System and method for detecting crimping pipe of power transmission line under complex field background

Abstract

The invention discloses a power transmission line crimping pipe detection system and method under a complex field background. The system is composed of a visible light image acquisition device and a crimping tube visible light image target detection device. The visible light image acquisition device is responsible for acquiring a visible light image of the crimping pipe and inputting the visible light image into the crimping pipe visible light image target detection device. The detection device is internally provided with a power transmission line crimping pipe detection model specially aiming at a field complex background, and is used for accurately detecting the crimping pipe. The detection model is based on the idea of a YOLO algorithm, but adopts a brand new backbone network, a feature fusion network and a wide area detection head. In order to solve the problem that a traditional network model is large in size, an MCBSC-C module and a GBSC module are introduced, a lightweight backbone network is constructed, and effective features can be rapidly extracted. The feature fusion network carries out multi-scale feature fusion through a GBSCSPPF module, and focuses on important feature maps in a concentrated manner through a CSA module.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a system and method for detecting transmission line crimped tubes under complex field backgrounds. Background Art

[0002] The crimping tube of the transmission line is an important hardware to achieve a stable connection between the conductor and the insulator string. Its crimping quality is directly related to the safe and stable operation of the line. Due to long-term service in complex outdoor environments, its electrical and mechanical properties are prone to failure. Even during the operation of the power grid, the rust of the crimping tube may cause line breakage and tower collapse accidents, which seriously threatens the safe and stable operation of the transmission line. Testing the crimping tube can help to reduce the failure of the crimping tube.

[0003] At present, the quality inspection of crimping tubes mainly relies on manual labor, but due to the shortcomings of manual inspection methods such as high labor intensity, low work efficiency, scattered inspection quality, and single means, it is very necessary to develop drone inspection technology. The relevant research on visual inspection of power transmission line hardware at home and abroad is mainly divided into two categories: machine learning and deep learning. The detection algorithm based on deep learning is very dependent on computer resources. The drone-mounted equipment is difficult to meet the current algorithm hardware resource requirements. Therefore, a method that can meet the detection accuracy requirements and reduce the demand for equipment hardware resources is needed. This invention proposes a lightweight YOLO-T algorithm to realize an accurate and fast crimping tube detection model. Summary of the invention

[0004] This detection model is based on the concept of the YOLO algorithm, but adopts a brand-new backbone network, feature fusion network and wide-area detection head. In view of the problem of large size of traditional network models, the present invention introduces the MCBSC-C module and the GBSC module to construct a lightweight backbone network that can quickly extract effective features. The feature fusion network performs multi-scale feature fusion through the GBSCSPPF module, and focuses on important feature maps through the CSA module. In addition, the lightweight wide-area detection head achieves effective detection of multi-scale targets while reducing the amount of calculation and the amount of parameters. By performing target detection on only two feature maps, the present invention greatly reduces the amount of calculation and the amount of parameters of the network, while maintaining high accuracy and high speed of the detection effect. The present invention can realize the recognition of crimping pipes in complex environments, has the characteristics of high recognition accuracy and rapid response, and is suitable for maintenance and safety inspection of power systems.

[0005] In order to achieve the above-mentioned purpose of the present invention, the present invention adopts the following technical solutions:

[0006] A method for detecting crimped tubes of transmission lines in a field with complex backgrounds is composed of a visible light image acquisition device and a visible light image target detection device; the visible light image acquisition device acquires visible light images of crimped tubes in a complex environment background and inputs them into the crimped tube image target detection device; the crimped tube visible light image target detection device contains a crimped tube image recognition model of YOLO-T for crimped tube detection; the method for detecting crimped tubes of transmission lines in a field with complex backgrounds includes a backbone network, a feature fusion network, and a wide-area detection head; the backbone network is composed of a GBSC1 module, an MCBSC-C4× module, an MCBSC-C2× module, a GBSC2 module, and a GBSC3 module in sequence, and the output feature maps of the output of the MCBSC-C4× module and the output of the GBSC3 are respectively selected and input into the feature fusion network. The feature fusion network is composed of a GBSCSPPF module, an Upsample module, and a Concat module in sequence, and the output feature maps of the Concat module and GBSC2 are respectively input into the wide-area detection head. The outputs of the MCBSC-C4× module and GBSC3 module of the backbone network are respectively passed to the GBSCSPPF module and the Upsample module to obtain the output feature map, which is fused through the Concat module to obtain the final output feature map of the feature fusion network.

