Power transmission line inspection image processing method and system and computer readable medium

Abstract

The invention discloses a power transmission line inspection image processing method and system and a computer readable medium, and belongs to the technical field of image processing.The method comprises the steps that firstly, it is found that GCANet does not consider the influence of adjacent pixels by analyzing a GCANeT network structure, and meanwhile, due to the fact that a power transmission line inspection scene has the characteristic of high resolution in fine rain weather, the influence of the adjacent pixels is not considered by GCANet; a better rain removal effect is difficult to obtain by using an original Decoder structure. Then, in order to solve the defects existing in the GCANet, a rain removal network (AAM-GCANet) in which an adjacency information auxiliary prediction module and a multi-scale feature enhancement decoding module are fused is provided, and then, in order to solve the defects existing in the GCANet, the invention provides a rain removal network (AAM-GCANet) in which an adjacency information auxiliary prediction module and a multi-scale feature enhancement decoding module are fused. The first pixel extraction module is used for extracting adjacent pixel information of a pixel point and assisting in predicting a fine rain component of the point, and the second pixel extraction module is used for completing a high-resolution decoding task by optimizing a decoding process and fusing multi-scale information in the decoding process. The method can be applied to power transmission line inspection image processing operation.

Description

Technical Field

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

[0002] Due to its complex and changeable characteristics, the transmission line scene is more susceptible to interference from factors such as drizzle, noise, and strong light than the indoor scene, resulting in poor imaging quality and adverse effects on subsequent environmental perception algorithms. For the problem of noise interference in transmission line images, the filtering-based denoising method can effectively filter out the noise. At the same time, the existing image enhancement research has also achieved good results for the strong light problem. However, the problem of drizzle in the transmission line scene is still a problem that needs to be solved in the field of visual perception of transmission line inspection.

[0003] The original rain removal algorithm achieves rain removal by combining the depth information of the image. However, when facing images in a single scene, the rain removal effect is greatly reduced due to the inability to accurately obtain the depth information of the image, and it is difficult to apply it to the actual transmission line inspection image. At present, the rain removal algorithm based on deep learning has the advantages of strong feature extraction ability and high rain removal accuracy, and has gradually become the mainstream in the field of image rain removal. For example, a rain streak removal method for power images disclosed in the patent document with publication number CN 117575951A includes: collecting actual inspection images in the power system to obtain power images; constructing a rain removal network model according to the feature aggregation module FAM, the continuous residual channel attention module RCA and the fusion module; using the power image to obtain several features and determine the objective function L; optimizing the rain removal network model according to the objective function L; determining the image of the power image after rain streaks are removed according to the optimized rain removal network model; the invention is based on constructing and optimizing the rain removal network model, and finally using the optimized rain removal network model to restore the rain inspection image collected by the drone inspection, thereby improving the execution efficiency of tasks such as later target recognition and detection. However, it does not solve the problem of high-resolution restoration of drizzle images. Summary of the invention

[0004] In view of this, the present invention aims at the deficiencies in the prior art and provides a method and system for processing transmission line inspection images.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a transmission line inspection image processing method, comprising the following steps:

[0006] S1: Use the GCANet network to extract features from the input transmission line inspection image under drizzling weather to obtain multi-level feature information;

[0007] S2: Use the gated aggregation network to fuse multi-level feature information to obtain fused feature information;

[0008] S3: Use the multi-scale feature enhancement decoding module and the adjacent information auxiliary prediction module to replace the Decoder module in the GCANet network to process the fused feature information and obtain the corrected transmission line inspection image.

[0009] Furthermore, in S1, a conventional convolution operation is performed on the transmission line inspection image to obtain a first feature map, local pixel information of the transmission line inspection image is fused with the first feature map to obtain a fused feature map, and a hole convolution process is performed on the fused feature map to obtain multi-level feature information.

