Overhead line insulator based on target detection algorithm and defect detection method thereof

Abstract

The invention provides an overhead line insulator based on a target detection algorithm and a defect detection method thereof, and is suitable for the field of intelligent inspection of power transmission lines and insulator image recognition in a power system. The method comprises the steps that firstly, insulator aerial images collected in the inspection process are classified and preprocessed, an insulator data set and an insulator defect data set are constructed, the two data sets are used for training a detection model to obtain an insulator detection model and an insulator defect detection model, the two detection models are cascaded, and the insulator detection model and the insulator defect detection model are obtained; and finally, inputting to-be-detected image data into the cascade detection model, outputting an insulator and a defect detection result thereof, positioning an insulator string and evaluating the working state of the insulator. According to the method, the insulator and the defect detection task of the insulator are efficiently and rapidly completed in a complex background, meanwhile, a lighter model can be deployed in various detection devices, and the flexibility of practical application of the model is improved.

Description

Technical Field

[0001] The present invention relates to the fields of power inspection and image recognition, and particularly to an overhead line insulator based on a target detection algorithm and a method for detecting its defects. Background Technique

[0002] As a key component in overhead transmission lines, insulators play a role in supporting conductors and preventing electric leakage, ensuring continuous power transmission and maintaining the safe operation of the power system. Insulators on overhead transmission lines are exposed to the external environment due to their special functions. After long-term use, insulators are easily affected by adverse factors such as weather, pollution, and aging, and defects such as breakage, fouling, flashover, and self-explosion may occur, which may cause serious consequences such as line failures and large-scale power outages, threatening the reliability and safety of the power system. Therefore, it is necessary to regularly inspect and maintain overhead transmission lines. Traditional power inspections mainly rely on staff to check one by one on-site, with a huge workload, and are easily affected by factors such as complex terrain and visual fatigue, posing risks of missed inspections and misjudgments. In order to reduce errors and achieve all-weather and uninterrupted inspections, using advanced devices such as drones and robots in cooperation with an automated inspection system for inspections has become the current mainstream development trend.

[0003] Using drones and robots for inspections and combining image processing, deep learning, and computer vision technologies can improve the efficiency and accuracy of automated inspections. There are mainly two existing target detection methods: one is the single-stage detection algorithm represented by the YOLO and RT-DETR series, and the other is the two-stage detection algorithm represented by the R-CNN and Faster R-CNN series. The core idea of the two-stage detection algorithm is to divide target detection into two stages, candidate region generation and classification and boundary regression, and it is often applied to detection scenarios with higher accuracy requirements. Compared with the traditional two-stage detection algorithm, the single-stage detection algorithm only needs one forward propagation to complete the target detection task, with a simpler network and faster detection speed. Therefore, in order to meet the real-time and accuracy requirements of insulator and its defect detection, the single-stage detection algorithm is often selected as the basic algorithm for this research topic. Summary of the Invention

[0004] For the above research background of overhead line insulators and their defect detection, the present invention proposes an overhead line insulator based on a target detection algorithm and a method for detecting its defects. The process of using this method includes the following steps:

[0005] Step 1, collect aerial images of insulators taken by drones, divide the pictures into two categories: insulators and insulator defects according to whether there are defects on the insulator surface, and perform data augmentation and preprocessing on the original images to construct two datasets of insulators and insulator defects. The specific process is as follows:

[0006] Control the UAV to continuously capture the insulators of the overhead transmission line at different heights and angles, and select the photos with clear target images from them as the original images. Use the open-source data annotation software labelimg to annotate the detection targets in the original images, including the positions of the insulator strings and the categories of various defects on the insulator surfaces. The images and label files correspond one by one. Perform data augmentation on the original images, and the label files are also processed in the same way. The specific data augmentation methods include: rotation, horizontal flipping, vertical flipping, random occlusion, and Gaussian blur. Divide the processed insulator images with defects into a training set, a validation set, and a test set according to the ratio of 8:2:2 to complete the construction of the insulator defect data set. Similarly, divide the insulator images without defects into a training set, a validation set, and a test set according to the ratio of 8:2:2 to complete the construction of the insulator data set.