[0007] The wide-area detection head is composed of GBSC4 module, Conv1 module, GBSC5 module and Conv2 module in sequence. The output feature maps of Conv1 module and Conv2 module are selected as the output of YOLO-T model respectively. The wide-area detection head takes the output feature maps of GBSC2 module of backbone network and Concat module of feature fusion network as input.

[0008] MCBSC-C4× module: The MCBSC-C4× module is composed of four MCBSC-C modules connected together. The MCBSC-C module has two branches. In the first branch, InputMaps is processed by the Maxpool module, then by the Avgpool module, and then the output feature map of the Avgpool module and the output feature map of the Maxpool module are processed by the multiplication module. In the second branch, the output feature map of Maxpool is processed by the Conv module, then by the BN module, and finally by the Swish. The output feature maps of the two branches are fused, and the output feature map of the multiplication module and the output feature map of Swish are processed by the addition module, and then processed by the CSA module to obtain the output feature map OutputMaps of the MCBSC-C module.

[0009] MCBSC-C2× module: The MCBSC-C2× module is composed of two MCBSC-C modules connected together. The MCBSC-C module has two branches. In the first branch, InputMaps is processed by the Maxpool module, then by the Avgpool module, and then the output feature map of the Avgpool module and the output feature map of the Maxpool module are processed by the multiplication module. In the second branch, the output feature map of Maxpool is processed by the Conv module, then by the BN module, and finally by the Swish. The output feature maps of the two branches are fused, and the output feature map of the multiplication module and the output feature map of Swish are processed by the addition module, and then processed by the CSA module to obtain the output feature map OutputMaps of the MCBSC-C module.

[0010] CSA module: The CSA module has two branches. The first branch: InputMaps is processed by the Conv module, then by the BN module, and then by the Weights module to perform channel weight analysis to obtain weight information. The second branch: InputMaps is processed by the Maxpool module, and then by the Avgpool module to obtain the output feature map. The two branch feature maps are fused, and the outputs of the BN module, the Weights module, and the Avgpool module are processed by the multiplication module to obtain the final output feature map of the CSA module.

[0011] GBSCSPPF module: The GBSCSPPF module has two branches. The first branch: InputMaps is processed by the GroupConv module, then comes out of the BN module, then processed by the Swish function, and finally processed by the CSA module. The second branch: The output feature map of the Group Conv module is input to the Maxpool1 module for processing, then processed by the Maxpool2 module, and finally processed by the Maxpool3 module. The output feature maps of the two branches are stacked, and the output feature maps of the CSA module, Maxpool1 module, Maxpool2 module, and Maxpool3 module are stacked to obtain the output feature map of the GBSCSPPF module.

[0012] GBSC module: consists of Group Conv module (group convolution), BN module (batch normalization), Swish activation function, and CSA module.

[0013] The wide-area detection head has two detection heads. The first detection head is composed of GBSC4 module and Conv1 module. The second detection head is composed of GBSC5 module and Conv2 module. This detection head directly obtains the detection frame and can achieve good detection results for targets of any scale. At the same time, since there are only two detection heads to directly obtain the detection frame, the calculation amount and parameter amount are lower.

[0014] The detection head implements target classification and regression tasks, removes redundant and overlapping bounding boxes through DIoU_NMS operations, retains the bounding boxes with the highest confidence, and outputs the processed bounding boxes and confidence scores.

[0015] An electronic device includes a processor, a memory and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a method for detecting transmission line pressure pipes in a complex background in the field as described above.

[0016] A non-volatile computer storage medium stores computer executable instructions, characterized in that the computer can execute the above-mentioned method for detecting transmission line crimped pipes in complex field backgrounds.

[0017] The present invention also provides a method for detecting a pressure pipe of a transmission line in a complex background in the field. The pressure pipe visible light image is collected, and the features of the pipe visible light image are extracted through the backbone network, and the pipe visible light image is converted into a multi-layer feature map. Two feature maps of different levels are selected and input into a feature fusion network, and feature fusion is performed through the feature fusion network. The fused feature map is input into a wide-area detection head; the wide-area detection head predicts the target category and position, directly generates a bounding box without an anchor box, calculates the confidence, and outputs the predicted result.

