Edge end power transmission tower corrosion diagnosis method and system based on corrosion significance perception

Abstract

The invention belongs to the technical field of image data processing, and provides a corrosion saliency perception-based edge end power transmission tower corrosion diagnosis method and system, and the technical scheme is that based on an obtained power transmission tower corrosion image and a corresponding corrosion saliency thermodynamic diagram, a constructed pruning trunk network and a pruning controller are trained in a combined manner; obtaining a trained pruning trunk network and a pruning controller; and deploying the trained trunk network capable of pruning and the pruning controller on edge equipment, and performing reasoning according to a power transmission tower image acquired in real time and the deployed trunk network capable of pruning and the pruning controller to obtain a power transmission tower corrosion diagnosis result. Wherein the reasoning process comprises the steps of extracting saliency thermodynamic diagram characteristics according to an image of a power transmission tower collected in real time, outputting an optimal pruning decision according to the saliency thermodynamic diagram characteristics and a constructed pruning controller, dynamically updating a channel of a backbone network according to the optimal pruning decision, and outputting a final diagnosis result based on the backbone network after the channel is updated. Computing resources can be adjusted according to the complexity of the input image.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image data processing, and particularly relates to an edge tower corrosion diagnosis method and system based on corrosion saliency perception. BACKGROUND

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

[0003] As a key infrastructure of power transmission network, the structural health state of a power transmission tower is directly related to the safe and stable operation of the power grid. The steel material power transmission tower is exposed to harsh natural environment for a long time, and is prone to corrosion. Corrosion can cause the effective cross-section of the component to decrease and the bearing capacity to decrease, thereby causing structural failure and even tower collapse and other serious accidents. Therefore, efficient and accurate corrosion detection of the power transmission tower has great economic and safety significance.

[0004] In recent years, with the rapid development of computer vision and deep learning technology, image recognition methods based on convolutional neural networks (CNN) have gradually replaced traditional manual inspection and image processing technology, and have become the mainstream solution for automatic diagnosis of power transmission tower corrosion. These methods can automatically learn abstract features of corrosion from images by training deep models, and show excellent accuracy and robustness.

[0005] However, deploying advanced CNN models on edge devices (such as Jetson Xavier, Nano, etc.) with severely limited computing resources, storage space and power consumption to achieve real-time online diagnosis still faces great challenges. Currently, the technical routes adopted to solve this problem mainly focus on the following two aspects: 1. Using lightweight pre-designed network models: directly using lightweight network architectures such as MobileNet, ShuffleNet, SqueezeNet, etc. as backbone. These models reduce the parameter quantity and computational complexity (FLOPs) to a certain extent through depth separable convolution, channel shuffle, etc., so that they can barely run on edge devices.

[0006] 2. Using model compression technology: after training large models (such as ResNet, VGG), static model compression techniques such as weight quantization, knowledge distillation and static model pruning are used. These techniques can compress the model size and speed up inference to a certain extent, but their compression strategies are fixed and one-size-fits-all. Once the compression is completed, the structure and computational complexity of the model no longer change after deployment.

[0007] The above schemes have a common defect: they all use a "fixed" model structure to process all input samples. This static processing mode ignores the huge differences in the inherent complexity of input images in real application scenarios, leading to inefficient allocation of computing resources and bottlenecks in real-time performance, including: Waste of resources for simple samples: When faced with images with simple backgrounds, no rust, or rust areas that are significant and single, the fixed lightweight model still performs the entire computing process, resulting in a large amount of redundant computation. This causes inference delay and power waste, and cannot meet the real-time processing requirements of extremely high frame rates.

[0008] Insufficient performance for complex samples: When faced with complex images with cluttered backgrounds, weak rust features, scattered distribution, or poor lighting conditions, the fixed lightweight model is difficult to perform accurate feature extraction and recognition due to its limited model capacity and representation ability, resulting in a significant decrease in diagnostic accuracy.

[0009] Lack of adaptive ability: Existing static compression or lightweight design is an offline optimization that cannot dynamically adjust computing resources according to real-time content of input images, lacking intelligent and on-demand allocation of computing power. SUMMARY

[0010] To solve at least one of the technical problems in the above background art, the present application provides an edge power transmission tower corrosion diagnosis method and system based on corrosion saliency perception, which considers the differences in processing images in real application scenarios under complex scenarios, dynamically adjusts computing resources, and can allocate data processing on demand.