[0007] Step 2: Obtain an improved detection model. Use the two data sets constructed in the previous step to iteratively train the model to obtain the optimal detection models for insulators and insulator defects, and cascade the two network models to jointly complete the detection work. And the steps before step 2 also include: Based on the YOLOv11s algorithm, use a lightweight neural network architecture to improve the feature extraction part of the YOLOv11s network, and use an efficient convolutional module and a small target detection layer to improve the feature fusion part of the YOLOv11s network. The lightweight neural network architecture is the FasterNet network, and the efficient convolutional module is the C3K2-Para module. The specific process is as follows:

[0008] (1) Use the FasterNet network architecture to replace the feature extraction part of the original YOLOv11s network. The input image data will pass through the FasterNet backbone, and the extracted image feature information will be sent to the SPPF pyramid pooling layer. The FasterNet network architecture consists of four stages. The input of the first stage is The input of the second stage is The input of the third stage is The input of the fourth stage is Each stage contains a FasterNet block. Before the FasterNet block, there is an embedding layer (a 4×4 regular convolution) or a fusion layer (a 2×2 regular convolution) for downsampling or channel adjustment. Each FasterNet block contains a PConv layer, two PWConv layers (or 1×1 regular convolutions), and there are normalization layers and activation layers between every two PWConv layers. After each PConv, there are two PWConv layers forming a structure similar to a T-shaped convolution. At the end of the FasterNet network architecture, there are a pooling layer, a 1×1 convolution layer, and a fully connected layer, which are jointly used for feature classification and transformation. To meet the lightweight requirements of the model, the most appropriate network depth and width are selected in the FasterNet-T0 embedding network.

[0009] (2) To strengthen multi-scale feature fusion and improve the network's detection ability for small target features, a P2 small target detection layer is introduced in the feature fusion part. At the same time, the three detection heads of the original YOLOv11s are changed to four detection heads, and a P2 small target detection head is added.

[0010] After the upsampling operation on the last 80×80 feature layer in the original YOLOv11s feature fusion part, a new upsampling operation on a 160×160 feature layer is added, which can combine the P2 feature map in the Backbone feature extraction part, and then further extract and fuse features through the C3K2-Para module. At the same time, based on the three detection heads of the original YOLOv11s, a P2 small target detection head is added, and the P2 small target detection head is combined with the detection heads of other layers (P3, P4, P5) to fuse feature information of different scales. In addition, with the increase in the number of upsampling layers, an additional step of processing the P3 feature map is added. After all upsampling operations are completed, a convolution operation is performed on the feature map, and the feature map after the convolution operation is concatenated with the P3 feature map, and then passed through the C3K2-Para module, and the output is used as the detection head of the P3 layer.

[0011] (3) Reconstruct all C3K2 modules in the feature fusion part. First, improve the Ghost Module module, then design the C3K-Para module, and finally construct the C3K2-Para module based on the C3K-Para module.

[0012] The Ghost Module module uses ordinary convolution to complete the generation of feature maps for some channels and group convolution operations, while the improved Ghost Module module uses Dynamic Conv (dynamic convolution) instead of ordinary convolution to complete the above operations in the Ghost Module module. The basic definition of Dynamic Conv is described by the following formula: 0≤πk f(x) ≤ 1,

[0013] where x and y represent the input and output respectively, g is the activation function, π k is the attention weight, and W and b are the weight matrix and bias vector respectively.