[0018] The beneficial effects of the present invention are as follows:

[0019] 1. A method for detecting pressure pipes for transmission lines in a complex background in the field is designed according to the main concept of the YOLO algorithm, and a new feature extraction network is used as the backbone network. The backbone network is constructed by an MCBSC-C module and a GBSC module. The backbone network of the present invention has the following advantages: more efficient parameter utilization, the GBSC module can obtain more feature maps through a simpler convolution operation, reduce the amount of calculation and speed up the reasoning speed while ensuring the accuracy of target detection; higher accuracy, the MCBSC-C module can improve the expression ability of feature map features, better reveal the information in the image features, and has a higher accuracy; high computational efficiency, by removing redundant parameters and layers, reducing complexity and computational space, and improving the computational efficiency of the model; easy to deploy, the lightness and high efficiency of the backbone network composed of the MCBSC-C module and the GBSC module make it easier to deploy in various restricted environments.

[0020] 2. The present invention uses the GBSCSPPF module optimized by the SPPF module in the feature fusion module, which has the following advantages: extracting rich feature information through three serial pooling operations to improve the recognition accuracy of the model, extracting features from different scales, finding detailed information of object targets at different scales, and utilizing the semantic information of rich features such as maximum pooling and feature fusion to enhance the model's perception of objects. The GBSC module is used to extract original scale features from the input feature map without affecting the computational complexity and parameter quantity of the network model, thereby ensuring that the detection effect of the network model is at least not lower than that when the GBSCSPPF module is not used in any scenario.

[0021] 3. The present invention adopts CSA module in the backbone network, feature fusion network and wide-area detection head. Advantages of this module: rich feature information, CSA module adopts the idea of ​​gradient diversion, and can obtain richer gradient information; strong feature fusion ability, good target detection performance, and fusion of low-level features lacking context information and semantic information with high-level information lacking details, which can use detail information and semantic information at different angles, thereby improving the accuracy, robustness and efficiency of target detection; excellent performance CSA module realizes channel attention mechanism through weights, reduces computational cost and model storage, and improves the operation efficiency of the model. CSA is a few-parameter attention mechanism. By calculating the attention weight, the model's attention is focused on the most relevant features, improving the accuracy and generalization ability of the model. At the same time, in each iteration, CSA will calculate the attention weight, and then apply it to the feature vector, and then update the parameters of the model. Since the attention weight is dynamic, the model can adjust itself to adapt to different tasks and data sets. CSA does not adopt a parameter-free method, and achieves high-speed reasoning of CSA by balancing the amount of calculation, the amount of parameters and the way of data reading. The pure non-parameter method will reduce the amount of calculation and the number of parameters, which will lead to repeated reading of data, resulting in the data reading speed becoming the bottleneck of the module reasoning speed. The traditional method has a large demand for the amount of calculation, resulting in the computing power of the device becoming the bottleneck of the module reasoning speed. The CSA module achieves fast module reasoning by balancing the amount of calculation and data reading.

[0022] 4. The present invention replaces the non-maximum suppression value NMS with DIoU_NMS in the head network. Compared with NMS, which is slightly rough and directly deletes the boxes with IoU greater than the threshold, DIoU_NMS not only considers the overlapping area, but also the distance between the center points on the basis of NMS. In this way, when two objects are too close, DIoU_NMS can also detect them. In complex environmental conditions, the use of DIoU_NMS can better exert excellent performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The figure is a schematic diagram of a transmission line crimping pipe detection system in a complex field background.

[0024] Figure 2 The schematic diagram of the structure of a model of a method for detecting crimped pipes of transmission lines in complex field backgrounds.

[0025] Figure 3 Schematic diagram of the MCBSC-C structure.

[0026] Figure 4 This is a schematic diagram of the GBSCSPPF module structure.

[0027] Figure 5 Schematic diagram of the CSA module structure.

[0028] Figure 6 Schematic diagram of the GBSC module structure. DETAILED DESCRIPTION

[0029] In order to have a deeper understanding of the technical solution, purpose of the invention and advantages of the invention, the present invention is described in detail below in conjunction with the accompanying drawings and preferred embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0030] Reference Figure 1 to Figure 6 A transmission line crimped tube detection system in a field with complex backgrounds, the system comprising a visible light image acquisition device and a visible light image target detection device; the visible light image acquisition device acquires visible light images of crimped tubes in complex environment backgrounds and inputs them into the crimped tube image target detection device; the crimped tube visible light image target detection device contains a crimped tube image recognition model of YOLO-T for crimped tube detection;

[0031] The image object detection device comprises:

[0032] A backbone network is used to extract features from the visible light image and obtain the first, second and third feature maps;

[0033] The feature fusion network fuses the first and second features obtained by the backbone network to obtain a fused feature map;

[0034] The wide-area detection head predicts the third feature map obtained by the backbone network and the fused feature map obtained by the feature fusion network to obtain the target category and position.

