A method and system for super-resolution reconstruction of transmission line inspection images

Abstract

The present invention discloses a method and system for super-resolution reconstruction of transmission line inspection images. The method comprises: acquiring at least one transmission line inspection image and performing Sahara blur processing on the transmission line inspection image to obtain at least one target inspection image; constructing a super-resolution reconstruction network based on convolutional shape convolution, and inputting the target inspection image into the super-resolution reconstruction network for training to obtain a target super-resolution reconstruction model; inputting the acquired real-time transmission line inspection image into the target super-resolution reconstruction model, which outputs a reconstructed image. The method can upgrade low-quality, blurry images to high-resolution images, thereby enhancing the clarity of image details and enabling target detection technology to more accurately identify and analyze the state of insulator discharge and heating.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method and system for super-resolution reconstruction of transmission line inspection images. Background Art

[0002] Insulators are essential components in power systems, primarily used to support and isolate high-voltage wires to prevent current leakage through air or other media, thereby ensuring the stability and safety of power transmission. With the continuous growth of electricity demand and the increasing complexity of power systems, higher requirements are being placed on the maintenance and inspection of insulators.

[0003] Traditional insulator inspection methods primarily involve manual inspections and scheduled maintenance. Manual inspections rely on experienced technicians using visual inspections and hand tools. This method is not only inefficient but also susceptible to human error, making it difficult to detect potential faults in a timely manner. While scheduled maintenance can prevent faults to a certain extent, it lacks real-time performance and often results in significant lags, failing to meet the high standards required by modern power systems.

[0004] With the rapid development of computer vision and deep learning technologies, image-based object detection technology is becoming a new trend in insulator inspection. This technology can automatically identify and locate insulators in images, using a UV imager to monitor their discharge and heating status in real time, significantly improving detection accuracy and efficiency. Especially during routine inspections of transmission lines, object detection technology can quickly identify insulator defects such as cracks, damage, and contamination, enabling timely action to prevent power outages and safety incidents caused by insulator failures.

[0005] However, despite the enormous potential of object detection technology for insulator inspection, several challenges remain. For example, the complex environment of power transmission lines, the low-resolution images produced by drones during filming, and the lack of clarity in ambient light during filming all contribute to the difficulty in detecting accurate objects. The scale of objects in images varies significantly, making existing detection algorithms unstable when handling multi-scale objects. Summary of the Invention

[0006] The present invention provides a method and system for super-resolution reconstruction of transmission line inspection images, which are used to solve the technical problem that the scale of the target in the image varies greatly and the performance of the existing detection algorithm is unstable when processing multi-scale targets.

[0007] In a first aspect, the present invention provides a method for super-resolution reconstruction of transmission line inspection images, comprising:

[0008] Acquire at least one transmission line inspection image, and perform Sahara blur processing on the transmission line inspection image to obtain at least one target inspection image;

[0009] A super-resolution reconstruction network is constructed based on convolutional shape convolution, and the target inspection image is input into the super-resolution reconstruction network as a training set for training to obtain a target super-resolution reconstruction model, wherein the loss function of the target super-resolution reconstruction model is expressed as:

[0010] ,

[0011] Where, is the loss function of the target super-resolution reconstruction model, is the multi-scale deformation feature alignment loss weight value, are the features of BConv branches of different sizes in the HDSA module, The standard deviation is Gaussian convolution kernel, is the L2 norm, For scale, is the multi-level attention consistent loss weight value, is the expected value, For the The attention weight of each channel, is a logarithmic function, Represents the feature map In the Variance calculation on channels, is a numerical stability term;

[0012] The acquired real-time transmission line inspection image is input into the target super-resolution reconstruction model, and the target super-resolution reconstruction model outputs a reconstructed image.

[0013] In a second aspect, the present invention provides a transmission line inspection image super-resolution reconstruction system, comprising:

[0014] an acquisition module configured to acquire at least one transmission line inspection image and perform Sahara blur processing on the transmission line inspection image to obtain at least one target inspection image;

[0015] The training module is configured to construct a super-resolution reconstruction network based on convolutional shape convolution, and input the target inspection image as a training set into the super-resolution reconstruction network for training to obtain a target super-resolution reconstruction model, wherein the loss function of the target super-resolution reconstruction model is expressed as:

[0016] ,

[0017] Where, is the loss function of the target super-resolution reconstruction model, is the multi-scale deformation feature alignment loss weight value, are the features of BConv branches of different sizes in the HDSA module, The standard deviation is Gaussian convolution kernel, is the L2 norm, For scale, is the multi-level attention consistent loss weight value, is the expected value, For the The attention weight of each channel, is a logarithmic function, Represents the feature map In the Variance calculation on channels, is a numerical stability term;

[0018] The output module is configured to input the acquired real-time transmission line inspection image into the target super-resolution reconstruction model, and the target super-resolution reconstruction model outputs a reconstructed image.