[0014] The improved Ghost Module first uses Dynamic Conv to generate intrinsic features with a small number of channels, then performs a series of cheap convolutional linear transformations based on Dynamic Conv on these intrinsic features in each channel to generate Ghost features, and finally concatenates the feature maps obtained in the first two steps to generate more feature maps. After obtaining the improved Ghost Module, the improved Ghost Module is used to replace the Bottleneck in the C3K module to obtain the C3K-Para module. When the input enters the C3K-Para module, it will first be divided into two branches. One branch will pass through two consecutive improved Ghost Modules after passing through the CBS module, and the other branch will pass through the CBS module. The outputs of the two branches are concatenated, and after concatenation, they pass through the CBS module again, and the result is the output of the C3K-Para module. After obtaining the C3K-Para module, the improved Ghost Module and the C3K-Para module are integrated into the C3K2 module, replacing the C3K and Bottleneck in the original C3K2 module to construct the C3K2-Para module. The specific structure of the C3K2-Para module is determined by parameter settings. When C3K = True, after the input passes through the CBS module, it will be evenly divided into two parts in the first dimension. One part directly enters the connection layer, and the other part becomes two branches. One branch also enters the connection layer, and the other branch contains two consecutive C3K-Para modules. The outputs passing through only the first C3K-Para module and the outputs passing through two consecutive C3K-Para modules enter the connection layer respectively. The above four parts of the output are concatenated in the connection layer and then input into the CBS module, and the result is the output of the C3K2-Para module when C3K = True. When C3K = False, after the input passes through the CBS module, it will be evenly divided into two parts in the first dimension. One part directly enters the connection layer, and the other part becomes two branches. One branch also enters the connection layer, and the other branch contains two consecutive improved Ghost Modules. The outputs passing through only the first improved Ghost Module and the outputs passing through two consecutive improved Ghost Modules enter the connection layer respectively. The above four parts of the output are concatenated in the connection layer and then input into the CBS module, and the result is the output of the C3K2-Para module when C3K = False. The setting of the C3K parameter in the network refers to the configuration file.

[0015] After completing the above steps, an improved detection model can be obtained. Then, use the two datasets constructed in Step 1 to iteratively train the detection model respectively. The number of iterations and the setting of training hyperparameters should enable the training results to converge stably. After sufficient iterative training, the optimal detection models for insulators and insulator defects can be obtained. Finally, cascade the two network models to jointly complete the detection work.

[0016] Step 3: Input the insulator image data to be detected into the cascade detection model, locate the insulator string, and output the detection results of the insulator surface defects, thus completing the tasks of locating and identifying the insulators and their defects. The specific process is as follows:

[0017] After the insulator image to be detected is input into the cascade detection model, the position of the insulator string and the surface defects of the insulator will be identified in sequence. There are two possibilities for the insulator image to be detected. If there are no defects on the insulator surface, the insulator defect detection model will not have an output result, and its final output result only shows the position of the insulator string. If there are defects on the insulator surface, both detection models will have output results, and its final output result includes the position of the insulator string, the position of the insulator surface defects, and the types of defects. Brief Description of the Drawings

[0018] Figure 1 is the flowchart of a method for detecting overhead line insulators and their defects based on object detection algorithms according to the present invention;

[0019] Figure 2 is the schematic diagram of the network structure of the improved YOLOv11s algorithm according to the present invention;

[0020] Figure 3 is the structural diagram of the FasterNet architecture according to the present invention;

[0021] Figure 4 is the structural diagram of the FasterNet block according to the present invention;

[0022] Figure 5 is the structural diagram of the Dynamic Conv according to the present invention;

[0023] Figure 6 is the structural diagram of the improved Ghost Module module according to the present invention;

[0024] Figure 7 is the structural diagram of the C3K-Para module according to the present invention;

[0025] Figure 8 is the structural diagram of the C3K2-Para module according to the present invention. Detailed Embodiments

[0026] The present invention will be described in detail below with reference to the accompanying drawings:

[0027] Figure 1 This is a flowchart of an overhead line insulator and its defect detection method based on a target detection algorithm according to the present invention. As shown in the flowchart, the detection method includes the following steps:

[0028] Step 1: Collect aerial images of insulators taken by a drone. Classify the pictures into two categories: insulators and insulator defects according to whether there are defects on the insulator surface. Perform data augmentation and preprocessing on the original images to construct two datasets of insulators and insulator defects. The specific process is as follows:

[0029] Control the drone to continuously take pictures of the overhead transmission line insulators at different heights and angles, and select the photos with clear target images as the original images. Use the open-source data annotation software labelimg to annotate the detection targets in the original images, including the positions of the insulator strings and the categories of various defects on the insulator surface. The images and label files correspond one by one. Perform data augmentation on the original images, and the label files are also processed in the same way. The specific data augmentation methods include: rotation, horizontal flipping, vertical flipping, random occlusion, and Gaussian blur. Divide the processed insulator images with defects into a training set, a validation set, and a test set according to the ratio of 8:2:2 to complete the construction of the insulator defect dataset. Similarly, divide the insulator images without defects into a training set, a validation set, and a test set according to the ratio of 8:2:2 to complete the construction of the insulator dataset.

[0030] Step 2: Obtain an improved detection model. Use the two datasets constructed in the previous step to iteratively train the model respectively to obtain the optimal detection models for insulators and insulator defects, and cascade the two network models to jointly complete the detection work. And the steps before step 2 also include: Based on the YOLOv11s algorithm, use a lightweight neural network architecture to improve the feature extraction part of the YOLOv11s network, and use an efficient convolutional module and a small target detection layer to improve the feature fusion part of the YOLOv11s network. The lightweight neural network architecture is the FasterNet network, and the efficient convolutional module is the C3K2-Para module. The specific process is as follows:

[0031] As Figure 2 shown, this is the schematic diagram of the network structure of the improved YOLOv11s algorithm in the present invention, and its main improved structure includes the following three points.

[0032] (1) Use the FasterNet network architecture to replace the feature extraction part of the original YOLOv11s network. The input image data will pass through the FasterNet backbone, and the extracted image feature information will be sent to the SPPF pyramid pooling layer. The FasterNet network architecture consists of four stages. The input of the first stage is The input of the second stage is The input of the third stage is The input of the fourth stage is Each stage contains a FasterNet block. Before the FasterNet block, there is an embedding layer (a 4×4 regular convolution) or a fusion layer (a 2×2 regular convolution) for downsampling or channel adjustment. Each FasterNet block contains a PConv layer, two PWConv layers (or 1×1 regular convolutions), and there are normalization layers and activation layers between every two PWConv layers. After each PConv, there are two PWConv layers forming a structure similar to a T-shaped convolution. At the end of the FasterNet network architecture, there are a pooling layer, a 1×1 convolution layer, and a fully connected layer, which are jointly used for feature classification and transformation. To meet the lightweight requirements of the model, the most appropriate network depth and width are selected in the FasterNet-T0 embedding network.

[0033] Figure 3 Figure 4 That is the above-mentioned FasterNet architecture structure diagram and FasterNet block structure diagram.

[0034] (2) To strengthen multi-scale feature fusion and improve the network's detection ability for small target features, a P2 small target detection layer is introduced in the feature fusion part. At the same time, the three detection heads of the original YOLOv11s are changed to four detection heads, and a P2 small target detection head is added.

[0035] After the upsampling operation on the last 80×80 feature layer in the original YOLOv11s feature fusion part, a new upsampling operation on a 160×160 feature layer is added, which can combine the P2 feature map in the Backbone feature extraction part, and then further extract and fuse features through the C3K2-Para module. At the same time, based on the three detection heads of the original YOLOv11s, a P2 small target detection head is added, and the P2 small target detection head is combined with the detection heads of other layers (P3, P4, P5) to fuse feature information of different scales. In addition, as the number of upsampling layers increases, an additional step of processing the P3 feature map is added. After all upsampling operations are completed, a convolution operation is performed on the feature map, and the feature map after the convolution operation is concatenated with the P3 feature map, and then passed through the C3K2-Para module, and the output is used as the detection head of the P3 layer.