[0035] Further, the backbone network is composed of a GBSC1 module, a MCBSC-C4× module, a MCBSC-C2× module, a GBSC2 module, and a GBSC3 module in sequence, the MCBSC-C4× module extracts the first feature map, the GBSC3 module obtains the second feature map, and the GBSC2 module obtains the third feature map;

[0036] The MCBSC-C4× module is composed of four MCBSC-C modules connected;

[0037] The MCBSC-C2× module is composed of two MCBSC-C modules connected;

[0038] The feature fusion network includes a GBSCSPPF module, an Upsample module, and a Concat module. The GBSCSPPF module is connected to the MCBSC-C4× module, and the Upsample module is connected to the GBSC3 module. The GBSCSPPF module is parallel to the Upsample module and is connected to the Concat module respectively. The Concat module performs feature fusion on the first and second features obtained by the backbone network to obtain a fused feature map.

[0039] The wide-area detection head includes a first branch consisting of a GBSC4 module and a Conv1 module and a second branch consisting of a GBSC5 module and a Conv2 module, the GBSC4 module of the first branch is connected to the GBSC2 module, the GBSC5 module of the second branch is connected to the Concat module, the Conv1 module of the first branch is used to output the target category, and the Conv2 module of the second branch is used to output the corresponding position;

[0040] Furthermore,

[0041] The MCBSC-C module includes two parallel branches and a processing branch;

[0042] The first branch: includes the InputMaps module, Maxpool module, Avgpool module, multiplication module and addition module connected in series;

[0043] The Maxpool module in the first branch is connected to the multiplication module;

[0044] The second branch: Maxpool module, Conv module processing, BN module and Swish module;

[0045] The output feature maps of the two branches are fused, and the output feature map of the multiplication module is added to the output feature map of Swish;

[0046] Processing branch: CSA module and OutputMaps module connected in series.

[0047] Furthermore,

[0048] The CSA module includes a main branch, a first branch, and a second branch;

[0049] Main branch: InputMaps module, Conv module, BN module, Multiplication module and Outmaps module;

[0050] First branch: a Weights module is also provided between the BN module and the multiplication module, and the Weights module performs channel weight analysis processing to obtain weight information;

[0051] Second branch: a Maxpool module and an Avgpool module are also provided between the InputMaps module and the multiplication module;

[0052] Among them, the multiplication module fuses the two branch feature maps, and the outputs of the BN module, the Weights module and the Avgpool module are multiplied to obtain the final output feature map of the CSA module.

[0053] Furthermore,

[0054] The GBSCSPPF module includes a main branch and three sub-branches in parallel;

[0055] Main branch: InputMaps module, Group Conv module processing, BN module, Swish function module processing, CSA module and Concat module;

[0056] Parallel sub-branch: the Group Conv module and the Concat module are sequentially connected in parallel to the Maxpool1 module, the Maxpool2 module and the Maxpool3 module;

[0057] The output feature maps of the two branches are stacked, and the output feature maps of the CSA module, Maxpool1 module, Maxpool2 module, and Maxpool3 module are stacked to obtain the output feature map of the GBSCSPPF module.

[0058] Furthermore, the GBSC module consists of a Group Conv module, a BN module, a Swish activation function, and a CSA module.

[0059] Further, the following steps are included:

[0060] Step S1: The backbone network is used to extract features from the visible light image to obtain the first, second and third feature maps;

[0061] Step S2: feature fusion network, which fuses the first and second features obtained by the backbone network to obtain a fused feature map;

[0062] Step S3: The wide area detection head predicts the third feature map obtained by the backbone network and the fused feature map obtained by the feature fusion network to obtain the target category and position.

[0063] like Figure 1As shown, a system for detecting crimped tubes of power lines in a field with complex backgrounds is composed of a visible light image acquisition device and a crimped tube visible light image target detection device; the visible light image acquisition device acquires the crimped tube visible light image and inputs it into the crimped tube visible light image target detection device; the crimped tube visible light image target detection device contains a visible light image recognition model for crimped tubes of power lines in a field with complex backgrounds, which is used for crimped tube detection.

