Transmission line group tower corrosion detection method, terminal and medium

Abstract

The invention discloses a transmission line group tower corrosion detection method, a terminal and a medium. The method comprises the following steps: constructing an original transmission line group tower corrosion image data set; constructing a transmission line group tower corrosion data set containing labeling information, and dividing the data set into a training set, a verification set and a test set; a hierarchical embedded coding and decoding network is constructed and trained; the hierarchical embedded encoding and decoding network structurally comprises a backbone network, a hierarchical embedded encoder and a multi-level decoder, and performing corrosion detection on the assembled tower by using a hierarchical embedded coding and decoding network. The method has the advantages that maintainers are helped to efficiently and accurately maintain the transmission line group tower corrosion area, the workload of the maintainers is relieved, and the safety accident risk is reduced.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular to a method, terminal and medium for detecting corrosion of transmission line towers. Background Art

[0002] With the rapid development of society, the power system has become one of the basic infrastructures of modern society. As a key component of power transmission in the power system, the safe operation of the transmission line tower is crucial to ensuring the stability of power supply and the sustainable development of social economy. However, due to long-term operation in the outdoor environment, the anti-rust coating on the steel structure body, connecting bolts and various hardware parts in the transmission line tower will gradually age and fall off, leading to corrosion problems. If the rusted parts of the transmission line tower are not handled in time, it will not only shorten the service life of the tower, but may also cause large-scale power outages or even personnel safety accidents, thus threatening the stability of the transmission system and the safety of the power network system. Regular corrosion detection of transmission line towers can greatly reduce the risk of the above accidents.

[0003] Traditional methods for detecting corrosion of transmission line towers mainly rely on manual inspections or surface inspections using specific instruments. However, these methods have problems such as low efficiency, high cost, and high safety risks, and are difficult to meet the corrosion detection needs of a large number of transmission line towers that need to be continuously maintained. In recent years, with the rapid development of computer vision and artificial intelligence technology, a new solution has been brought to the corrosion detection of transmission line towers. By using a dedicated drone with a high-resolution camera and combining it with a deep learning target detection algorithm, all-round and efficient corrosion detection of transmission line towers can be achieved. This detection method based on visual perception has the advantages of non-contact, strong real-time performance, and high precision. It can greatly improve the accuracy and efficiency of corrosion detection, reduce labor costs, and ensure the safety and reliability of power transmission. However, most of the existing general deep learning target detection algorithms require prior processing (such as non-maximum suppression), which makes it difficult to adjust the parameters and train the model. In addition, the recognition rate of general target detection algorithms for irregular corrosion and micro-corrosion is still low, which is difficult to meet the accuracy requirements of transmission line tower corrosion detection. Summary of the invention

[0004] The present application provides a transmission line tower corrosion detection method, terminal and medium, which have the advantage of helping maintenance personnel to efficiently and accurately perform maintenance on the rusted areas of transmission line towers, reducing their workload and reducing the risk of safety accidents.

[0005] The technical solution of this application is as follows:

[0006] On the one hand, the present application provides a method for detecting corrosion of a transmission line tower, comprising the following steps:

[0007] Step 1: Construct an original transmission line tower corrosion image dataset;

[0008] Step 2: Construct a transmission line tower corrosion dataset with labeled information and divide it into training set, validation set, and test set;

[0009] Step 3: construct a hierarchical chimeric codec network and train it; wherein the structure of the hierarchical chimeric codec network includes: a backbone network, a hierarchical chimeric encoder and a multi-level decoder;

[0010] Step 4: Apply the hierarchical chimeric codec network to perform rust detection on the tower group.

[0011] Furthermore, step 1 includes the following steps:

[0012] Step 11: Use a drone to collect images of the transmission line towers. During the collection process, the drone takes pictures of each tower from various angles;

[0013] Step 12: Perform image enhancement on the acquired image.