[0010] Furthermore, the multi-level feature information includes low-level feature information, middle-level feature information, and high-level feature information.

[0011] Furthermore, in S2, a gated aggregation network is used to learn the feature information of each level in the multi-level feature information to obtain the weight of the feature information of each level, and the fused feature information is determined according to the feature information of each level and the corresponding weight.

[0012] Furthermore, in S3, the multi-scale feature enhancement decoding module processes the fused feature information to obtain decoded feature information for restoring the high-resolution drizzle image, and the adjacent information auxiliary prediction module processes the decoded feature information to obtain a corrected transmission line inspection image.

[0013] Furthermore, before performing conventional convolution operation on the transmission line inspection image to obtain the first feature map, it also includes: performing separate convolution on the transmission line inspection image to obtain a second feature map, each feature point in the second feature map integrates features of its surrounding (2r-1) size, where r = expansion rate.

[0014] Furthermore, the multi-scale feature enhancement decoding module processes the fused feature information, including: decoding the fused feature information multiple times, generating a feature compensation plane with a resolution of H through a 3×3 convolution operation for the feature layer obtained in each decoding, and splicing the feature compensation plane with the feature layer for the next decoding, and then splicing the final feature layer with the feature compensation planes retained in the decoding process to obtain the decoded feature information for restoring the high-resolution drizzle picture.

[0015] Furthermore, the adjacent information auxiliary prediction module processes the decoded feature information, including: performing original drizzle component prediction and adjacent information prediction on the decoded feature information; the original drizzle component prediction information is obtained by convolving the decoded feature information through a conventional 3*3 convolution layer, and the adjacent information prediction information is obtained by weighted summing {x0, x1..., x6, x7} through an adaptive attention mechanism, {x0, x1..., x6, x7} represents the correlation between a certain pixel point in the decoded feature information and its adjacent pixel points; the original drizzle component prediction information is fused with the adjacent information prediction information to obtain a corrected transmission line inspection image.

[0016] The present invention also provides a transmission line inspection image processing system, comprising:

[0017] Feature extraction module: configured to use the GCANet network to extract features from the input transmission line inspection image under drizzling weather to obtain multi-level feature information;

[0018] Feature fusion module: configured to use a gated aggregation network to fuse multi-level feature information to obtain fused feature information;

[0019] Decoding module: It is configured to use the multi-scale feature enhancement decoding module and the adjacent information auxiliary prediction module to replace the Decoder module in the GCANet network to process the fused feature information to obtain the corrected transmission line inspection image.

[0020] Furthermore, the feature extraction module performs conventional convolution operations on the transmission line inspection image to obtain a first feature map, fuses local pixel information of the transmission line inspection image with the first feature map to obtain a fused feature map, and performs hole convolution processing on the fused feature map to obtain multi-level feature information, wherein the multi-level feature information includes low-level feature information, mid-level feature information, and high-level feature information.

[0021] Furthermore, the feature fusion module uses a gated aggregation network to learn feature information of each level in the multi-level feature information to obtain weights of the feature information of each level, and determines the fused feature information according to the feature information of each level and the corresponding weights.

[0022] Furthermore, the multi-scale feature enhancement decoding module processes the fused feature information to obtain decoded feature information for restoring the high-resolution drizzle picture, and the adjacent information auxiliary prediction module processes the decoded feature information to obtain a corrected transmission line inspection image.

[0023] Furthermore, it also includes a separation convolution module, which is configured to perform separation convolution on the transmission line inspection image to obtain a second feature map before performing a conventional convolution operation on the transmission line inspection image to obtain a first feature map, wherein each feature point in the second feature map is integrated with features of its surrounding (2r-1) size, where r = expansion rate.