[0064] The architecture diagram of a method for detecting transmission line crimping pipes in complex field backgrounds is as follows: Figure 2 As shown in the figure, it includes a backbone network, a feature fusion network, and a wide-area detection head. Taking the compressed tube visible light image with an input size of 640×640×3 as an example, the backbone network is used to extract the features of the compressed tube visible light image, convert the compressed tube visible light image into a multi-layer feature map, select two feature maps of different levels and input them into the feature fusion network, the feature fusion network performs feature fusion, and the fused feature map is input into the wide-area detection head, which predicts the target category and position, generates a bounding box, calculates the confidence, and outputs the predicted result.

[0065] The backbone network is composed of GBSC1 module, MCBSC-C4× module, MCBSC-C2× module, GBSC2 module and GBSC3 module in sequence. The output feature maps of the MCBSC-C4× module and the GBSC3 output are respectively selected and input into the feature fusion network.

[0066] The feature fusion network is composed of GBSCSPPF module, Upsample module and Concat module in sequence. The output feature maps of Concat module and GBSC2 are input to the wide area detection head respectively. The outputs of MCBSC-C4× module and GBSC3 module of the backbone network are respectively input to GBSCSPPF module and Upsample module to obtain the output feature maps, which are fused through Concat module to obtain the final output feature map of the feature fusion network. The output feature map is used as part of the input of the wide area detection head.

[0067] The wide-area detection head is composed of GBSC4 module, Conv1 module, GBSC5 module and Conv2 module in sequence. The output feature maps of Conv1 module and Conv2 module are selected as the output of YOLO-T model respectively. The wide-area detection head takes the output feature maps of GBSC2 module of backbone network and Concat module of feature fusion network as input.

[0068] The backbone network samples the MCBSC-C module and GBSC module as the basic modules for network feature extraction. The GBSC module uses simple linear operations to replace ordinary convolution to generate more similar feature maps, so that more feature maps can be generated with fewer parameters. The GBSC module uses the Group Conv module (grouped convolution) and the CSA module. The Group Conv module effectively reduces the number of parameters and the amount of calculation of the network model. The CSA module is a low-parameter attention mechanism module, which enables the model to focus on important spatial and channel features, and improves the model detection effect to a limited extent. The reasonable use of the GBSC module in the backbone network improves the utilization of parameters while also improving the detection effect of the network model.

[0069] MCBSC-C4× module: The MCBSC-C4× module is composed of four MCBSC-C modules connected together. The MCBSC-C module has two branches. In the first branch, InputMaps is processed by the Maxpool module, then by the Avgpool module, and then the output feature map of the Avgpool module and the output feature map of the Maxpool module are processed by the multiplication module. In the second branch, the output feature map of Maxpool is processed by the Conv module, then by the BN module, and finally by the Swish. The output feature maps of the two branches are fused, and the output feature map of the multiplication module and the output feature map of Swish are processed by the addition module, and then processed by the CSA module to obtain the output feature map OutputMaps of the MCBSC-C module.

[0070] MCBSC-C2× module: The MCBSC-C2× module is composed of two MCBSC-C modules connected together. The MCBSC-C module has two branches. In the first branch, InputMaps is processed by the Maxpool module, then by the Avgpool module, and then the output feature map of the Avgpool module and the output feature map of the Maxpool module are processed by the multiplication module. In the second branch, the output feature map of Maxpool is processed by the Conv module, then by the BN module, and finally by the Swish. The output feature maps of the two branches are fused, and the output feature map of the multiplication module and the output feature map of Swish are processed by the addition module, and then processed by the CSA module to obtain the output feature map OutputMaps of the MCBSC-C module.

[0071] CSA module: The CSA module has two branches. The first branch: InputMaps is processed by the Conv module, then by the BN module, and then by the Weights module to perform channel weight analysis to obtain weight information. The second branch: InputMaps is processed by the Maxpool module, and then by the Avgpool module to obtain the output feature map. The two branch feature maps are fused, and the outputs of the BN module, the Weights module, and the Avgpool module are processed by the multiplication module to obtain the final output feature map of the CSA module. The CSA low-parameter attention mechanism reduces the complexity of the model by reducing the number of parameters, making the model easier to train and reducing the risk of overfitting. Fewer parameters means that the model occupies less storage space, which is particularly important for models deployed on embedded devices or mobile devices. Due to the small number of parameters, the amount of calculation is also reduced accordingly, so the model has a faster inference speed and is suitable for real-time applications. The reduction in the amount of calculation means lower energy consumption, which is particularly important for battery-powered devices. The low-parameter attention mechanism can be flexibly applied to different network architectures and tasks, while effectively balancing the balance between computing power and data.