[0014] Furthermore, step 2 includes the following steps:

[0015] Step 21: annotate the corroded area in the original transmission line tower corrosion image dataset to obtain an annotation file containing target frame information, and any image is matched with a corresponding annotation file; the annotation file includes category information and position information of the target frame, wherein the position information is composed of four data, namely, the horizontal coordinate of the center point of the target frame, the vertical coordinate of the center point of the target frame, the width of the target frame, and the height of the target frame; the original transmission line tower corrosion image dataset and the corresponding annotation file constitute a transmission line tower corrosion image dataset containing annotation information;

[0016] Step 22: Randomly divide the transmission line tower corrosion image dataset into a training set, a validation set, and a test set.

[0017] Furthermore, in the hierarchical embedded codec network:

[0018] The backbone network is a feature extraction network of the CNN architecture, which is used to extract compressed feature maps of the original image at different scales;

[0019] The hierarchical embedding encoder extracts global information from compressed feature maps at different scales. The hierarchical embedding encoder takes differentiated data operations for feature maps of different sizes, specifically: embedding the smallest feature map containing the richest global information within the layer, and then embedding the output of the intra-layer embedding with the selected larger feature map between layers, and finally obtaining the output of the hierarchical embedding encoder.

[0020] The multi-level decoder generates the output of the predicted target box, including category information and location information. The category information is whether it is a rusted area, and the location information is the horizontal and vertical coordinates of the center of the target box and the width and height of the target box. The multi-level decoder takes the randomly initialized learnable tensor and the encoder output with global information as input, and obtains the output box information through multi-level self-decoupling operations and cross-domain decoupling operations. The length of the learnable tensor is the number of network output prediction boxes.

[0021] Further, the hierarchical embedding encoder includes an intra-layer embedding module and an inter-layer embedding module;

[0022] The minimum feature map tensor is sent to the intra-layer embedding module to obtain the intra-layer embedding output of the corresponding layer. The intra-layer embedding module multiplies the tensor by three different parameter matrices to obtain the query matrix Q, the key matrix K and the value matrix V. Then, the query matrix Q and the key matrix K are multiplied to obtain the attention matrix. The attention matrix is ​​then normalized by the Softmax function, and the normalized attention matrix is ​​multiplied by the value matrix V to obtain the intra-layer embedding output with the same size as the input of the intra-layer embedding module. The calculation process of the intra-layer embedding output is expressed by the formula:

[0023]

[0024] in, represents the embedded output within the layer, Q, K, V represent the query matrix, key matrix and value matrix respectively, C t Represents the dimension of the input tensor;

[0025] The intra-layer mosaic output is converted into a mosaic feature map through dimensional transformation, and used together with the other two feature maps as the input of the inter-layer mosaic module to obtain the inter-layer mosaic output, i.e., the encoder output;

[0026] The inter-layer mosaic module includes an upsampling unit, a downsampling unit, a splicing unit and a C3 unit, wherein the C3 unit is a data processing structure composed of a cascade of convolution, batch normalization and Silu activation functions.

[0027] Furthermore, after completing the construction of the hierarchical chimeric codec network, the predicted target box output by the network is bipartite graph matched with the true target box using the Hungarian algorithm, and a number of predicted target boxes corresponding to the true target box are selected; then, the classification loss and position loss between the selected prediction box and the true box are calculated, and the two are added to form the total loss function; among them, the classification loss uses the cross entropy loss, and its expression is as follows:

[0028]

[0029] Among them, Loss cls represents the classification loss, y represents the true box, represents the prediction box, N represents the total number of categories, c i Indicates the current category;

[0030] The position loss uses IoU loss, which is expressed as follows:

[0031]

[0032] Among them, Loss obj represents the position loss, y represents the real box, represents the predicted box, IoU represents the intersection-over-union ratio between the real box and the predicted box, ρ represents the Euclidean distance between the center points of the real box and the predicted box, and k represents the diagonal distance of the minimum closure area that contains both the predicted box and the real box;

[0033] Through the gradient descent algorithm, the parameters in the network are continuously iteratively updated, so that the loss function value is continuously reduced.