[0019] According to a third aspect, 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, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the method for super-resolution reconstruction of transmission line inspection images according to any embodiment of the present invention.

[0020] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the program instructions are executed by a processor, the processor executes the steps of the transmission line inspection image super-resolution reconstruction method of any embodiment of the present invention.

[0021] The transmission line inspection image super-resolution reconstruction method and system of the present application have the following beneficial effects:

[0022] 1. A multi-scale feature collaboration mechanism is adopted. By setting up three different scale Bconv modules of 9x9, 5x5, and 3x3 in parallel, the collaborative capture of local image details, regional features, and global context information is achieved. This feature extraction method effectively solves the problem of detail loss in complex texture restoration caused by traditional single-scale convolution. The HDSA module is used to adaptively assign weights to multi-scale features, and the deformable attention mechanism is used to achieve feature space alignment and semantic association. Compared with conventional channel splicing or weighted fusion, the feature fusion effect is improved under low-quality input. Through residual learning of cascade processing and reconstruction modules, the noise amplification effect is suppressed while maintaining high-frequency details, which significantly improves the PSNR indicator compared to traditional residual networks.

[0023] 2. The SRBConv loss function uses a multi-scale Gaussian blur comparison term to enforce the preservation of high-frequency details at different scales, while enhancing feature diversity by maximizing channel variance. Dynamic weight parameters achieve an adaptive balance between detail preservation and feature optimization, ultimately improving image restoration quality while maintaining the efficiency and robustness of model training.

[0024] 3. Through four-way feature decoupling, the traditional square convolution kernel is decomposed into four orthogonal branch convolutions in horizontal (left / right) and vertical (up / down) directions, forming a convoluted receptive field. This design achieves the equivalent perceptual range of a 9x9 convolution while maintaining the computational efficiency of a 3x3 convolution. By extracting direction-sensitive features, each branch uses narrow convolution kernels (kx1 and 1xk) to specifically capture edge and texture features in specific directions. Compared to standard convolution, the error rate in directional texture reconstruction tasks is reduced. Through four-branch channel splitting, the number of parameters is controlled to 1 / 3 of that of a standard convolution while maintaining feature diversity, achieving a balance between computational efficiency and feature expression capabilities.

[0025] 4. Combining the spatial adaptability of convolutional convolution with the global relationship modeling of self-attention, the dynamically generated offset enables attention calculation to focus on semantically related deformation areas, which improves the reasoning speed in image restoration tasks compared to traditional Transformers; the convolutional convolution path retains local geometric details, the self-attention path establishes long-range dependencies, and the dual-path features achieve complementary enhancement through a gating mechanism; through the temperature scaling coefficient k and the learned offset amplitude control, the network can automatically adjust the degree of deformation according to the complexity of the input features, while maintaining stability while improving the modeling ability of complex deformations; the zero initialization of the residual projection layer ensures that the initial stage of the network is equivalent to the identity mapping, and combined with the batch normalization layer, it can improve the convergence speed of deep networks and effectively alleviate the gradient vanishing problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0027] Figure 1 A flowchart of a method for super-resolution reconstruction of transmission line inspection images provided by one embodiment of the present invention;

[0028] Figure 2 A block diagram of a transmission line inspection image super-resolution reconstruction process according to a specific embodiment of the present invention;

[0029] Figure 3 A flow chart of the internal processing of the HDSA module according to one embodiment of the present invention;

[0030] Figure 4 This is a structural block diagram of a transmission line inspection image super-resolution reconstruction system provided by one embodiment of the present invention;

[0031] Figure 5 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0033] See also Figure 1 , which shows a flowchart of a method for super-resolution reconstruction of transmission line inspection images of the present application.

[0034] like Figure 1 As shown in FIG, the super-resolution reconstruction method of the transmission line inspection image specifically includes the following steps:

[0035] Step S101: Acquire at least one transmission line inspection image, and perform Sahara blur processing on the transmission line inspection image to obtain at least one target inspection image.