[0072] GBSC module: It consists of Group Conv module (group convolution), BN module (batch normalization), Swish activation function, and CSA module. While maintaining feature extraction, the GBSC module can reduce the spatial and channel redundancy between features in the convolutional neural network, thereby compressing the model and improving its performance, and enhancing the input feature expression capability.

[0073] A model of a method for detecting a pressure pipe of a transmission line in a complex background in the field is a model that has been trained by a training set, verified by a validation set, and tested by a test machine. Different GIS infrared images are collected by an infrared image acquisition device, and the data volume is expanded through data enhancement operations such as rotation, cropping, scaling, and shifting to make a data set, which is divided into a training set, a validation set, and a test set, with a ratio of about 8:1:1. Labelimg software is used to annotate the data sets respectively, and the annotated training set is used to start training the GIS infrared image recognition model based on the improved YOLOv5, and the trained model is verified and tested with the validation set and the test set.

[0074] The wide-area detection head consists of two detection heads. The wide-area detection head inputs the output feature map of GBSC2 of the backbone network and the output feature map of Concat of the feature fusion network. The GBSC4 module and Conv1 adapt to different detection tasks by adjusting the number of channels and kernel size. The detection head implements target classification and regression tasks, and removes redundant and overlapping bounding boxes through DIoU_NMS operation, retains the bounding box with the highest confidence, and outputs the processed bounding box and confidence score. Compared with NMS, DIoU_NMS is slightly rough in directly deleting bounding boxes with IoU greater than the threshold. On the basis of NMS, DIoU_NMS not only considers the overlapping area, but also the distance of the center point position, and uses DIoU instead of IoU as the judgment criterion of NMS.

[0075] A model for detecting a pressure-connected pipe of a transmission line in a complex background in the wild is proposed. The hyperparameters (training steps, learning rate, optimal network structure, etc.) of the model are adjusted through a validation set to obtain the best performance. The test set is used to verify whether the model for detecting a pressure-connected pipe of a transmission line in a complex background in the wild has generalization performance after the hyperparameter training of the model on the training set and the validation set. When using the test set to evaluate the model for detecting a pressure-connected pipe of a transmission line in a complex background in the wild, the evaluation references include PR curve, mean average precision (mAP) of all classes, and frames per second (FPS). Finally, the GIS infrared image recognition model based on the improved YOLOv5 that can meet the evaluation requirements is placed in the GIS infrared image target detection device to identify the substation GIS.

[0076] The present invention also provides a method for detecting crimped tubes of power transmission lines in complex backgrounds in the wild, which collects visible light images of the crimped tubes, uses a backbone network to extract features of the visible light images of the crimped tubes, converts the visible light images of the crimped tubes into multi-layer feature maps, selects two feature maps of different levels and inputs them into a feature fusion network, performs feature fusion through the feature fusion network, and inputs the fused feature maps and the output feature maps of the backbone network GBSC2 into a wide-area detection head; the wide-area detection head predicts target categories and positions, generates bounding boxes, calculates confidence levels, and outputs prediction results.

[0077] An electronic device includes a processor, a memory and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a method for detecting transmission line pressure pipes in a complex background in the field as described above.

[0078] A non-volatile computer storage medium stores computer executable instructions, characterized in that the computer can execute the above-mentioned method for detecting transmission line crimped pipes in complex field backgrounds.

Claims

1. A field complex background transmission line crimping pipe detection system, characterized in that: The system is composed of a visible light image acquisition device and a visible light image target detection device; the visible light image acquisition device acquires a visible light image of a compressed tube under a complex environment background and inputs it into the compressed tube image target detection device; The crimped tube visible light image target detection device contains the crimped tube image recognition model of YOLO-T, which is used for crimped tube detection; The image object detection device comprises: A backbone network is used to extract features from the visible light image and obtain the first, second and third feature maps; A feature fusion network is used to fuse the first and second feature maps obtained by the backbone network to obtain a fused feature map. The wide-area detection head predicts the third feature map obtained by the backbone network and the fused feature map obtained by the feature fusion network to obtain the target category and position.