[0036] (3) Reconstruct all C3K2 modules in the feature fusion part. First, improve the Ghost Module module, then design the C3K-Para module, and finally construct the C3K2-Para module based on the C3K-Para module.

[0037] The Ghost Module uses ordinary convolution to complete the generation of feature maps for some channels and group convolution operations, while the improved Ghost Module uses Dynamic Conv (dynamic convolution) instead of ordinary convolution to complete the above operations in the Ghost Module. The basic definition of Dynamic Conv is described by the following formula: 0 ≤ π k (x) ≤ 1,

[0038] where x and y represent the input and output respectively, g is the activation function, and π k is the attention weight, and W and b are the weight matrix and bias vector respectively.

[0039] The improved Ghost Module first uses Dynamic Conv to generate inherent features with a small number of channels, then performs a series of cheap convolutional linear transformations based on Dynamic Conv on these inherent features in each channel to generate Ghost features, and finally concatenates the feature maps obtained in the previous two steps to generate more feature maps. After obtaining the improved Ghost Module, the improved Ghost Module is used to replace the Bottleneck in the C3K module to obtain the C3K-Para module. When the input enters the C3K-Para module, it will first be divided into two branches. One branch will pass through two consecutive improved Ghost Modules after passing through the CBS module, and the other branch will pass through the CBS module. The outputs of the two branches are concatenated, and after concatenation, they pass through the CBS module again, and the result is the output of the C3K-Para module. After obtaining the C3K-Para module, the improved Ghost Module and the C3K-Para module are integrated into the C3K2 module, replacing the C3K and Bottleneck in the original C3K2 module to construct the C3K2-Para module. The specific structure of the C3K2-Para module is determined by the parameter settings. When C3K = True, after the input passes through the CBS module, it will be evenly divided into two parts in the first dimension. One part directly enters the connection layer, and the other part becomes two branches. One branch also enters the connection layer, and the other branch contains two consecutive C3K-Para modules. The output of only the first C3K-Para module and the output of passing through two consecutive C3K-Para modules enter the connection layer respectively. The above four parts of the output are concatenated in the connection layer and then input into the CBS module, and the result is the output of the C3K2-Para module when C3K = True. When C3K = False, after the input passes through the CBS module, it will be evenly divided into two parts in the first dimension. One part directly enters the connection layer, and the other part becomes two branches. One branch also enters the connection layer, and the other branch contains two consecutive improved Ghost Modules. The output of only the first improved Ghost Module and the output of passing through two consecutive improved Ghost Modules enter the connection layer respectively. The above four parts of the output are concatenated in the connection layer and then input into the CBS module, and the result is the output of the C3K2-Para module when C3K = False. The setting of the C3K parameter in the network refers to the configuration file.

[0040] Figure 6 Namely, it is the structure diagram of the above-mentioned improved Ghost Module. Figure 7 Namely, it is the structure diagram of the above-mentioned C3K-Para module. Figure 8 Namely, it is the structure diagram of the above-mentioned C3K2-Para module.

[0041] After completing the above steps, an improved detection model can be obtained. Then, use the two datasets constructed in Step 1 to iteratively train the detection model respectively. The number of iterations and the settings of training hyperparameters should enable the training results to converge stably. After sufficient iterative training, the optimal detection models for insulators and insulator defects can be obtained. Finally, cascade the two network models to jointly complete the detection work.

[0042] Step 3: Input the insulator image data to be detected into the cascade detection model, locate the insulator string, and output the detection results of the insulator surface defects, thus completing the tasks of locating and identifying the insulators and their defects. The specific process is as follows:

[0043] After the insulator image to be detected is input into the cascade detection model, the position of the insulator string and the surface defects of the insulator will be identified in sequence. There are two possibilities for the insulator image to be detected. If there are no defects on the insulator surface, the insulator defect detection model will not have an output result, and its final output result only shows the position of the insulator string. If there are defects on the insulator surface, both detection models will have output results, and its final output result includes the position of the insulator string, the position of the insulator surface defects, and the defect types.