[0036] In this step, edge detection is performed on a transmission line inspection image, and convolution is performed on the transmission line inspection image using a Gaussian blur kernel to obtain a fuzzy transmission line inspection image. The expression for convolution of the transmission line inspection image using the Gaussian blur kernel is:

[0037] ,

[0038] ,

[0039] Where, is the midpoint of the fuzzy transmission line inspection image The pixel value at The midpoint of the transmission line inspection image The pixel value at is the Gaussian kernel, is the convolution operation, is the standard deviation of the Gaussian kernel, which controls the degree of blur;

[0040] The transmission line inspection image and the fuzzy transmission line inspection image are fused to obtain a target inspection image, which is expressed as:

[0041] ,

[0042] ,

[0043] Where, The midpoint of the target inspection image The pixel value at is the edge strength of each pixel in the image, obtained by the Sobel operator, is the weight factor, is the maximum edge intensity value in the fuzzy transmission line inspection image.

[0044] Step S102: construct a super-resolution reconstruction network based on convolutional shape convolution, and input the target inspection image as a training set into the super-resolution reconstruction network for training to obtain a target super-resolution reconstruction model.

[0045] In this step, the loss function of the target super-resolution reconstruction model is expressed as:

[0046] ,

[0047] Where, is the loss function of the target super-resolution reconstruction model, is the multi-scale deformation feature alignment loss weight value, are the features of Bconv (convolutional large-scale convolution) branches of different sizes in the HDSA module, The standard deviation is Gaussian convolution kernel, is the L2 norm, For scale, is the multi-level attention consistent loss weight value, is the expected value, For the The attention weight of each channel, is a logarithmic function, Represents the feature map In the Variance calculation on channels, is a numerically stable term.

[0048] It should be noted that the target super-resolution reconstruction model includes:

[0049] Input layer: The input is a low-resolution feature map of (1x128x128).

[0050] The parallel convolutional convolution layer connected to the input layer includes a first convolutional convolution layer with a convolution kernel of 9x9, a second convolutional convolution layer with a convolution kernel of 5x5, and a third convolutional convolution layer with a convolution kernel of 3x3. It is used to input the low-resolution input image into three convolutional convolution layers of 9x9, 5x5, and 3x3 in parallel to extract image information in different areas.

[0051] HDSA modules connected with parallel convolutional convolutional layers; and

[0052] The reconstruction layer connected to the HDSA module is used to input the output of the fusion layer into the reconstruction layer. Through two convolutions with a convolution kernel of 3, the number of channels of the image is reduced from 64 to 1 to keep the size consistent with the original input image.

[0053] Step S103: inputting the acquired real-time transmission line inspection image into the target super-resolution reconstruction model, and the target super-resolution reconstruction model outputs a reconstructed image.

[0054] In this step, see Figure 2The real-time transmission line inspection image is input into the first convolutional convolution layer with a convolution kernel of 9x9, the second convolutional convolution layer with a convolution kernel of 5x5, and the third convolutional convolution layer with a convolution kernel of 3x3, respectively. The first, second, and third convolution results are output accordingly. The first, second, and third convolution results are concatenated to obtain a first output image. The first output image is input into the HDSA (High-Dimensional Self-Attention) module. The HDSA module performs deep attention feature fusion on the first feature map to obtain a second output image. The second output image is sequentially subjected to convolutional convolution with a kernel size of 5 and convolutional convolution with a kernel size of 3 to obtain a third output image. The third output image and the real-time transmission line inspection image are input into the reconstruction layer. The reconstruction layer maps the high-level features in the third output image to the real-time transmission line inspection image to obtain a reconstructed image.

[0055] Specifically, the detailed expansion of the convolution operation:

[0056] ,

[0057] Where, is the left-direction convolution kernel, is the right-direction convolution kernel, is the upward convolution kernel, is the convolution kernel in the downward direction, is the residual projection kernel, It is a channel splicing operation;

[0058] The output of all convolutional layers is concatenated, and the expression is:

[0059] ,

[0060] Where, is the first feature map, This is the feature map obtained by the first convolutional convolution layer with a convolution kernel of 9x9. This is the feature map obtained by the first convolutional convolution layer with a convolution kernel of 5x5. This is the feature map obtained by the first convolutional convolution layer with a convolution kernel of 3x3;

[0061] The specific operation method is:

[0062] ,

[0063] Where, It is the feature map after concatenating three-way convolution operations.