[0011] To achieve the above purpose, the present application adopts the following technical solutions: The first aspect of the present application provides an edge power transmission tower corrosion diagnosis method based on corrosion saliency perception, comprising the following steps: Obtain power transmission tower corrosion images and generate a corrosion saliency heat map corresponding to each image; Based on the obtained power transmission tower corrosion images and corresponding corrosion saliency heat maps, jointly train the constructed prunable backbone network and pruning controller to obtain the trained prunable backbone network and pruning controller; Deploy the trained prunable backbone network and pruning controller to an edge device, and perform inference based on the real-time collected power transmission tower images and the deployed prunable backbone network and pruning controller to obtain the power transmission tower corrosion diagnosis result, wherein the inference process includes: extracting saliency heat map features from the real-time collected power transmission tower images, outputting optimal pruning decisions based on the saliency heat map features and the constructed pruning controller, dynamically updating the channels of the backbone network based on the optimal pruning decisions, and outputting the final diagnosis result based on the channel updated backbone network.

[0012] Further, the obtained power transmission tower corrosion image and the corresponding corrosion saliency heat map are used to jointly train the constructed prunable backbone network and the pruning controller, and the joint training specifically includes: Statistical features are extracted based on the obtained power transmission tower corrosion image and the corresponding corrosion saliency heat map, and a state vector is constructed based on the statistical features; The constructed prunable backbone network is subjected to sparse training to obtain a prunable backbone network subjected to sparse training; The constructed state vector is taken as a state, and the pruning controller is trained in the environment of the prunable backbone network M subjected to pre-training and sparsification by using a reinforcement learning paradigm based on a set reward function.

[0013] Further, the extracted statistical features are used to construct the state vector, and the state vector includes a salient region proportion, an average saliency intensity, a saliency distribution entropy, and a maximum connected region proportion.

[0014] Further, the prunable backbone network is composed of a feature extraction network and a detection head, the prunable backbone network uses ResNet-18 as a basic skeleton, ResNet-18 is improved to adapt to channel pruning, a channel pruning gate is explicitly inserted after each residual block of ResNet-18, the gamma coefficient of each BN layer is directly associated with the output channel of the previous convolution layer to serve as an importance score of the channel, the detection head receives multi-scale feature maps from the feature extraction network, and outputs a bounding box coordinate, a confidence, and a corrosion category.

[0015] Further, the constructed prunable backbone network is subjected to sparse training, and the sparse training includes: An L1 regularization term for all BN layer gamma coefficients is additionally added to the loss function; A global distribution of all BN layer gamma values is counted, a pruning threshold is set, all channels with a gamma value lower than the pruning threshold are determined as redundant channels, and these channels and the corresponding convolution kernels in the next layer are physically removed from the network.

[0016] Further, the set reward function is: , wherein, is an accuracy reward, the detection accuracy of the pruned model on the current image, is an efficiency reward, and β is a set weight.

[0017] Further, the reward function based on the setting adopts the paradigm of reinforcement learning to train the pruning controller, including: extracting a state vector based on the backbone network that has been trained by sparsification, the controller making a decision according to the state, outputting a pruning ratio, temporarily pruning the backbone network according to the decision, and obtaining a final pruning strategy by maximizing the expected reward.

[0018] The second aspect of the present application provides an edge power transmission tower corrosion diagnosis system based on corrosion saliency perception, comprising: An image acquisition module is configured to acquire power transmission tower corrosion images and generate a corrosion saliency heat map corresponding to each image; An offline training module is configured to train the constructed pruning backbone network and pruning controller jointly based on the acquired power transmission tower corrosion images and corresponding corrosion saliency heat maps, and obtain a trained pruning backbone network and pruning controller; An online inference module is configured to deploy the trained pruning backbone network and pruning controller to an edge device, perform inference according to real-time collected power transmission tower images and the deployed pruning backbone network and pruning controller, and obtain a power transmission tower corrosion diagnosis result, wherein the inference process comprises: extracting saliency heat map features from the real-time collected power transmission tower images, outputting an optimal pruning decision according to the saliency heat map features and the constructed pruning controller, dynamically updating the channels of the backbone network according to the optimal pruning decision, and outputting a final diagnosis result based on the updated backbone network.