Claims

1. An overhead line insulator based on a target detection algorithm and its defect detection method, characterized in that The method includes the following steps: Step 1, collect the aerial images of insulators taken by drones, classify the pictures into two categories: insulators and insulator defects according to whether there are defects on the insulator surface, and perform data augmentation and preprocessing on the original images to construct two datasets of insulators and insulator defects. Step 2, obtain an improved detection model, use the two datasets constructed in the previous step to iteratively train the model respectively to obtain the optimal detection models for insulators and insulator defects, and cascade the two network models to jointly complete the detection work. Step 3, input the image data of the insulator to be detected into the cascaded detection model, locate the insulator string, output the detection result of the insulator surface defect, and complete the tasks of locating and identifying the insulator and its defects.

2. The overhead line insulator and its defect detection method based on the target detection algorithm according to claim 1, characterized in that Collect the aerial images of insulators taken by drones, classify the pictures into two categories: insulators and insulator defects according to whether there are defects on the insulator surface, and perform data augmentation and preprocessing on the original images to construct two datasets of insulators and insulator defects. Specifically: Control the drone to continuously take pictures of the overhead transmission line insulators at different heights and angles, and select the photos with clear target images as the original images. Use the open-source data annotation software labelimg to annotate the detection targets in the original images, including the positions of the insulator strings and the categories of each defect on the insulator surface, and the images and label files correspond one by one. Perform data augmentation on the original images, and the label files are also processed in the same way. The specific data augmentation methods include: rotation, horizontal flipping, vertical flipping, random occlusion, and Gaussian blur. Divide the processed insulator images with defects into a training set, a validation set, and a test set according to the ratio of 8:2:2 to complete the construction of the insulator defect dataset. Similarly, divide the insulator images without defects into a training set, a validation set, and a test set according to the ratio of 8:2:2 to complete the construction of the insulator dataset.

3. For an overhead line insulator and its defect detection method based on the object detection algorithm according to claim 1, before obtaining an improved detection model, using the two datasets constructed in the previous step to iteratively train the model respectively to obtain the optimal detection models for insulators and insulator defects, and cascading the two network models to jointly complete the detection work step, the step further includes: Based on the YOLOv11s as the basic algorithm, use a lightweight neural network architecture to improve the feature extraction part of the YOLOv11s network, and use an efficient convolutional module and a small target detection layer to improve the feature fusion part of the YOLOv11s network. The lightweight neural network architecture is the FasterNet network, and the efficient convolutional module is the C3K2-Para module.

4. The overhead line insulator based on the object detection algorithm and its defect detection method according to claim 3, characterized in that: Use a lightweight neural network architecture to improve the feature extraction part of the YOLOv11s network. Specifically: Replace the feature extraction part of the original YOLOv11s network with the FasterNet network architecture. The input image data will pass through the FasterNet backbone, and the extracted image feature information will be sent to the SPPF pyramid pooling layer. The FasterNet network architecture consists of four stages. The input of the first stage is The input of the second stage is The input of the third stage is The input of the fourth stage is Each stage contains a FasterNet block. There is an embedding layer (a 4×4 regular convolution) or a fusion layer (a 2×2 regular convolution) before the FasterNet block for downsampling or channel adjustment. There is a PConv layer, two PWConv layers (or 1×1 regular convolutions) in each FasterNet block, and there are normalization layers and activation layers between every two PWConv layers. A structure similar to a T-shaped convolution is formed by two PWConv layers connected after each PConv. At the end of the FasterNet network architecture, there are a pooling layer, a 1×1 convolution layer, and a fully connected layer, which are jointly used for feature classification and transformation. To meet the lightweight requirements of the model, the most suitable network depth and width are selected in the FasterNet-T0 embedded network.