[0034] Furthermore, step 4 includes: configuring the trained hierarchical chimeric codec network to the terminal, detecting the image of the transmission line tower to be inspected taken by a dedicated drone in the terminal, and obtaining an output image containing a rusted target frame.

[0035] On the other hand, the present application provides a transmission line tower corrosion detection terminal, including a memory and a processor, wherein the memory stores a computer-readable program, and when the computer-readable program is called and executed by the processor, the transmission line tower corrosion detection method as described above is implemented.

[0036] On the other hand, the present application provides a computer-readable medium, in which a computer-readable program is stored. When the computer-readable program is called and executed by a computer, the transmission line tower corrosion detection method as described above is implemented.

[0037] In summary, the beneficial effects of this application are:

[0038] 1. The present invention proposes an end-to-end encoding and decoding network and applies it to tower corrosion detection;

[0039] Most of the existing rust detection algorithms use prior anchor boxes and non-maximum suppression, both of which rely on human prior knowledge. For irregular rust and tiny rust, it is often difficult to define a suitable prior anchor box. At the same time, non-maximum suppression will also make the model complicated, making it difficult to adjust hyperparameters and optimize. The hierarchical chimeric codec network proposed in the present invention is an end-to-end network. There is no need to deliberately define a priori anchor boxes for rust images before network training, nor is there any need to perform non-maximum suppression on redundant prediction boxes. The network uses the global modeling capability of the hierarchical chimeric encoder and the step-by-step decoupling capability of the multi-level decoder to regard rust detection as a collective prediction problem of rust categories and rust locations. It directly outputs quantitative prediction boxes from the original input image, which simplifies the input and output form of the network and has the characteristics of high efficiency and simplicity.

[0040] 2. The present invention proposes a hierarchical mosaic encoder and a multi-level decoder. In the hierarchical mosaic encoder, only the smallest feature map containing the richest global information is mosaicked within the layer, and then the output after the intra-layer mosaic is mosaicked with the selected larger feature map between layers. This differentiated data operation can obtain rich deep global information in the feature map while introducing a small amount of computation, thereby enhancing the performance of the encoder. In the multi-level decoder, self-decoupling operations and cross-domain decoupling operations respectively realize self-learning of learnable sequences and interactive learning with encoder outputs with global information, and the design of the multi-level structure also makes the set of rusted target frames that the network finally predicts and outputs more reliable and robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of the method for detecting corrosion of a transmission line tower assembly according to the present invention;

[0042] Figure 2 A schematic diagram of a hierarchical mosaic codec network in an embodiment of the present invention;

[0043] Figure 3 Schematic diagram of a hierarchical mosaic encoder in an embodiment of the present invention;

[0044] Figure 4 Schematic diagram of a multi-stage decoder in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The specific implementation of the present application is described in detail below with reference to the accompanying drawings.

[0046] A specific embodiment of the present application provides a method for detecting corrosion of a transmission line tower. Figure 1 As shown, the following steps are included:

[0047] Step 1: Construct an original transmission line tower corrosion image dataset. Step 1 includes the following steps:

[0048] Step 11: The dedicated drone performs systematic, high-resolution image acquisition of the transmission line towers. During the acquisition process, the drone takes pictures of each tower from various angles to ensure maximum coverage of the towers.

[0049] Step 12: Enhance the collected images. In order to make the network trained in the subsequent process more robust, the collected image set needs to be enhanced so that it can be expanded on the basis of the original data. Data enhancement methods include geometric transformation enhancement and pixel transformation enhancement. Among them, geometric transformation enhancement includes flipping, rotating, cropping, scaling and translating the original image. Pixel transformation enhancement includes adding salt and pepper noise or Gaussian noise to the original image, adjusting HSV contrast, brightness, saturation and white balance. After geometric transformation enhancement and pixel transformation enhancement are performed on the collected image set, the newly added pictures after the transformation are added to the collected image set to form the original transmission line tower corrosion image dataset.