[0024] Furthermore, the multi-scale feature enhancement decoding module processes the fused feature information, including: decoding the fused feature information multiple times, generating a feature compensation plane with a resolution of H through a 3×3 convolution operation for each feature layer obtained by decoding, splicing the feature compensation plane with the feature layer for the next decoding, and then splicing the final feature layer with each feature compensation plane retained in the decoding process to obtain decoded feature information for restoring a high-resolution drizzle picture;

[0025] The adjacent information auxiliary prediction module processes the decoded feature information, including: performing original drizzle component prediction and adjacent information prediction on the decoded feature information; the original drizzle component prediction information is obtained by convolving the decoded feature information through a conventional 3*3 convolution layer, and the adjacent information prediction information is obtained by weighted summing {x0, x1..., x6, x7} through an adaptive attention mechanism, {x0, x1..., x6, x7} represents the correlation between a certain pixel point in the decoded feature information and its adjacent pixel points; the original drizzle component prediction information is fused with the adjacent information prediction information to obtain a corrected transmission line inspection image.

[0026] The present invention also provides a computer device, comprising one or more processors, wherein the processor stores a program, and when the program is executed by the processor, the power transmission line inspection image processing method as described in any one of claims 1 to 8 is implemented.

[0027] The present invention also provides a computer-readable medium storing a computer program, and when the computer program is executed, the above-mentioned transmission line inspection image processing method is implemented.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The power transmission line inspection image processing method of the present invention first uses the GCANet network to extract features according to the input power transmission line inspection image in drizzling weather. The GCANet network uses smooth hole convolution instead of hole convolution to further increase the receptive field; then the feature fusion is performed through the gated aggregation network. For the different levels of input features, the gated aggregation network obtains the feature weights of each level through learning, and performs weighted summation on the features of each level; finally, the decoder part of the GCANeT network is improved, and a multi-scale feature enhancement decoding module and an adjacent information auxiliary prediction module are added to obtain an improved new network AAM-GCANet, which optimizes the decoding process, fuses the multi-scale information in the decoding process, and completes the high-resolution decoding task.

[0030] Based on GCANet, the present invention proposes a rain removal network (AAM-GCANet) that integrates an adjacent information auxiliary prediction module and a multi-scale feature enhancement decoding module. It mainly includes two parts: Encoder and Decoder. The improved Decoder part of the present invention is composed of a multi-scale feature enhancement decoding module and an adjacent information auxiliary prediction module. The former is used to extract the adjacent pixel information of the pixel point to assist in predicting the fine rain component of the point, and the latter completes the high-resolution decoding task by optimizing the decoding process and integrating the multi-scale information in the decoding process. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of a method for processing power transmission line inspection images according to an embodiment of the present invention;

[0032] Figure 2 It is a structural diagram of the GCANet algorithm in an embodiment of the present invention;

[0033] Figure 3 This is the overall framework diagram of AAM-GCANet in an embodiment of the present invention;

[0034] Figure 4 is a schematic diagram of a smoothed atrous convolution in an embodiment of the present invention;

[0035] Figure 5 : is a comparison diagram before and after the improvement of the decoder structure in an embodiment of the present invention;

[0036] Figure 6 Schematic diagram of an adjacency information auxiliary prediction module in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0038] Example 1

[0039] Reference Figure 1 The power transmission line inspection image processing method according to the embodiment of the present invention comprises the following steps:

[0040] Step S1: Feature extraction: The GCANet network is used to extract features from the input transmission line inspection image under drizzling weather to obtain multi-level feature information, specifically:

[0041] Step S11 separation convolution: By analyzing the smooth hole convolution process, it can be seen that the four adjacent pixels output after the convolution are independent of the pixels that depend on each other in the previous layer, that is, there is no adjacent equidependency between pixels of the same color. Therefore, before the smooth hole convolution, a separation convolution operation with a kernel of (2r-1) is added, and the convolution parameters are shared by all channels. "Separable" is the idea of ​​separable convolution, and "sharing" means that all channels share the convolution weights. Separate convolution is performed on the transmission line inspection image to obtain the second feature map. Each feature point in the second feature map is integrated with the features of the size of (2r-1) around it, where r = expansion rate.