[0064] Further, see Figure 3 , the HDSA module performs deep attention feature fusion on the first feature map to obtain a second output image including:

[0065] Perform feature projection on the first feature map, and the expression is:

[0066]

[0067] Where, is the query vector, is the key vector, is a value vector, is the weight matrix of the query vector, is the weight matrix of the key vector, is the weight matrix of the value vector, is the first feature map;

[0068] The query projection formula is:

[0069] ,

[0070] The key projection formula is:

[0071] ,

[0072] The value projection formula is:

[0073] ,

[0074] The query vector and key vector The spatial dimension of is flattened, and the expression is:

[0075] ,

[0076] ,

[0077] ,

[0078] ,

[0079] Where, , represents the spatial position after flattening, is the total number of spatial points in the feature map, is the flattened query matrix, is the number of output batches, for The number of channels, 、 are the length and width of the output position, respectively. is the flattened bond matrix, for The number of channels, is the feature map batch;

[0080] Calculate the attention weights of the flattened query matrix and the flattened key matrix respectively to obtain the attention matrix, which is expressed as:

[0081] ,

[0082] Where, is the attention matrix, is the normalized attention weight, is the transpose of the flattened bond matrix;

[0083] Perform deformable convolution calculations based on the value vector, specifically including:

[0084] right Keep the original dimension for deformable convolution, the expression is:

[0085] ,

[0086] Where, It is the feature map after deformable convolution processing;

[0087] The offset is generated by:

[0088] ,

[0089] Where, is the result of the first offset generation, is the first offset convolution kernel;

[0090] ,

[0091] Where, The result of the second offset is generated. is the second offset convolution kernel, is the convolution operation;

[0092] ,

[0093] Where, is the deformation offset;

[0094] The coordinate calculation formula for deformable convolution is:

[0095] ,

[0096] Where, is the convolution kernel position, For the Location The offset in direction, For the Location The offset in direction, is the weight of the deformable convolution kernel, is the feature map after the input feature map is projected, For full slicing operations, take all elements of the dimension. is the step length, All are relative offsets within the nucleus;

[0097] The attention matrix performs weighted fusion on the output of the deformable convolution to obtain a second output image, wherein the expression of the second output image is:

[0098] ,

[0099] Where, is the second output image, is the activation function, is batch normalization, is the projected convolution kernel, is the convolution operation, is the attention-weighted feature, The original input does not refer to the initial input image of the entire network, but the initial feature map input into the HDSA module by the previous module. x is added to the output of the fused feature through the residual connection, which ensures that the model can retain the original input information in the deep part of the network, thereby optimizing the feature fusion process and increasing the stability during training and the integrity of information transmission.

[0100] It should be noted that the second output image is sequentially subjected to convolutional convolution with a kernel size of 5 and convolutional convolution with a kernel size of 3 to obtain a third output image including:

[0101] The second output image is input into the convolution with a kernel size of 5, and the first result is output, which is expressed as:

[0102] ,

[0103] Where, is a convolutional convolution with a kernel size of 5, is the second output image, is batch normalization, is the activation function, For channel splicing operation, is the left vertical convolution feature, is the right vertical convolution feature, is the upward horizontal convolution feature, is the downward horizontal convolution feature;

[0104] Among them, the expression of the left vertical convolution feature is:

[0105] ,

[0106] Where, After the left convolution operation, the position in the output feature map is The corresponding value, is the number of output batches, is the number of output channels, 、 are the length and width of the output position, respectively. is the vertical displacement within the nucleus, is the left convolution, is the number of input channels, Indicates that the second output image is Sampling is carried out at Represents the position offset of the convolution kernel in the vertical direction (ie, height direction), the displacement of the convolution kernel when it slides along the vertical axis of the image. It represents the sliding step size of the convolution kernel at that position. Indicates the position of the convolution kernel in the horizontal direction. Since the width of the convolution kernel is fixed to 1 in the horizontal direction (that is, the convolution kernel does not expand in the horizontal direction), there is no horizontal offset and it only slides in the vertical direction.

[0107] The expression of the right vertical convolution feature is:

[0108] ,

[0109] Where, To perform horizontal flip of the convolution kernel for left convolution;

[0110] The expression of the upward vertical convolution feature is:

[0111] ,

[0112] Where, After the upward convolution operation, the position in the output feature map is The corresponding value, is the horizontal offset within the nucleus, Indicates that the second output image is Sampling was carried out at Represents the position offset of the convolution kernel in the horizontal direction (i.e., width direction), the displacement of the convolution kernel when it slides along the vertical axis of the image, It represents the sliding step of the convolution kernel at this position. Indicates the position of the convolution kernel in the vertical direction. Since the height of the convolution kernel is fixed to 1 in the vertical direction (that is, the convolution kernel does not expand in the vertical direction), there is no longitudinal offset and it only slides in the horizontal direction.