2. A field complex background transmission line crimping pipe detection system as claimed in claim 1, characterized in that: The backbone network is composed of a GBSC1 module, a MCBSC-C4× module, a MCBSC-C2× module, a GBSC2 module, and a GBSC3 module in sequence, wherein the MCBSC-C4× module extracts the first feature map, the GBSC3 module obtains the second feature map, and the GBSC2 module obtains the third feature map; The MCBSC-C4× module is composed of four MCBSC-C modules connected; The MCBSC-C2× module is composed of two MCBSC-C modules connected; The feature fusion network includes a GBSCSPPF module, an Upsample module, and a Concat module. The GBSCSPPF module is connected to the MCBSC-C4× module, and the Upsample module is connected to the GBSC3 module. The GBSCSPPF module is parallel to the Upsample module and is connected to the Concat module respectively. The Concat module performs feature fusion on the first and second features obtained by the backbone network to obtain a fused feature map. The wide-area detection head includes a first branch consisting of a GBSC4 module and a Conv1 module and a second branch consisting of a GBSC5 module and a Conv2 module. The GBSC4 module of the first branch is connected to the GBSC2 module, and the GBSC5 module of the second branch is connected to the Concat module. The Conv1 module of the first branch is used to output the target category, and the Conv2 module of the second branch is used to output the corresponding position.

3. A field complex background transmission line crimping pipe detection system according to claim 1, characterized in that: The MCBSC-C module includes two parallel branches and a processing branch; The first branch: includes the InputMaps module, Maxpool module, Avgpool module, multiplication module and addition module connected in series; The Maxpool module in the first branch is connected to the multiplication module; The second branch: Maxpool module, Conv module processing, BN module and Swish module; The output feature maps of the two branches are fused, and the output feature map of the multiplication module is added to the output feature map of Swish; Processing branch: CSA module and OutputMaps module connected in series.

4. A field complex background transmission line crimping pipe detection system according to claim 3, characterized in that: The CSA module includes a main branch, a first branch, and a second branch; Main branch: InputMaps module, Conv module, BN module, Multiplication module and Outmaps module; First branch: a Weights module is also provided between the BN module and the multiplication module, and the Weights module performs channel weight analysis processing to obtain weight information; Second branch: a Maxpool module and an Avgpool module are also provided between the InputMaps module and the multiplication module; Among them, the multiplication module fuses the two branch feature maps, and the outputs of the BN module, the Weights module and the Avgpool module are multiplied to obtain the final output feature map of the CSA module.

5. The field complex background transmission line crimping pipe detection system according to claim 2 is characterized in that: The GBSCSPPF module includes a main branch and three sub-branches in parallel; Main branch: InputMaps module, Group Conv module processing, BN module, Swish function module processing, CSA module and Concat module; Parallel sub-branch: the Group Conv module and the Concat module are sequentially connected in parallel to the Maxpool1 module, the Maxpool2 module and the Maxpool3 module; The output feature maps of the two branches are stacked, and the output feature maps of the CSA module, Maxpool1 module, Maxpool2 module, and Maxpool3 module are stacked to obtain the output feature map of the GBSCSPPF module.

6. The field complex background transmission line crimping pipe detection system according to claim 1 is characterized in that: The GBSC module consists of a Group Conv module, a BN module, a Swish activation function, and a CSA module.

7. A method for detecting crimped pipes of power transmission lines in a field with complex backgrounds, used in a system for detecting crimped pipes of power transmission lines in a field with complex backgrounds as claimed in any one of claims 1 to 6, characterized in that: The following steps are involved: Step S1: The backbone network is used to extract features from the visible light image to obtain the first, second and third feature maps; Step S2 : Feature fusion network, which fuses the first and second features obtained by the backbone network to obtain a fused feature map; Step S3: The wide area detection head predicts the third feature map obtained by the backbone network and the fused feature map obtained by the feature fusion network to obtain the target category and position.

8. An electronic device, characterized in that: It includes a processor, a memory and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a method for detecting transmission line pressure pipes in a complex outdoor background as described in claim 7.

9. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions execute the method for detecting crimped pipes of transmission lines in complex field backgrounds as described in claim 7.

Images