[0019] The third aspect of the present application provides a computer readable storage medium.

[0020] A computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the edge power transmission tower corrosion diagnosis method based on corrosion saliency perception as described above.

[0021] The fourth aspect of the present application provides a computer device.

[0022] A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of the edge power transmission tower corrosion diagnosis method based on corrosion saliency perception as described above.

[0023] Compared with the prior art, the present application has the following advantages: The application combines the characteristics of the power transmission tower corrosion image and the corrosion saliency heat map, trains the constructed prunable main network and the pruning controller jointly, adopts a dynamic adjustment strategy to continuously adjust the prunable main network, considers the huge difference in the internal complexity of the input image in the actual application scene, dynamically adjusts the computing resources, can allocate data processing on demand, improves the efficient use of computing resources, and avoids a large amount of redundant calculation.

[0024] Advantages of the additional aspects of the application will be partially given in the following description, partially will become apparent from the following description, or will be understood by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0025] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description of the exemplary embodiments of the application and the explanation thereof serve to explain the application, and do not constitute an improper limitation of the application.

[0026] Figure 1 The figure is a flow chart of the edge power transmission tower corrosion diagnosis method based on corrosion saliency perception provided by the embodiment of the application. DETAILED DESCRIPTION

[0027] The application will be further described below in conjunction with the drawings and embodiments.

[0028] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs.

[0029] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form, and in addition, it should be understood that when the terms "comprise" and / or "include" are used in the specification, they mean the presence of a feature, step, operation, device, component and / or combination thereof.

[0030] Embodiment one As shown in Figure 1 The embodiment provides an edge power transmission tower corrosion diagnosis method based on corrosion saliency perception, which comprises the following steps: Step 1: Obtain the power transmission tower corrosion image, and generate the corrosion saliency heat map corresponding to each image; In this embodiment, images of corrosion on transmission towers are acquired and finely annotated (e.g., pixel-level segmentation labels or bounding box labels). A pre-trained saliency detection model (e.g., U-Net) or traditional algorithms (e.g., FT, HC) is used to generate a corresponding saliency map for each image in the training set. This heatmap has the same size as the original image, and each pixel value represents the saliency (0~1) of corrosion at that location.

[0031] Step 2: Based on the acquired images of transmission tower corrosion and the corresponding corrosion saliency heatmap, the constructed pruning backbone network and pruning controller are jointly trained to obtain the trained pruning backbone network and pruning controller. Specifically, the steps include the following: Step 201: Extract statistical features based on the acquired images of corrosion on the transmission tower and the corresponding heat map of corrosion significance, and construct a state vector based on the statistical features; In this embodiment, in the corrosion saliency heatmap, each pixel value represents the degree of corrosion presence at that location. Instead of directly using a high-dimensional heatmap, compact and information-rich statistical features are extracted from it as a state vector s describing the "diagnostic difficulty" of the image. This state vector will serve as the input to the pruning controller.

[0032] Specifically, the extracted statistical features include the proportion of significant regions, the average intensity of significance, the significance distribution entropy, and the proportion of the largest connected regions; Among them, the proportion of significant areas The calculation formula is: , in, The pixels representing the rusted areas in the rust saliency heatmap. The set binarization threshold, This represents the total number of pixels; the higher the percentage of significant areas, the larger and more obvious the rusted area, and the more "simple" the image may be, tending towards pruning.

[0033] Among them, the average intensity of significance The calculation formula is: , in, It is greater than in heatmap M In the part where the average significance level is higher, it indicates that the model has a higher confidence level in the rusted area, and the image may be more "simple".

[0034] Among them, the saliency distribution entropy Calculate the normalized histogram p of the heatmap M, and then calculate the entropy: , wherein, is the probability value of the i interval in the histogram, a high entropy value indicates that the saliency is dispersed in many places, and the background may be cluttered; a low entropy value indicates that the saliency is concentrated, and the image is cleaner. Low-entropy images tend to be pruned.