[0113] The expression of the downward vertical convolution feature is:

[0114] ,

[0115] Where, To perform horizontal flipping of the convolution kernel for upward convolution;

[0116] The first result is input into the first adaptive convolution module, and the second result is output.

[0117] ,

[0118] ,

[0119] Where, is the second calculation result, is the first adaptive convolution module, is the residual weight coefficient, is the convolution operation, is the weighted input;

[0120] The second result is input into the convolutional convolution with a kernel size of 3, and the third result is output. The third result is input into the second adaptive convolution module, and the third output image is output. The expression is:

[0121] ,

[0122] Where, is the third output image, is the second adaptive convolution module, is a convolution with a kernel size of 3.

[0123] The expression for reconstructing the image is:

[0124] ,

[0125] ,

[0126] ,

[0127] ,

[0128] ,

[0129] ,

[0130] Where, To reconstruct the image, Reconstructing the image. For real-time transmission line inspection images, is the tanh output, the output range is [-1, 1], is the hyperbolic tangent function, is the feature map of the final output convolution, is the output feature after LeakyReLU activation function processing, is a 3x3 convolution kernel with no bias. is the second bias term, is the result feature of batch normalization, is the slope of the negative semi-axis, is a learnable scaling parameter, is the mean along the channel dimension, is a learnable offset parameter, To prevent division by zero for small constants, is the variance along the channel dimension, is the dimensionality reduction convolution operation, is the convolution kernel parameter, is the output feature of the dimensionality reduction convolution, is the first bias term, is the third output image.

[0131] Traditional super-resolution reconstruction networks can achieve good reconstruction effects on single blurred images, but since most of their objects are used to restore old images or single images generated by low-resolution visible light devices, they do not consider the reliability of actual applications in which the shooting equipment may be blurred due to insufficient light or due to limited resource deployment on drones, making these models difficult to promote in actual applications in the power field. Therefore, the method of this application uses a new convolutional convolution layer. By using padding in four different directions (horizontal and vertical), it captures texture information in more directions of the insulator, and can upgrade low-quality, blurred images to high-resolution images, thereby enhancing the clarity of image details, so that target detection technology can more accurately identify and analyze the discharge and heating state of the insulator.

[0132] See also Figure 4 , which shows a structural block diagram of a transmission line inspection image super-resolution reconstruction system of the present application.

[0133] like Figure 4 As shown, the transmission line inspection image super-resolution reconstruction system 200 includes an acquisition module 210 , a training module 220 and an output module 230 .

[0134] The acquisition module 210 is configured to acquire at least one transmission line inspection image and perform Sahara blur processing on the transmission line inspection image to obtain at least one target inspection image;

[0135] The training module 220 is configured to construct a super-resolution reconstruction network based on convolutional shape convolution, and input the target inspection image as a training set into the super-resolution reconstruction network for training to obtain a target super-resolution reconstruction model, wherein the loss function of the target super-resolution reconstruction model is expressed as:

[0136] ,

[0137] Where, is the loss function of the target super-resolution reconstruction model, is the multi-scale deformation feature alignment loss weight value, are the features of BConv branches of different sizes in the HDSA module, The standard deviation is Gaussian convolution kernel, is the L2 norm, For scale, is the multi-level attention consistent loss weight value, is the expected value, For the The attention weight of each channel, is a logarithmic function, Represents the feature map In the Variance calculation on channels, is a numerical stability term;

[0138] The output module 230 is configured to input the acquired real-time transmission line inspection image into the target super-resolution reconstruction model, and the target super-resolution reconstruction model outputs a reconstructed image.

[0139] It should be understood that Figure 4 Modules and references documented in Figure 1 Therefore, the operations and features described above for the method and the corresponding technical effects also apply to Figure 4 The modules in it will not be described in detail here.

