely configured relative to a processor, which can be connected to the SA-GAN-based super-resolution reconstruction system for transmission line inspection images via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0108] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 3 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 3 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the SA-GAN-based super-resolution reconstruction method for transmission line inspection images described in the above embodiment. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the SA-GAN-based super-resolution reconstruction system for transmission line inspection images. The output device 340 may include a display screen or other display device.

[0109] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0110] In one implementation, the above-described electronic device is applied to a super-resolution reconstruction system for transmission line inspection images based on SA-GAN, serving as a client, and includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0111] The original images of power line inspections collected by drones are obtained, and bicubic interpolation is performed to reduce the resolution to obtain low-resolution images. A dataset containing low-resolution images and the original high-resolution images is constructed and divided into training set and validation set.

[0112] An SA-GAN super-resolution reconstruction network is constructed, and the network is iteratively trained and validated using the training set and validation set to obtain a super-resolution reconstruction model. The SA-GAN network includes a generator, which sequentially includes an input layer, a residual layer, a spatial adaptive multi-scale feature modulation module, an upsampling layer, a global perception attention mechanism module, a channel attention mechanism module, and an output layer.

[0113] The spatially adaptive multi-scale feature modulation module enhances the input features through channel grouping, multi-scale convolution processing, spatially adaptive scale weight calculation and weighted fusion;

[0114] The global perception attention mechanism module enhances the perception of the overall structure of the image through global channel attention and deep convolutional spatial attention mechanisms;

[0115] The channel attention mechanism module integrates local convolutional details and global channel attention weighted features to optimize feature representation;

[0116] The low-resolution transmission line inspection image to be reconstructed is input into the trained super-resolution reconstruction model, and the corresponding super-resolution image is output.

[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A super-resolution reconstruction method for transmission line inspection images based on SA-GAN, characterized in that, include: The original images of power line inspections collected by drones are obtained, and bicubic interpolation is performed to reduce the resolution to obtain low-resolution images. A dataset containing low-resolution images and the original high-resolution images is constructed and divided into training set and validation set. An SA-GAN super-resolution reconstruction network is constructed, and the network is iteratively trained and validated using the training set and validation set to obtain a super-resolution reconstruction model. The SA-GAN network includes a generator, which sequentially includes an input layer, a residual layer, a spatial adaptive multi-scale feature modulation module, an upsampling layer, a global perception attention mechanism module, a channel attention mechanism module, and an output layer. The spatially adaptive multi-scale feature modulation module enhances the input features through channel grouping, multi-scale convolution processing, spatially adaptive scale weight calculation and weighted fusion; The global perception attention mechanism module enhances the perception of the overall structure of the image through global channel attention and deep convolutional spatial attention mechanisms; The channel attention mechanism module integrates local convolutional details and global channel attention weighted features to optimize feature representation; The low-resolution transmission line inspection image to be reconstructed is input into the trained super-resolution reconstruction model, and the corresponding super-resolution image is output.

2. The method for super-resolution reconstruction of transmission line inspection images based on SA-GAN according to claim 1, characterized in that, The steps of the spatial adaptive multi-scale feature modulation module in processing the input feature map specifically include: The input feature map with C channels is evenly divided into N groups; Perform independent 3x3 convolution operations on each group to generate N feature maps of different scales, as shown in the expression: ， In the formula, For the first Output feature maps of each scale group For batch indexing, The input channel index variable cycles through 0 to 15. This represents the offset of the convolution kernel in the x-direction. This represents the offset of the convolution kernel in the y-direction. For the first Convolutional kernel weights for each scale group For the first The output feature map of each scale group after passing through 16 residual blocks For the first Convolutional bias for each scale group; The adaptive selection weights at N scales for each spatial location are calculated using a 1×1 convolutional layer and a softmax function, expressed as follows: ， ， In the formula, For normalized scale weights, To calculate the exponent for the unnormalized scale weight score, The weight scale index is (0,1,2,3). For unnormalized scale weight scores, This is the input channel for calculating the scale selector weights. The 1×1 convolution weights are for the scale selector. This is the output feature map after 16 residual blocks. For the bias term of the scale selector; The feature maps of N different scale groups are weighted and summed according to the adaptive selection of weights to obtain the fused feature map, expressed as: ， In the formula, The fused feature map To normalize the weights, For the first Output feature maps of each scale group; Calculate the channel attention weight map of the fused feature map and multiply it element-wise with the fused feature map to obtain the final output.