[0050] Step 2: Construct a transmission line tower corrosion dataset with labeled information and divide it into training set, validation set, and test set. Step 2 includes the following steps:

[0051] Step 21: The technicians annotate the corroded areas in the original transmission line tower corrosion image dataset to obtain an annotation file containing target frame information. Any image is matched with a corresponding annotation file; the annotation file includes the category information and position information of the target frame, wherein the position information is composed of four data, namely, the horizontal coordinate of the center point of the target frame, the vertical coordinate of the center point of the target frame, the width of the target frame, and the height of the target frame; the original transmission line tower corrosion image dataset and the corresponding annotation file constitute a transmission line tower corrosion image dataset containing annotation information.

[0052] Step 22: Randomly divide the transmission line tower corrosion image dataset into training set, validation set and test set to ensure that the training and evaluation of the network model can be carried out on different data to enhance the credibility of the network model.

[0053] Step 3: Construct a hierarchical chimeric codec network and train it; Figure 2 As shown, the structure of the hierarchical chimeric codec network includes: a backbone network, a hierarchical chimeric encoder and a multi-level decoder.

[0054] In the hierarchical chimeric codec network:

[0055] The backbone network is a feature extraction network of the CNN architecture, which is used to extract compressed feature maps of the original image at different scales, thereby facilitating efficient interaction of deep information in subsequent codecs.

[0056] The hierarchical embedding encoder extracts global information from compressed feature maps at different scales to assist the subsequent decoder in making accurate predictions. The hierarchical embedding encoder performs differentiated data operations on feature maps of different sizes, specifically: embedding the smallest feature map containing the richest global information within a layer, then embedding the output after the embedding within a layer with the selected larger feature map between layers, and finally obtaining the hierarchical embedding encoder output; the hierarchical embedding encoder further learns the deep coupling information in the output feature map of the backbone network through inter-layer data embedding. At the same time, since only the intra-layer data embedding is applied to the smallest feature map, the rich deep global information in the feature map is obtained with a small amount of computation, which greatly enhances the performance of the encoder.

[0057] The multi-level decoder generates the output of the predicted target box, including category information and location information. The category information is whether it is a rusted area, and the location information is the horizontal and vertical coordinates of the center of the target box and the width and height of the target box. The multi-level decoder takes the randomly initialized learnable tensor and the encoder output with global information as input, and obtains the output box information through multi-level self-decoupling operations and cross-domain decoupling operations. The length of the learnable tensor is the number of network output prediction boxes.

[0058] The backbone network selects the DarkNet-53 architecture. The input image size is 640×640×3. After the first channel transformation of the backbone network, the size becomes 640×640×32. The 3rd, 4th, and 5th layers of the DarkNet-53 architecture are selected as the input of the hierarchical chimeric encoder. The sizes of the feature maps of these three layers are 80×80×256, 40×40×512, and 20×20×1024, respectively.

[0059] It should be noted that before the minimum feature map containing high-level features of size 20×20×1024 is fed into the intra-layer embedding module, it needs to be converted into a minimum feature map tensor of size 400×1024 through dimensionality transformation.

[0060] The hierarchical embedding encoder includes an intra-layer embedding module and an inter-layer embedding module, and its structure is as follows: Figure 3 shown.

[0061] First, the minimum feature map tensor is sent to the intra-layer embedding module to obtain the intra-layer embedding output of the corresponding layer. The intra-layer embedding module multiplies the tensor by three different parameter matrices of size 1024×1024 to obtain the query matrix Q, the key matrix K and the value matrix V. Then, the query matrix Q and the key matrix K are multiplied to obtain the attention matrix, and then the attention matrix is ​​normalized by the Softmax function. Then, the normalized attention matrix is ​​multiplied by the value matrix V to obtain the intra-layer embedding output with the same size as the input of the intra-layer embedding module; the calculation process of the intra-layer embedding output is expressed by the formula:

[0062]

[0063] in, represents the embedded output within the layer, Q, K, V represent the query matrix, key matrix and value matrix respectively, C t Represents the dimensions of the input tensor.