[0140] In other embodiments, embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor is caused to execute the method for super-resolution reconstruction of transmission line inspection images in any of the above method embodiments;

[0141] As an embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, and the computer-executable instructions are configured as follows:

[0142] Acquire at least one transmission line inspection image, and perform Sahara blur processing on the transmission line inspection image to obtain at least one target inspection image;

[0143] A super-resolution reconstruction network is constructed based on convolutional shape convolution, and the target inspection image is input into the super-resolution reconstruction network as a training set for training to obtain a target super-resolution reconstruction model, wherein the loss function of the target super-resolution reconstruction model is expressed as:

[0144] ,

[0145] Where, is the loss function of the target super-resolution reconstruction model, is the multi-scale deformation feature alignment loss weight value, are the features of BConv branches of different sizes in the HDSA module, The standard deviation is Gaussian convolution kernel, is the L2 norm, For scale, is the multi-level attention consistent loss weight value, is the expected value, For the The attention weight of each channel, is a logarithmic function, Represents the feature map In the Variance calculation on channels, is a numerical stability term;

[0146] The acquired real-time transmission line inspection image is input into the target super-resolution reconstruction model, and the target super-resolution reconstruction model outputs a reconstructed image.

[0147] The computer-readable storage medium may include a program storage area and a data storage area. The program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created based on the use of the transmission line inspection image super-resolution reconstruction system. Furthermore, the computer-readable storage medium may include high-speed random access memory and may also include storage, 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 storage remote from the processor. Such remote storage may be connected to the transmission line inspection image super-resolution reconstruction system via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0148] Figure 5 Schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Figure 5 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, the memory 320, the input device 330 and the output device 340 may be connected via a bus or other means. Figure 5 The example of a bus connection is shown. Memory 320 is the aforementioned computer-readable storage medium. Processor 310 executes the various server functional applications and data processing by running the non-volatile software programs, instructions, and modules stored in memory 320, thereby implementing the above-described method embodiment for super-resolution reconstruction of transmission line inspection images. Input device 330 can receive input digital or character information and generate key signal input related to user settings and function control of the transmission line inspection image super-resolution reconstruction system. Output device 340 may include a display device such as a display screen.

[0149] The electronic device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0150] As an embodiment, the electronic device is applied to a transmission line inspection image super-resolution reconstruction system and is used for a client, 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, and the instructions are executed by the at least one processor to enable the at least one processor to:

[0151] Acquire at least one transmission line inspection image, and perform Sahara blur processing on the transmission line inspection image to obtain at least one target inspection image;

[0152] A super-resolution reconstruction network is constructed based on convolutional shape convolution, and the target inspection image is input into the super-resolution reconstruction network as a training set for training to obtain a target super-resolution reconstruction model, wherein the loss function of the target super-resolution reconstruction model is expressed as:

[0153] ,

[0154] Where, is the loss function of the target super-resolution reconstruction model, is the multi-scale deformation feature alignment loss weight value, are the features of BConv branches of different sizes in the HDSA module, The standard deviation is Gaussian convolution kernel, is the L2 norm, For scale, is the multi-level attention consistent loss weight value, is the expected value, For the The attention weight of each channel, is a logarithmic function, Represents the feature map In the Variance calculation on channels, is a numerical stability term;

[0155] The acquired real-time transmission line inspection image is input into the target super-resolution reconstruction model, and the target super-resolution reconstruction model outputs a reconstructed image.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for super-resolution reconstruction of transmission line inspection images, characterized in that: include: Acquire at least one transmission line inspection image, and perform Gaussian blur and image fusion processing on the transmission line inspection image to obtain at least one target inspection image; A super-resolution reconstruction network is constructed based on convolutional shape convolution, and the target inspection image is input into the super-resolution reconstruction network as a training set for training to obtain a target super-resolution reconstruction model, wherein the loss function of the target super-resolution reconstruction model is expressed as: ， Where, is the loss function of the target super-resolution reconstruction model, is the multi-scale deformation feature alignment loss weight value, are the features of BConv branches of different sizes in the HDSA module, The standard deviation is Gaussian convolution kernel, is the L2 norm, For scale, is the multi-level attention consistent loss weight value, is the expected value, For the The attention weight of each channel, is a logarithmic function, Represents the feature map In the Variance calculation on channels, is a numerical stability term; The convolutional convolution decomposes the traditional square convolution kernel into four orthogonal branch convolutions in horizontal and vertical directions through four-way feature decoupling; The target super-resolution reconstruction model includes: Input layer; A parallel convolutional convolution layer connected to the input layer, wherein the parallel convolutional convolution layer includes a first convolutional convolution layer with a convolution kernel of 9x9, a second convolutional convolution layer with a convolution kernel of 5x5, and a third convolutional convolution layer with a convolution kernel of 3x3; an HDSA module connected to the parallel convolutional convolutional layer; and a reconstruction layer connected to the HDSA module; Inputting the acquired real-time transmission line inspection image into the target super-resolution reconstruction model, and outputting the target super-resolution reconstruction model to obtain a reconstructed image, wherein inputting the acquired real-time transmission line inspection image into the target super-resolution reconstruction model, and outputting the target super-resolution reconstruction model to obtain a reconstructed image comprises: The real-time transmission line inspection image is respectively input into the first convolutional convolution layer with a convolution kernel of 9x9, the second convolutional convolution layer with a convolution kernel of 5x5, and the third convolutional convolution layer with a convolution kernel of 3x3, and the first convolution result, the second convolution result, and the third convolution result are obtained as the corresponding outputs; splicing the first convolution result, the second convolution result, and the third convolution result to obtain a first output image; Inputting the first output image into the HDSA module, and performing deep attention feature fusion on the first output image according to the HDSA module to obtain a second output image; Apply convolutional convolution with a kernel size of 5 and convolutional convolution with a kernel size of 3 to the second output image to obtain a third output image; The third output image and the real-time transmission line inspection image are input into a reconstruction layer, and high-level features in the third output image are mapped into the real-time transmission line inspection image according to the reconstruction layer, so as to finally obtain a reconstructed image.