[0035] wherein, the maximum connected region ratio: On the binary thermal map (M>threshold_T), the largest connected region is found, and the ratio of its area to the image area is calculated. A high value indicates that there is a dominant, continuous corrosion region; a low value indicates that the corrosion is scattered, making diagnosis more difficult. High values tend to be pruned.

[0036] Finally, the state vector constructed is represented as The complexity of the current image is effectively encoded by the state vector.

[0037] Step 202, sparse training of the constructed prunable backbone network to obtain a prunable backbone network M after sparse training; In this embodiment, the prunable backbone network aims to balance high precision and low latency, and its core consists of a feature extraction network and a detection head.

[0038] The prunable backbone network uses ResNet-18 as the basic skeleton, and improves ResNet-18 to adapt to channel pruning. After each residual block of ResNet-18, an explicit channel pruning gate is inserted. The pruning gate is not an additional complex module, but is realized by utilizing the inherent scaling factor (γ parameter) in the batch normalization layer. Specifically, the γ coefficient of each BN layer is directly associated with the output channels of the previous convolution layer as the importance score of the channel. These γ coefficients will be subject to sparse constraints during training, so that the network automatically learns the importance and redundancy of the channels. In order to complete the positioning and classification of the corrosion area, a lightweight YOLO detection head is used, such as based on YOLOv3 or YOLOX. The detection head receives multi-scale feature maps from the feature extraction network and outputs bounding box coordinates, confidence and corrosion class.

[0039] As a further implementation, the constructed prunable backbone network is subjected to sparse training, which includes: Specifically, a structured pruning method based on the scaling factor of the BN layer is used, and the specific steps are as follows: Sparse training: In the model training stage, in addition to the conventional classification and regression loss, an L1 regularization term for all BN layer γ coefficients is additionally added to the loss function. This forces the network to push the γ values corresponding to unimportant channels to zero during optimization.

[0040] In this embodiment, the norm criterion is used, such as L1-Norm, which is to calculate the sum of the absolute values of all weights of each convolution kernel (corresponding to an output channel) in a convolution layer. The smaller the weight norm (size) of a channel, the closer the activation value is to zero, and the smaller the contribution to the next layer, so it is less important, and the unimportant channel can be removed by setting a threshold.

[0041] Channel pruning: the global distribution of all BN layer gamma values is counted; a pruning threshold is set (for example, after sorting all gamma values, the value of the kth percentile is taken as the threshold), and all channels with gamma values below the threshold are determined as redundant channels. Subsequently, these channels and their corresponding convolution kernels in the next layer are physically removed from the network to form a smaller and more compact "subnetwork".

[0042] Pruning backbone network fine-tuning: pruning operation temporarily reduces the accuracy of the model, in order to restore its performance, a period of fine-tuning training is performed on the pruned lightweight model using training data. In this stage, the L1 regularization constraint is removed, and only the detection loss is used to allow the model weights to adapt to the new compact structure.

[0043] The loss function formula is: where λ is the sparsity penalty coefficient, used to control the strength of pruning.

[0044] Step 203, taking the constructed state vector as the state, based on the environment of the pre-trained and sparsified prunable backbone network M, using the reinforcement learning paradigm to train the pruning controller based on the set reward function; In this embodiment, the reinforcement learning paradigm is used to train the pruning controller, specifically, the feature vector s extracted from the saliency heat map in the environment of the pre-trained and sparsified prunable backbone network M is taken as the state, and the specific action is the pruning ratio a; The set reward function specifically includes the accuracy reward R_accuracy and the efficiency reward R_efficiency, and is expressed as: , wherein, R_accuracy is the accuracy reward, the detection accuracy of the pruned model on the current image can be measured by the confidence or IoU of the detection head, for example, the difference or relative value of mAP is used; R_efficiency is the efficiency reward, which encourages a higher pruning rate. R_efficiency = a (the action itself) or (1 - FLOPs_after_pruning / FLOPs_original) can be used, and β is the weight coefficient, which can be preset in advance.

[0045] Specific training process adopts policy gradient algorithm such as proximal policy optimization or REINFORCE, extracts state vector based on the backbone network which has been trained by sparsification, and the controller makes decisions according to the state and outputs pruning ratio, temporarily prunes the backbone network according to the decisions, and obtains the final pruning strategy by maximizing the expected reward.