[0064] The intra-layer embedding module can learn the correlation between tensor values ​​at different positions from the data interaction of tensors, thereby obtaining rich deep global information in the smallest feature map. The intra-layer embedding output of size 400×1024 is converted back to a embedding feature map of size 20×20×1024 through dimensional transformation, and is used together with the other two feature maps of size 40×40×512 and 80×80×256 as the input of the inter-layer embedding module to obtain the inter-layer embedding output, i.e., the encoder output.

[0065] The inter-layer mosaic module includes an upsampling unit, a downsampling unit, a splicing unit and a C3 unit, among which upsampling and downsampling can expand and reduce the scale of the feature map, thereby ensuring the scale alignment requirements of subsequent feature map splicing; splicing is a common operation for fusing different feature maps, that is, stacking different feature maps of the same size in the channel direction to form a new spliced ​​feature map; the C3 unit is a data processing structure composed of convolution, batch normalization and Silu activation functions cascaded in sequence. This unit can efficiently perform feature mosaicking on the spliced ​​feature map and specify the number of output channels as 256.

[0066] First, the inter-layer mosaic module upsamples the mosaic feature map of size 20×20×1024 to 40×40×1024, then concatenates it with the input feature map of size 40×40×512, and then performs feature mosaicking through the C3 unit to obtain an upsampled mosaic feature map of size 40×40×256.

[0067] The upsampled mosaic feature map is upsampled again, concatenated with the input feature map of size 80×80×256 and passed through the C3 unit to obtain an upsampled mosaic feature map of size 80×80×256. The upsampled mosaic feature map of size 80×80×256 is then downsampled to 40×40×256, then concatenated with the upsampled mosaic feature map of the same size, and then passed through the C3 unit for feature concatenation to obtain a downsampled mosaic feature map of size 40×40×256. The downsampled mosaic feature map is downsampled again, concatenated with the mosaic feature map of size 20×20×1024 and passed through the C3 unit to obtain an inter-layer mosaic output feature map of size 20×20×256. Finally, in order to facilitate the feeding of data into the subsequent decoder, the feature map is dimensionally transformed to obtain a feature map tensor of size 400×256, which is the output of the hierarchical mosaic encoder.

[0068] The structure of the multi-stage decoder is as follows Figure 4 As shown in , its input includes a set of randomly initialized learnable tensors and the output tensor of the hierarchical chimeric encoder. Among them, the randomly initialized learnable tensor represents the randomly initialized target box information, the length of the tensor represents the number of predicted target boxes, and the dimension of the sequence represents the learnability of the target box. The length of the tensor can be specified arbitrarily, and the dimension of the sequence needs to be aligned with the dimension of the output tensor of the hierarchical chimeric encoder, that is, the length of the learnable tensor is set to 60, and its dimension is fixed to 256.

[0069] First, the initialized learnable tensor is multiplied by three different parameter matrices of size 256×256 to obtain the query matrix Q, key matrix K and value matrix V, and the query matrix Q and the key matrix K are multiplied to obtain the attention matrix. The attention matrix is ​​then normalized by the Softmax function and multiplied by the value matrix V to obtain the self-decoupling tensor. The self-decoupling tensor is then residually connected and batch normalized with the initialized learnable tensor to obtain the residual self-decoupling tensor.

[0070] The residual self-decoupling tensor and the output tensor of the hierarchical chimeric encoder are multiplied by three different parameter matrices of size 256×256. The former obtains the query matrix Q, and the latter obtains the key matrix K and the value matrix V. Then, according to the steps of multiplying Q and K, Softmax normalization, multiplying with V, residual connection and batch normalization, the cross-domain decoupling tensor can be obtained. If the multi-level decoder contains 3 levels, the cross-domain decoupling tensor is used to replace the original initialized learnable tensor from the second level, and the output tensor of the hierarchical chimeric encoder is used equally at each level. Then the above steps are repeated 3 times, and the cross-domain decoupling tensor of the third level is the final prediction tensor. The prediction tensor is flattened and passed through two multi-layer perceptrons (MLPs) respectively, and the category information and position information of 60 predicted target boxes can be obtained respectively. The category information includes rust and background, and the position information includes the horizontal coordinate of the center point of the target box, the vertical coordinate of the center point of the target box, the width of the target box and the height of the target box.