2. The method for super-resolution reconstruction of transmission line inspection images according to claim 1, characterized in that: The acquiring of at least one transmission line inspection image, and performing Gaussian blur and image fusion processing on the transmission line inspection image to obtain at least one target inspection image includes: Edge detection is performed on a transmission line inspection image, and the transmission line inspection image is convolved with a Gaussian blur kernel to obtain a fuzzy transmission line inspection image. The expression for convolving the transmission line inspection image with a Gaussian blur kernel is: ， ， Where, is the midpoint of the fuzzy transmission line inspection image The pixel value at The midpoint of the transmission line inspection image The pixel value at is the Gaussian kernel, is the convolution operation, is the standard deviation of the Gaussian kernel, which controls the degree of blur; The transmission line inspection image and the fuzzy transmission line inspection image are fused to obtain a target inspection image, which is expressed as: ， ， Where, The midpoint of the target inspection image The pixel value at is the edge strength of each pixel in the image, is the weight factor, is the maximum edge intensity value in the fuzzy transmission line inspection image.

3. The method for super-resolution reconstruction of transmission line inspection images according to claim 1, characterized in that: The step of performing deep attention feature fusion on the first output image according to the HDSA module to obtain a second output image includes: Perform feature projection on the first output image, and the expression is: Where, is the query vector, is the key vector, is a value vector, is the weight matrix of the query vector, is the weight matrix of the key vector, is the weight matrix of the value vector, is the first output image; The query vector and key vector The spatial dimension of is flattened, and the expression is: ， ， ， ， Where, , represents the spatial position after flattening, is the total number of spatial points in the feature map, is the flattened query matrix, is the number of output batches, for The number of channels, 、 are the length and width of the output position, respectively. is the flattened bond matrix, for The number of channels, is the feature map batch; Calculate the attention weights of the flattened query matrix and the flattened key matrix respectively to obtain the attention matrix, which is expressed as: ， Where, is the attention matrix, is the normalized attention weight, is the transpose of the flattened bond matrix; Perform a deformable convolution calculation according to the value vector, and perform weighted fusion on the output of the deformable convolution according to the attention matrix to obtain a second output image, wherein the expression of the second output image is: ， Where, is the second output image, is the activation function, is batch normalization, is the projected convolution kernel, is the convolution operation, is the attention-weighted feature, is the original input.

4. The method for super-resolution reconstruction of transmission line inspection images according to claim 1, characterized in that: The step of sequentially subjecting the second output image to a convolutional convolution with a kernel size of 5 and a convolutional convolution with a kernel size of 3 to obtain a third output image includes: The second output image is input into the convolution with a kernel size of 5, and the first result is output, which is expressed as: ， Where, is a convolutional convolution with a kernel size of 5, is the second output image, is batch normalization, is the activation function, For channel splicing operation, is the left vertical convolution feature, is the right vertical convolution feature, is the upward horizontal convolution feature, is the downward horizontal convolution feature; Among them, the expression of the left vertical convolution feature is: ， Where, After the left convolution operation, the value corresponding to the position in the output feature map is is the number of output batches, is the number of output channels, 、 are the length and width of the output position, respectively. is the vertical displacement within the nucleus, is the left convolution, is the number of input channels, Indicates that the second output image is Sampling was carried out at The expression of the right vertical convolution feature is: ， Where, To perform horizontal flip of the convolution kernel for left convolution; The expression of the upward vertical convolution feature is: ， Where, After the upward convolution operation, the value corresponding to the position in the output feature map is is the horizontal offset within the nucleus, Indicates that the second output image is Sampling was carried out at The expression of the downward vertical convolution feature is: ， Where, To perform horizontal flipping of the convolution kernel for upward convolution; The first result is input into the first adaptive convolution module, and the second result is output. ， ， Where, is the second calculation result, is the first adaptive convolution module, is the residual weight coefficient, is the convolution operation, is the weighted input; The second result is input into the convolutional convolution with a kernel size of 3, and the third result is output. The third result is input into the second adaptive convolution module, and the third output image is output. The expression is: ， Where, is the third output image, is the second adaptive convolution module, is a convolution with a kernel size of 3.