[0042] Step S12: Smoothed dilated convolution: In order to solve the problem that dilated convolution is difficult to apply to rain removal tasks, GCANet proposes smoothed dilated convolution. First, a conventional convolution operation is performed on the transmission line inspection image to obtain a first feature map, and the local pixel information of the transmission line inspection image is fused with the first feature map to obtain a fused feature map. The fused feature map is subjected to dilated convolution to obtain multi-level feature information, which includes low-level feature information, mid-level feature information, and high-level feature information. This method allows a single pixel to also contain local information, solving the problem of dilated convolution losing information continuity, such as Figure 4 As shown in the figure, the difference between the features extracted before and after smoothing shows that the smoothed hole convolution extracts the fusion information of adjacent pixels and maintains the continuity of information between pixels.

[0043] Step S2: feature fusion: use the gated aggregation network to fuse the multi-level feature information to obtain the fused feature information, specifically:

[0044] Step S21: After learning the feature information using the smooth dilated convolution operation, the algorithm obtains multi-level feature information. GCANet extracts features from low, medium, and high levels, and uses the gated aggregation network to achieve feature fusion. The gated aggregation network is used to learn the feature information of each level in the multi-level feature information to obtain the weight of the feature information of each level, and the fused feature information is determined based on the feature information of each level and the corresponding weight. l , F m , F h}, the gated aggregation network obtains the feature weights of each level through learning {M l , M m , M h}, perform weighted summation of features at each level.

[0045] F o =M l *F l +M m *F m +M h *F h

[0046] In the formula, F o Represents the fused features.

[0047] Step S3 decoding: Use the multi-scale feature enhancement decoding module and the adjacent information auxiliary prediction module to replace the Decoder module in the GCANet network to process the fused feature information to obtain the corrected transmission line inspection image, specifically:

[0048] The multi-scale feature enhancement decoding module processes the fused feature information to obtain the decoded feature information used to restore the high-resolution drizzle picture. The adjacent information auxiliary prediction module processes the decoded feature information to obtain the corrected transmission line inspection image.

[0049] The multi-scale feature enhancement decoding module processes the fused feature information, including: decoding the fused feature information multiple times, generating a feature compensation plane with a resolution of H through a 3×3 convolution operation for the feature layer obtained in each decoding, and splicing the feature compensation plane with the feature layer for the next decoding, and then splicing the final feature layer with the feature compensation planes of different scales retained in the decoding process to obtain the decoded feature information for restoring the high-resolution drizzle picture.

[0050] The adjacent information auxiliary prediction module processes the decoded feature information, including: original drizzle component prediction and adjacent information prediction for the decoded feature information; the original drizzle component prediction information is obtained by convolving the decoded feature information through a conventional 3*3 convolution layer, and the adjacent information prediction information is obtained by weighted summing {x0, x1..., x6, x7} through an adaptive attention mechanism, {x0, x1..., x6, x7} represents the correlation between a certain pixel point in the decoded feature information and its adjacent pixel points; the original drizzle component prediction information is fused with the adjacent information prediction information to obtain a corrected transmission line inspection image.

[0051] The improved decoder part of the present invention is composed of a multi-scale feature enhancement decoding module and an adjacent information auxiliary prediction module. The multi-scale feature enhancement decoding module reconstructs feature maps of different resolutions to the same size during the decoding process and constructs residual structure fusion features, thereby enriching the decoding features for restoring high-resolution drizzle images. The adjacent information auxiliary prediction module predicts the influence of 8 adjacent pixels on the drizzle component, corrects the drizzle component prediction result after weighting, obtains the final drizzle component, and completes rain removal.