[0071] After completing the construction of the hierarchical chimeric codec network, the Hungarian algorithm is used to perform bipartite graph matching on the 60 predicted target boxes output by the network and the true target boxes, and the predicted boxes corresponding to the true target boxes in the 60 predicted target boxes are selected; then, the classification loss and position loss between the selected prediction boxes and the true boxes are calculated, and the two are added to form the total loss function; among them, the classification loss uses the cross entropy loss, and its expression is as follows:

[0072]

[0073] Among them, Loss cls represents the classification loss, y represents the true box, represents the prediction box, N represents the total number of categories, c i Indicates the current category;

[0074] The position loss uses IoU loss, which is expressed as follows:

[0075]

[0076] Among them, Loss obj represents the position loss, y represents the real box, represents the predicted box, IoU represents the intersection-over-union ratio between the real box and the predicted box, ρ represents the Euclidean distance between the center points of the real box and the predicted box, and k represents the diagonal distance of the minimum closure area that contains both the predicted box and the real box;

[0077] Through the gradient descent algorithm, the parameters in the network are continuously iterated and updated, so that the loss function value is continuously reduced. When the loss function value hardly changes or the change value is less than a set threshold, it means that the prediction model has converged, the training process is over, and the model parameters are fixed.

[0078] Step 4: Apply the hierarchical chimeric codec network to perform rust detection on the tower group.

[0079] The trained hierarchical chimeric codec network is configured to the terminal, and the image of the transmission line tower to be inspected taken by a dedicated drone is detected in the terminal to obtain an output image containing the rusted target frame.

[0080] Another specific embodiment of the present application provides a transmission line tower corrosion detection terminal, including a memory and a processor, wherein the memory stores a computer-readable program, and when the computer-readable program is called and executed by the processor, the transmission line tower corrosion detection method as described above is implemented.

[0081] Another specific embodiment of the present application provides a computer-readable medium, in which a computer-readable program is stored. When the computer-readable program is called and executed by a computer, the transmission line tower group corrosion detection method as described above is implemented.

[0082] The parts not involved in the present invention are the same as the prior art or are implemented by using the prior art.

[0083] The above is only a preferred implementation of the present application. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present application, and these all fall within the scope of protection of the present application.

Claims

1. A method for detecting corrosion of a transmission line tower, characterized in that: The following steps are involved: Step 1: Construct an original transmission line tower corrosion image dataset; Step 2: Construct a transmission line tower corrosion dataset with labeled information and divide it into training set, validation set, and test set; Step 3: construct a hierarchical chimeric codec network and train it; wherein the structure of the hierarchical chimeric codec network includes: a backbone network, a hierarchical chimeric encoder and a multi-level decoder; Step 4: Apply the hierarchical chimeric codec network to perform rust detection on the tower group.

2. The method for detecting corrosion of a transmission line tower according to claim 1, characterized in that: Step 1 includes the following steps: Step 11: Use a drone to collect images of the transmission line towers. During the collection process, the drone takes pictures of each tower from various angles; Step 12: Perform image enhancement on the acquired image.

3. The method for detecting corrosion of a transmission line tower according to claim 1, characterized in that: Step 2 includes the following steps: Step 21: annotate the corroded area in the original transmission line tower corrosion image dataset to obtain an annotation file containing target frame information, and any image is matched with a corresponding annotation file; the annotation file includes category information and position information of the target frame, wherein the position information is composed of four data, namely, the horizontal coordinate of the center point of the target frame, the vertical coordinate of the center point of the target frame, the width of the target frame, and the height of the target frame; the original transmission line tower corrosion image dataset and the corresponding annotation file constitute a transmission line tower corrosion image dataset containing annotation information; Step 22: Randomly divide the transmission line tower corrosion image dataset into a training set, a validation set, and a test set.