5. The overhead line insulator and its defect detection method based on the target detection algorithm according to claim 3, characterized in that: Use a small target detection layer to improve the feature fusion part of the YOLOv11s network. Specifically: To strengthen multi-scale feature fusion and improve the network's detection ability for small-object features, a P2 small-object detection layer is introduced in the feature fusion part. At the same time, the three detection heads of the original YOLOv11s are changed to four detection heads, and a P2 small-object detection head is added. After the upsampling operation on the last 80×80 feature layer in the feature fusion part of the original YOLOv11s, a new upsampling operation on a 160×160 feature layer is added, which can combine the P2 feature map in the Backbone feature extraction part, and then further extract and fuse features through the C3K2-Para module. At the same time, a P2 small-object detection head is added on the basis of the three detection heads of the original YOLOv11s, and the P2 small-object detection head is combined with the detection heads of other layers (P3, P4, P5) to fuse feature information of different scales. In addition, with the increase in the number of upsampling layers, an additional step of processing the P3 feature map is added. After all upsampling operations are completed, a convolution operation is performed on the feature map, and the feature map after the convolution operation is concatenated with the P3 feature map, and then passed through the C3K2-Para module, and the output is used as the detection head of the P3 layer.

6. The overhead line insulator based on the target detection algorithm and its defect detection method according to claim 3, characterized in that: Improve the feature fusion part of the YOLOv11s network using an efficient convolution module, specifically: Reconstruct all C3K2 modules in the feature fusion part. First, improve the Ghost Module, then design the C3K-Para module, and finally construct the C3K2-Para module based on the C3K-Para module. The Ghost Module uses ordinary convolution to complete the generation of the feature map of some channels and grouped convolution operations, while the improved Ghost Module uses Dynamic Conv (dynamic convolution) to replace ordinary convolution to complete the above operations in the Ghost Module. The basic definition of Dynamic Conv is described by the following formula: where x and y represent the input and output respectively, g is the activation function, and π k is the attention weight, and W and b are the weight matrix and bias vector respectively. The improved Ghost Module first uses Dynamic Conv to generate a small number of channels of intrinsic features, then performs a series of cheap convolutional linear transformations based on Dynamic Conv on these intrinsic features in each channel to generate Ghost features, and finally concatenates the feature maps obtained in the first two steps to generate more feature maps. After obtaining the improved Ghost Module, the improved Ghost Module is used to replace the Bottleneck in the C3K module to obtain the C3K-Para module. When the input enters the C3K-Para module, it will first be divided into two branches. One branch will pass through two consecutive improved Ghost Modules after passing through the CBS module, and the other branch will pass through the CBS module. The outputs of the two branches are concatenated, and after concatenation, they pass through the CBS module again, and the result is the output of the C3K-Para module. After obtaining the C3K-Para module, the improved Ghost Module and the C3K-Para module are integrated into the C3K2 module, replacing the C3K and Bottleneck in the original C3K2 module to construct the C3K2-Para module. The specific structure of the C3K2-Para module is determined by the parameter settings. When C3K = True, after the input passes through the CBS module, it will be evenly divided into two parts in the first dimension. One part directly enters the connection layer, and the other part becomes two branches. One branch also enters the connection layer, and the other branch contains two consecutive C3K-Para modules. The outputs passing through only the first C3K-Para module and the outputs passing through two consecutive C3K-Para modules enter the connection layer respectively. The above four parts of the outputs are concatenated in the connection layer and then input into the CBS module, and the result is the output of the C3K2-Para module when C3K = True. When C3K = False, after the input passes through the CBS module, it will be evenly divided into two parts in the first dimension. One part directly enters the connection layer, and the other part becomes two branches. One branch also enters the connection layer, and the other branch contains two consecutive improved Ghost Modules. The outputs passing through only the first improved Ghost Module and the outputs passing through two consecutive improved Ghost Modules enter the connection layer respectively. The above four parts of the outputs are concatenated in the connection layer and then input into the CBS module, and the result is the output of the C3K2-Para module when C3K = False. The setting of the C3K parameter in the network refers to the configuration file.

Images