[0046] Step 3: deploy the trained prunable backbone network and pruning controller to the edge device, perform inference according to the real-time collected transmission tower image and the deployed prunable backbone network and pruning controller, and obtain the transmission tower corrosion diagnosis result. The application constructs a cooperative recognition mode of "back-end server coarse positioning and edge device accurate judgment", the back-end server is responsible for performing the calculation-intensive corrosion saliency heat map generation and sparsification training process, and outputs the trained model and pruning strategy. The edge device deploys the finally obtained lightweight prunable backbone network. During inference, the device dynamically activates the important channels retained after pruning according to the features of the input image, and skips the redundant channels, thereby realizing adaptive and rapid inference.

[0047] In the embodiment, the inference according to the real-time collected transmission tower image and the deployed prunable backbone network and pruning controller comprises: According to the real-time collected transmission tower image, the saliency heat map features are extracted, the optimal pruning decision is output according to the saliency heat map features and the constructed pruning controller, the important channels corresponding to the backbone network are dynamically activated according to the optimal pruning decision, the unimportant channels are shielded, and finally the final diagnosis result is output through the backbone network.

[0048] Embodiment two The embodiment provides an edge end transmission tower corrosion diagnosis system based on corrosion saliency perception, comprising: An image acquisition module is configured to acquire transmission tower corrosion images and generate a corrosion saliency heat map corresponding to each image. An offline training module is configured to train the constructed prunable backbone network and pruning controller based on the acquired transmission tower corrosion images and corresponding corrosion saliency heat maps, and obtain the trained prunable backbone network and pruning controller. An online inference module is configured to deploy the trained prunable backbone network and pruning controller to an edge device, perform inference according to real-time collected transmission tower images and the deployed prunable backbone network and pruning controller, and obtain a transmission tower corrosion diagnosis result, wherein the inference process comprises: extracting saliency heat map features from real-time collected transmission tower images, outputting an optimal pruning decision according to the saliency heat map features and the constructed pruning controller, dynamically updating the channels of the backbone network according to the optimal pruning decision, and outputting a final diagnosis result based on the updated backbone network.

[0049] It should be noted that the specific implementation of the edge power transmission tower corrosion diagnosis system based on corrosion salience perception of the embodiments of the present application is similar to the specific implementation of the edge power transmission tower corrosion diagnosis method based on corrosion salience perception of the embodiments of the present application, and specific reference can be made to the description in the method part. In order to reduce redundancy, this part will not be repeated here.

[0050] Embodiment three The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps in the edge power transmission tower corrosion diagnosis method based on corrosion salience perception.

[0051] Embodiment four The embodiment provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor realizes the steps in the edge power transmission tower corrosion diagnosis method based on corrosion salience perception when executing the program.

[0052] Embodiment five The embodiment provides a program product, which is a computer program product, including a computer program, and the computer program is executed by a processor to realize the steps in the edge power transmission tower corrosion diagnosis method based on corrosion salience perception.

[0053] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer usable program code.

[0054] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The function of the device specified in one flow or multiple flows and / or blocks Figure 1 The function of the device specified in one flow or multiple flows and / or blocks

[0055] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 one or more flow or block Figure 1 one or more flow or block

[0056] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions executed on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 one or more flow or block Figure 1 one or more flow or block

[0057] Those of ordinary skill in the art can understand that all or part of the flow of the above-mentioned embodiment method can be completed by a computer program instructing relevant hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the flow of the above-mentioned embodiment method. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM), a random access memory (RAM), or the like.

[0058] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those of ordinary skill in the art can make various changes and modifications to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A corrosion diagnosis method for edge-end transmission towers based on corrosion saliency perception, characterized in that, Includes the following steps: Acquire images of corrosion on transmission towers and generate a heat map of corrosion saliency for each image; Based on the acquired images of transmission tower corrosion and the corresponding corrosion saliency heatmaps, the constructed pruning backbone network and pruning controller are jointly trained to obtain the trained pruning backbone network and pruning controller. The trained pruning backbone network and pruning controller are deployed to edge devices. Based on real-time acquired images of the transmission tower and the deployed pruning backbone network and pruning controller, inference is performed to obtain the corrosion diagnosis results of the transmission tower. The inference process includes: extracting salient heat map features from the real-time acquired images of the transmission tower; outputting the optimal pruning decision based on the salient heat map features and the constructed pruning controller; dynamically updating the channels of the backbone network based on the optimal pruning decision; and outputting the final diagnosis result based on the updated backbone network.