5. The method for super-resolution reconstruction of transmission line inspection images according to claim 1, characterized in that: The expression of the reconstructed image is: ， ， ， ， ， ， Where, To reconstruct the image, Reconstructing the image. For real-time transmission line inspection images, is the tanh output, the output range is [-1, 1], is the hyperbolic tangent function, is the feature map of the final output convolution, is the output feature after LeakyReLU activation function processing, is a 3x3 convolution kernel with no bias. is the second bias term, is the result feature of batch normalization, is the slope of the negative semi-axis, is a learnable scaling parameter, is the mean along the channel dimension, is a learnable offset parameter, To prevent division by zero for small constants, is the variance along the channel dimension, is the dimensionality reduction convolution operation, is the convolution kernel parameter, is the output feature of the dimensionality reduction convolution, is the first bias term, is the third output image.

6. A transmission line inspection image super-resolution reconstruction system, characterized in that: include: an acquisition module configured to acquire at least one transmission line inspection image, and perform Gaussian blur and image fusion processing on the transmission line inspection image to obtain at least one target inspection image; The training module is configured to construct a super-resolution reconstruction network based on convolutional shape convolution, and input the target inspection image as a training set into the super-resolution reconstruction network for training to obtain a target super-resolution reconstruction model, wherein the loss function of the target super-resolution reconstruction model is expressed as: ， Where, is the loss function of the target super-resolution reconstruction model, is the multi-scale deformation feature alignment loss weight value, are the features of BConv branches of different sizes in the HDSA module, The standard deviation is Gaussian convolution kernel, is the L2 norm, For scale, is the multi-level attention consistent loss weight value, is the expected value, For the The attention weight of each channel, is a logarithmic function, Represents the feature map In the Variance calculation on channels, is a numerical stability term; The convolutional convolution decomposes the traditional square convolution kernel into four orthogonal branch convolutions in horizontal and vertical directions through four-way feature decoupling; The target super-resolution reconstruction model includes: Input layer; A parallel convolutional convolution layer connected to the input layer, wherein the parallel convolutional convolution layer includes a first convolutional convolution layer with a convolution kernel of 9x9, a second convolutional convolution layer with a convolution kernel of 5x5, and a third convolutional convolution layer with a convolution kernel of 3x3; an HDSA module connected to the parallel convolutional convolutional layer; and a reconstruction layer connected to the HDSA module; An output module is configured to input the acquired real-time transmission line inspection image into the target super-resolution reconstruction model, and the target super-resolution reconstruction model outputs a reconstructed image, wherein the inputting the acquired real-time transmission line inspection image into the target super-resolution reconstruction model, and the target super-resolution reconstruction model outputs a reconstructed image includes: The real-time transmission line inspection image is respectively input into the first convolutional convolution layer with a convolution kernel of 9x9, the second convolutional convolution layer with a convolution kernel of 5x5, and the third convolutional convolution layer with a convolution kernel of 3x3, and the first convolution result, the second convolution result, and the third convolution result are obtained as the corresponding outputs; splicing the first convolution result, the second convolution result, and the third convolution result to obtain a first output image; Inputting the first output image into the HDSA module, and performing deep attention feature fusion on the first output image according to the HDSA module to obtain a second output image; Apply convolutional convolution with a kernel size of 5 and convolutional convolution with a kernel size of 3 to the second output image to obtain a third output image; The third output image and the real-time transmission line inspection image are input into a reconstruction layer, and high-level features in the third output image are mapped into the real-time transmission line inspection image according to the reconstruction layer, so as to finally obtain a reconstructed 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, and the instructions are executed 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 a processor, the method according to any one of claims 1 to 5 is implemented.

Images