[0052] Step S31 Multi-scale feature enhancement decoding module: The Encoder-Decoder structure has the characteristic of achieving resolution restoration, and is therefore widely used in pixel-level image processing tasks such as image deraining, depth estimation, and segmentation. The typical decoding structure usually achieves upsampling operations by cascading deconvolution layers and convolution layers. Figure 5 As shown in (a), in order to enhance the decoding features and improve the accuracy of the rain removal algorithm, the present invention proposes a multi-scale feature enhancement decoding module (Multi-scale feature enhancement Decoding module, MSFE-Decoder) based on the residual structure. Figure 5 (b) as shown.

[0053] Compared with the original Decoder structure, the multi-scale feature enhancement decoding module proposed in the present invention makes full use of the features in each decoding process and generates a feature compensation plane with a resolution of H through a 3×3 convolution operation. Taking the feature layer obtained by the first decoding operation as an example, the feature resolution of this layer is H / 2. X1 is downsampled to obtain a feature compensation plane with the same resolution. After concatenation (concat), it is input into the next decoding process and the plane is retained. The subsequent decoding process is the same. It is worth noting that for the decoding features of different resolutions at different stages, the resolution of the generated feature compensation plane is H, which is achieved by adjusting the convolution kernel parameters. The final decoded features and the feature decoding planes of different scales retained in the decoding process are concatenated and output, which is the output of the multi-scale feature enhancement decoding module.

[0054] Y = concat([X1,X2,X3,X4])

[0055] Where, X i represents the feature compensation plane or the final decoding output produced by the i-th decoding process.

[0056] In summary, the MSFE-Decoder proposed in the present invention strengthens the features of each decoding through the residual structure and retains feature compensation planes of different scales during the decoding process to enhance the multi-scale expression capability of the features of the decoding process, thereby improving the feature understanding capability of the network when processing high-resolution transmission line images.

[0057] The multi-scale feature enhancement decoding module uses the multi-scale features of the middle layer in the decoding process to enhance the final decoding features, where the middle layer multi-scale features are {X1, X2, X3}. The present invention verifies the impact of the middle layer multi-scale feature enhancement on the algorithm deraining performance by controlling variables. The original GCANet without {X1, X2, X3} feature enhancement is used as the baseline model. MSFE-Decoder-1 indicates the use of X1 feature to enhance the decoding process, MSFE-Decoder-2 indicates the use of X1 feature and X2 feature to enhance the decoding process, and MSFE-Decoder-3 indicates the use of X1 feature, X2 feature and X3 feature to enhance the decoding process at the same time.

[0058] The experimental results are shown in Table 1. When using only MSFE-Decoder as an improvement strategy and using all multi-scale features to enhance decoding, the best deraining performance is obtained, where SSIM and PNSR reach 0.9891 and 32.4436 respectively. Obviously, the MSFE-Decoder strategy is effective. When only X1 features are used to enhance decoding, SSIM and PNSR are improved by 0.0137 and 1.3894 respectively, which also verifies the effectiveness of the MSFE-Decoder strategy. At the same time, it is not difficult to see that the X1 feature and X2 feature have a more obvious performance improvement on the deraining algorithm, while the X3 feature has a smaller improvement. This is because the X2 feature and the X3 feature are obtained from intermediate features at different stages of the decoding process but with the same resolution. The X3 feature has a smaller effect on enhancing the multi-scale expression ability of the deraining algorithm.

[0059] Table 1 Experimental results for verifying the effectiveness of the MSFE-Decoder strategy

[0060]

[0061] Step S32 Adjacent information assists the prediction module: For pixel-level image processing tasks, there is a strong correlation between local pixels, especially in drizzle weather, the drizzle component of a certain pixel has a strong correlation with the drizzle component of its adjacent pixel. In order to explore the correlation of drizzle components between pixels, the present invention proposes an adjacent information assists the prediction module (AIP-module). For a certain pixel in the decoded feature information of the drizzle transmission line scene image, there are 8 adjacent pixels, and its drizzle component is affected by the adjacent pixels such as Figure 6 As shown in (a), {x0,x1...,x6,x7} represents the correlation between the adjacent pixels and the point.