4. The method for detecting corrosion of a transmission line tower according to claim 1, characterized in that: In the hierarchical chimeric codec network: The backbone network is a feature extraction network of the CNN architecture, which is used to extract compressed feature maps of the original image at different scales; The hierarchical embedding encoder extracts global information from compressed feature maps at different scales. The hierarchical embedding encoder takes differentiated data operations for feature maps of different sizes, specifically: embedding the smallest feature map containing the richest global information within the layer, and then embedding the output of the intra-layer embedding with the selected larger feature map between layers, and finally obtaining the output of the hierarchical embedding encoder. The multi-level decoder generates the output of the predicted target box, including category information and location information. The category information is whether it is a rusted area, and the location information is the horizontal and vertical coordinates of the center of the target box and the width and height of the target box. The multi-level decoder takes the randomly initialized learnable tensor and the encoder output with global information as input, and obtains the output box information through multi-level self-decoupling operations and cross-domain decoupling operations. The length of the learnable tensor is the number of network output prediction boxes.

5. The method for detecting corrosion of a transmission line tower according to claim 1, characterized in that: The hierarchical embedding encoder includes an intra-layer embedding module and an inter-layer embedding module; The minimum feature map tensor is sent to the intra-layer embedding module to obtain the intra-layer embedding output of the corresponding layer. The intra-layer embedding module multiplies the tensor by three different parameter matrices to obtain the query matrix Q, the key matrix K and the value matrix V. Then, the query matrix Q and the key matrix K are multiplied to obtain the attention matrix. The attention matrix is ​​then normalized by the Softmax function, and the normalized attention matrix is ​​multiplied by the value matrix V to obtain the intra-layer embedding output with the same size as the input of the intra-layer embedding module. The calculation process of the intra-layer embedding output is expressed by the formula: in, represents the embedded output within the layer, Q, K, V represent the query matrix, key matrix and value matrix respectively, C t Represents the dimension of the input tensor; The intra-layer mosaic output is converted into a mosaic feature map through dimensional transformation, and used together with the other two feature maps as the input of the inter-layer mosaic module to obtain the inter-layer mosaic output, i.e., the encoder output; The inter-layer mosaic module includes an upsampling unit, a downsampling unit, a splicing unit and a C3 unit, wherein the C3 unit is a data processing structure composed of a cascade of convolution, batch normalization and Silu activation functions.

6. The method for detecting corrosion of a transmission line tower according to claim 1, characterized in that: After completing the construction of the hierarchical chimeric codec network, the predicted target box output by the network is bipartite graph matched with the true target box using the Hungarian algorithm, and several predicted target boxes corresponding to the true target boxes are selected; then, the classification loss and position loss between the selected prediction box and the true box are calculated, and the two are added to form the total loss function; among them, the classification loss uses the cross entropy loss, and its expression is as follows: Among them, Loss cls represents the classification loss, y represents the true box, represents the prediction box, N represents the total number of categories, c i Indicates the current category; The position loss uses IoU loss, which is expressed as follows: Among them, Loss obj represents the position loss, y represents the real box, represents the predicted box, IoU represents the intersection-over-union ratio between the real box and the predicted box, ρ represents the Euclidean distance between the center points of the real box and the predicted box, and k represents the diagonal distance of the minimum closure area that contains both the predicted box and the real box; Through the gradient descent algorithm, the parameters in the network are continuously iteratively updated, so that the loss function value is continuously reduced.

7. The method for detecting corrosion of a transmission line tower according to claim 1, characterized in that: Step 4 includes: configuring the trained hierarchical chimeric codec network to the terminal, detecting the image of the transmission line tower to be inspected taken by a dedicated drone in the terminal, and obtaining an output image containing a rusted target frame.

8. A transmission line tower corrosion detection terminal, characterized in that: It includes a memory and a processor, the memory stores a computer-readable program, and when the computer-readable program is called and executed by the processor, the transmission line tower group corrosion detection method according to any one of claims 1-7 is implemented.

9. A computer-readable medium, characterized in that The computer-readable medium stores a computer-readable program, and when the computer-readable program is called and executed by a computer, the method for detecting corrosion of a transmission line tower as described in any one of claims 1 to 7 is implemented.

Images