2. The corrosion diagnosis method for edge-end transmission towers based on corrosion saliency perception as described in claim 1, characterized in that, The method involves jointly training the constructed pruning backbone network and pruning controller based on the acquired images of transmission tower corrosion and the corresponding corrosion saliency heatmaps, specifically including: Statistical features are extracted from the acquired images of corrosion on the transmission towers and the corresponding corrosion saliency heatmaps, and a state vector is constructed based on the statistical features. The constructed pruning backbone network is subjected to sparse training to obtain the sparsely trained pruning backbone network. Using the constructed state vector as the state, and based on a pre-trained and sparse pruning backbone network M, a reinforcement learning paradigm is used to train the pruning controller according to the set reward function.

3. The corrosion diagnosis method for edge-end transmission towers based on corrosion saliency perception as described in claim 2, characterized in that, The extracted statistical features are used to construct a state vector, which includes the proportion of salient regions, the average strength of salientity, the salient distribution entropy, and the proportion of the largest connected regions.

4. The corrosion diagnosis method for edge-end transmission towers based on corrosion saliency perception as described in claim 1, characterized in that, The pruning backbone network consists of a feature extraction network and a detection head. The pruning backbone network uses ResNet-18 as the basic skeleton and improves ResNet-18 to adapt to channel pruning. After each residual block of ResNet-18, a channel pruning gate is explicitly inserted, which directly associates the γ coefficient of each BN layer with the output channel of the previous convolutional layer as the importance score of that channel. The detection head receives multi-scale feature maps from the feature extraction network and outputs bounding box coordinates, confidence scores, and corrosion categories.

5. The corrosion diagnosis method for edge-end transmission towers based on corrosion saliency perception as described in claim 1, characterized in that, The constructed pruning backbone network is then subjected to sparse training, including: An L1 regularization term for the γ coefficients of all BN layers was added to the loss function; The global distribution of γ values ​​in all BN layers is statistically analyzed. A pruning threshold is set, and all channels with γ values ​​below the pruning threshold are identified as redundant channels. These channels and their corresponding convolutional kernels in the next layer are then physically removed from the network.

6. The corrosion diagnosis method for edge-end transmission towers based on corrosion saliency perception as described in claim 1, characterized in that, The reward function is defined as follows: ， in, As a precision bonus, the pruned model's detection accuracy on the current image is... β is the set weight for the efficiency reward.

7. The corrosion diagnosis method for edge-end transmission towers based on corrosion saliency perception as described in claim 1, characterized in that, The pruning controller is trained using a reinforcement learning paradigm based on a set reward function. This includes: extracting state vectors from a sparsely trained backbone network; the controller making decisions based on the states and outputting a pruning ratio; temporarily pruning the backbone network based on the decisions; and obtaining the final pruning strategy by maximizing the expected reward.

8. A corrosion diagnosis system for edge-end transmission towers based on corrosion saliency perception, characterized in that, include: The image acquisition module is used to acquire images of corrosion on the transmission tower and generate a heat map of corrosion salience for each image. The offline training module is used to jointly train the constructed pruning backbone network and pruning controller based on the acquired images of transmission tower corrosion and the corresponding corrosion saliency heat map, so as to obtain the trained pruning backbone network and pruning controller. The online inference module is used to deploy the trained pruning backbone network and pruning controller to edge devices. It performs inference based on real-time acquired images of the transmission tower and the deployed pruning backbone network and pruning controller to obtain the transmission tower corrosion diagnosis results. The inference process includes: extracting salient heat map features from the real-time acquired images of the transmission tower; outputting the optimal pruning decision based on the salient heat map features and the constructed pruning controller; dynamically updating the channels of the backbone network based on the optimal pruning decision; and outputting the final diagnosis result based on the updated backbone network.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the corrosion diagnosis method for edge-end transmission towers based on corrosion saliency perception as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the edge-end transmission tower corrosion diagnosis method based on corrosion saliency perception as described in any one of claims 1-7.

Images