[0062] The AIP-module network structure proposed by the present invention is as follows Figure 6 As shown in (b), the network is divided into two parts, one is the original drizzle component M1 prediction network, and the other is the neighbor information M2 prediction network. M1 is obtained through conventional 3*3 convolution, while M2 as neighbor auxiliary information is obtained by weighted summation of {x0,x1...,x6,x7} through the adaptive attention mechanism. The adaptive attention mechanism obtains the weights of each adjacent pixel {w0,w1,w2...,w6,w7} through convolution and pooling operations, and performs dot multiplication with the input {x0,x1...,x6,x7}.

[0063] M2=g({x0,x1......,x6,x7},{w1,w2......,w6,w7})

[0064] Finally, the result of the adjacency information auxiliary module is fused with the original drizzle component prediction result to obtain the final drizzle component information in the drizzle transmission line scene image.

[0065] M=M1+M2

[0066] After performing difference processing with the original image, the drizzle restored image can be obtained.

[0067] The adjacent information auxiliary prediction module quantifies the influence of adjacent pixels on the drizzle component and uses the prediction results as constraint information to assist the prediction. Obviously, the number of adjacent pixels k is different, and the strength of the constraint information is also different. Therefore, the present invention uses k=0, that is, not using the AIP-module strategy, as the baseline model of the experiment, and verifies k=4, that is, the semi-enclosed case of adjacent pixels, and k=8, that is, the fully enclosed case. The impact of the AIP-module strategy on the performance of the deraining algorithm is shown in Table 2.

[0068] The experimental results show that when the adjacent pixels are in the four directions of up, down, left, and right, that is, k = 4, the AIP-module strategy improves the performance of the deraining algorithm, and the SSIM and PNSR are improved by 0.0187 and 1.8879 respectively. This result shows that it is effective to use the adjacent pixel information to assist in predicting the drizzle component. When k = 8, the AIP-module strategy enables the algorithm to obtain the best deraining performance, and the SSIM and PNSR reach 0.9913 and 32.1354 respectively.

[0069] Table 2 Comparison of rain removal performance at different k values

[0070]

[0071] This also shows that the AIP-module strategy proposed in the present invention is effective, and the higher the intensity of the adjacency information and the more adjacent pixels, the higher the improvement in the performance of the deraining algorithm.

[0072] Under the same hardware conditions, the rain removal performance of the AAM-GCANet proposed in this invention and other typical rain removal algorithms in the transmission line scenario in drizzling weather is compared and analyzed. The rain removal performance of each algorithm in the verification data set is shown in Table 3.

[0073] Table 3 Performance comparison of various rain removal algorithms

[0074]

[0075] The experimental results show that the DCP deraining algorithm based on the atmospheric scattering model and dark channel prior has the worst deraining performance, with a structural similarity SSIM of only 0.6864, which is due to the diverse backgrounds and complex light sources of the Foggy Cityscapes dataset. At the same time, although AOD-Net uses convolutional neural network to extract features instead of traditional dark channel features to enhance the feature expression ability and achieve better deraining effect, it still has not gotten rid of the defects of the atmospheric scattering model itself, and the deraining performance is limited. Compared with the first two deraining algorithms that rely on the atmospheric scattering model, GCANet, which directly focuses on the correlation between drizzle images and clear images, relies on its Encoder-Decoder features to achieve better performance. The AAM-GCANet proposed in the present invention achieves the best deraining effect, with SSIM increased by 4.3% and PSNR increased by 4.7797. This is because the multi-scale feature enhancement decoding module used in the present invention achieves deraining performance improvement by enhancing the decoding features in the deraining process and strengthening the multi-scale expression ability of the algorithm. At the same time, the prediction accuracy of the drizzle component of each pixel in the image is further improved through the adjacency information auxiliary prediction module.