3. The method for super-resolution reconstruction of transmission line inspection images based on SA-GAN according to claim 1, characterized in that, The processing of the global awareness attention mechanism module includes: Global average pooling is performed on the input feature map, and global channel attention weights are generated through a first fully connected layer and a second fully connected layer. These weights are then multiplied by the input feature map to obtain the first target feature map. The expression for the first fully connected layer is: ， In the formula, This is the output of the first fully connected layer. It is the ReLU activation function. This is the input channel index for the first fully connected layer. This is the weight matrix of the first fully connected layer. As input to the global average pooling result, This is the bias term for the first fully connected layer; The expression for the second fully connected layer is: ， In the formula, For global channel attention weights, It is the Sigmoid activation function. This is the input channel index for the second fully connected layer. This is the weight matrix for the second fully connected layer. This is the output of the first fully connected layer. For the bias term of the second fully connected layer, This is the weight matrix for the second fully connected layer; The expression for the first target feature map is: ， In the formula, The feature map after applying channel attention. The original input feature map, It is the element-wise multiplication operator. For global channel attention weights; Perform a depthwise convolution on the obtained first target feature map, generate a spatial attention weight map using the Sigmoid function, multiply it again by the first target feature map, and output the result, expressed as: ， ， In the formula, This is the final output of the GPA module. The feature map after channel attention processing; Spatial attention weights, For the Sigmoid function, This represents the offset of the convolution kernel in the x-direction. This represents the offset of the convolution kernel in the y-direction. The weights of the depthwise convolution. This is the bias term for depthwise convolution.

4. The method for super-resolution reconstruction of transmission line inspection images based on SA-GAN according to claim 1, characterized in that, The processing of the channel attention mechanism module includes: The input feature map is subjected to a 3×3 convolution to extract local features, expressed as follows: ， In the formula, For local convolutional feature output, For the global fusion attention mechanism channel index, This represents the offset of the convolution kernel in the x-direction. This represents the offset of the convolution kernel in the y-direction. This is the output feature map of the GPA module. This is the bias term for local convolution; Global average pooling is performed on the input feature map, and global channel attention weights are generated through two fully connected layers. These weights are then multiplied by the input feature map to obtain the globally weighted features. The expression for the global channel attention weights is as follows: ， ， In the formula, For LGCAM channel attention weights, For the Sigmoid function, This is the input channel index for the second fully connected layer of LGCAM. For the weights of the second fully connected layer, For the bias of the second fully connected layer, The output of the first fully connected layer. It is the ReLU activation function. This is the input channel index for the first fully connected layer in LGCAM. For the weights of the first fully connected layer, This is the global pooling result of LGCAM. Bias for the first fully connected layer; The local features are added to the global weighted features to obtain the final output, expressed as: ， In the formula, This is the final output of the LGCAM module. Preserving spatial details of local features for local convolutional features. Output features for the GPA module. Features that are globally semantically weighted for global channel attention weights.

5. The method for super-resolution reconstruction of transmission line inspection images based on SA-GAN according to claim 1, characterized in that, The generator loss function of the SA-GAN network is a multi-scale attention feature matching loss, expressed as: ， ， ， In the formula, For multi-scale attention feature matching loss function, This represents the total number of network depth levels. This represents the total number of spatial scale levels. For L1 norm operators, For multi-scale attention weight functions, For multi-scale feature extraction functions, It is the element-wise multiplication operator. The super-resolution output image generated by the generator. For a true high-resolution reference image, To separately and Perform feature extraction. For network depth level indexing, For spatial scale level indexing, It is a 3x3 two-dimensional convolution. The channel grouping function uniformly divides the channel dimensions of the input features into S groups, and then selects the s-th group. For residual block functions, It is the product of the feature map height and the feature map width. For height indexing, Indexed in the width direction. Represented as a slice of the feature map. The Softmax normalization function transforms the input vector into a probability distribution. This is the transpose of the original feature map.

6. A super-resolution reconstruction system for transmission line inspection images based on SA-GAN, characterized in that, include: The acquisition module is configured to acquire the original images of power line inspection collected by the UAV, perform bicubic interpolation to reduce the resolution, obtain low-resolution images, and construct a dataset containing low-resolution images and the original high-resolution images, which is divided into training set and validation set. The construction module is configured to build an SA-GAN super-resolution reconstruction network. The network is iteratively trained and validated using the training set and validation set to obtain a super-resolution reconstruction model. The SA-GAN network includes a generator, which sequentially includes an input layer, a residual layer, a spatial adaptive multi-scale feature modulation module, an upsampling layer, a global perception attention mechanism module, a channel attention mechanism module, and an output layer. The spatially adaptive multi-scale feature modulation module enhances the input features through channel grouping, multi-scale convolution processing, spatially adaptive scale weight calculation and weighted fusion; The global perception attention mechanism module enhances the perception of the overall structure of the image through global channel attention and deep convolutional spatial attention mechanisms; The channel attention mechanism module integrates local convolutional details and global channel attention weighted features to optimize feature representation; The output module is configured to input the low-resolution transmission line inspection image to be reconstructed into the trained super-resolution reconstruction model and output the corresponding super-resolution image.

7. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 5.

Images