[0076] Example 2

[0077] The embodiment of the present invention is a transmission line inspection image processing system, comprising:

[0078] Feature extraction module: configured to use the GCANet network to extract features from the input transmission line inspection image under drizzling weather to obtain multi-level feature information;

[0079] The specific configuration is to perform the feature extraction in step S1 of Example 1;

[0080] Feature fusion module: configured to use a gated aggregation network to fuse multi-level feature information to obtain fused feature information;

[0081] The specific configuration is to perform the feature fusion in step S2 of embodiment 1;

[0082] Decoding module: configured to use a multi-scale feature enhancement decoding module and an adjacent information auxiliary prediction module to replace the Decoder module in the GCANet network to process the fused feature information and obtain a corrected transmission line inspection image;

[0083] The specific configuration is to execute the decoding in step S3 of Example 1.

[0084] Example 3

[0085] An embodiment of the present invention provides a computer device, including one or more processors, wherein the processor stores a program, and when the program is executed by the processor, the power transmission line inspection image processing method in embodiment 1 is implemented.

[0086] Example 4

[0087] The embodiment of the present invention is a computer-readable medium storing a computer program. When the computer program is executed, the power transmission line inspection image processing method in Embodiment 1 is implemented.

[0088] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, a person skilled in the art can still modify or make equivalent substitutions to the specific implementations of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention are within the scope of protection of the claims of the present invention to be approved.

Claims

1. A method for processing transmission line inspection images, characterized in that: The following steps are involved: S1: Use the GCANet network to extract features from the input transmission line inspection image under drizzling weather to obtain multi-level feature information; S2: Use the gated aggregation network to fuse multi-level feature information to obtain fused feature information; S3: Use the multi-scale feature enhancement decoding module and the adjacent information auxiliary prediction module to replace the Decoder module in the GCANet network to process the fused feature information and obtain the corrected transmission line inspection image.

2. The method for processing power line inspection images according to claim 1, characterized in that: In S1, a conventional convolution operation is performed on the transmission line inspection image to obtain a first feature map, local pixel information of the transmission line inspection image is fused with the first feature map to obtain a fused feature map, and a hole convolution process is performed on the fused feature map to obtain multi-level feature information.

3. The method for processing power line inspection images according to claim 1, characterized in that: The multi-level feature information includes low-level feature information, middle-level feature information, and high-level feature information.

4. The method for processing power line inspection images according to claim 1, characterized in that: In S2, a gated aggregation network is used to learn the feature information of each level in the multi-level feature information to obtain the weight of the feature information of each level, and the fused feature information is determined according to the feature information of each level and the corresponding weight.

5. The method for processing power line inspection images according to claim 1, characterized in that: In S3, the multi-scale feature enhancement decoding module processes the fused feature information to obtain the decoded feature information used to restore the high-resolution drizzle picture, and the adjacent information auxiliary prediction module processes the decoded feature information to obtain the corrected transmission line inspection image.

6. The method for processing power line inspection images according to claim 2, characterized in that: Before performing conventional convolution operation on the transmission line inspection image to obtain the first feature map, the method further includes: performing separate convolution on the transmission line inspection image to obtain a second feature map, wherein each feature point in the second feature map incorporates features of the size of (2r-1) around it, wherein r=expansion rate.

7. The method for processing power line inspection images according to claim 5, characterized in that: The multi-scale feature enhancement decoding module processes the fused feature information, including: decoding the fused feature information multiple times, generating a feature compensation plane with a resolution of H through a 3×3 convolution operation for the feature layer obtained in each decoding, and splicing the feature compensation plane with the feature layer for the next decoding, and then splicing the final feature layer with the feature compensation planes retained in the decoding process to obtain the decoded feature information for restoring the high-resolution drizzle picture.

8. The method for processing power line inspection images according to claim 5, characterized in that: The adjacent information auxiliary prediction module processes the decoded feature information, including: original drizzle component prediction and adjacent information prediction for the decoded feature information; the original drizzle component prediction information is obtained by convolving the decoded feature information through a conventional 3*3 convolution layer, and the adjacent information prediction information is obtained by weighted summing {x0, x1..., x6, x7} through an adaptive attention mechanism, {x0, x1..., x6, x7} represents the correlation between a certain pixel point in the decoded feature information and its adjacent pixel points; the original drizzle component prediction information is fused with the adjacent information prediction information to obtain a corrected transmission line inspection image.

9. A transmission line inspection image processing system, characterized in that: include: Feature extraction module: configured to use the GCANet network to extract features from the input transmission line inspection image under drizzling weather to obtain multi-level feature information; Feature fusion module: configured to use a gated aggregation network to fuse multi-level feature information to obtain fused feature information; Decoding module: It is configured to use the multi-scale feature enhancement decoding module and the adjacent information auxiliary prediction module to replace the Decoder module in the GCANet network to process the fused feature information to obtain the corrected transmission line inspection image.

10. The power transmission line inspection image processing system according to claim 9, characterized in that: The feature extraction module performs conventional convolution operations on the transmission line inspection image to obtain a first feature map, fuses local pixel information of the transmission line inspection image with the first feature map to obtain a fused feature map, and performs hole convolution processing on the fused feature map to obtain multi-level feature information, wherein the multi-level feature information includes low-level feature information, mid-level feature information, and high-level feature information.

11. The power transmission line inspection image processing system according to claim 9, characterized in that: The feature fusion module uses a gated aggregation network to learn feature information of each level in the multi-level feature information to obtain the weight of the feature information of each level, and determines the fused feature information according to the feature information of each level and the corresponding weight.

12. The power transmission line inspection image processing system according to claim 9, characterized in that: The multi-scale feature enhancement decoding module processes the fused feature information to obtain decoded feature information for restoring the high-resolution drizzle picture, and the adjacent information auxiliary prediction module processes the decoded feature information to obtain a corrected transmission line inspection image.

13. The power transmission line inspection image processing system according to claim 10, characterized in that: It also includes a separation convolution module, which is configured to perform separation convolution on the transmission line inspection image to obtain a second feature map before performing a conventional convolution operation on the transmission line inspection image to obtain a first feature map. Each feature point in the second feature map is integrated with features of its surrounding (2r-1) size, where r = expansion rate.

14. The power transmission line inspection image processing system according to claim 12, characterized in that: The multi-scale feature enhancement decoding module processes the fused feature information, including: decoding the fused feature information multiple times, generating a feature compensation plane with a resolution of H through a 3×3 convolution operation for each feature layer obtained by decoding, splicing the feature compensation plane with the feature layer for the next decoding, and then splicing the final feature layer with each feature compensation plane retained in the decoding process to obtain decoded feature information for restoring a high-resolution drizzle picture; The adjacent information auxiliary prediction module processes the decoded feature information, including: performing original drizzle component prediction and adjacent information prediction on the decoded feature information; the original drizzle component prediction information is obtained by convolving the decoded feature information through a conventional 3*3 convolution layer, and the adjacent information prediction information is obtained by weighted summing {x0, x1..., x6, x7} through an adaptive attention mechanism, {x0, x1..., x6, x7} represents the correlation between a certain pixel point in the decoded feature information and its adjacent pixel points; the original drizzle component prediction information is fused with the adjacent information prediction information to obtain a corrected transmission line inspection image.

15. A computer device comprising one or more processors, wherein the processors store a program, characterized in that: When the program is executed by the processor, the power transmission line inspection image processing method according to any one of claims 1 to 8 is implemented.

16. A computer readable medium storing a computer program, characterized in that: When the computer program is executed, the power transmission line inspection image processing method according to any one of claims 1 to 8 is implemented.

Images