Insulator defect detection method based on multi-scale characteristics and channel perception

Abstract

The invention discloses an insulator defect detection method based on multi-scale features and channel perception, and belongs to the technical field of computer-aided algorithms. The method comprises the following steps: firstly, constructing a multi-scale feature fusion module, and fully extracting context information by combining multi-scale feature information and through a series of parallel deep convolution; secondly, a self-adaptive channel sensing down-sampling module is embedded in the backbone network, so that small-scale feature information is prevented from being lost during down-sampling of a feature map, and the sensing capability of the model to important channel information is improved; and finally, on the basis of adding a small-scale target detection layer, designing a lightweight shared phantom convolution detection head so as to improve the multi-scale target detection capability of the model under the condition of reducing the model parameter quantity. According to the method, a good effect is achieved on insulator defects of different types and multiple scales, the average detection precision is 91.4% and is improved by 5.2% compared with a reference model, and the overall performance is superior to that of other insulator defect detection algorithms.

Description

Technical Field

[0001] The invention relates to an insulator defect detection method based on multi-scale features and channel perception, and belongs to the technical field of computer-aided algorithms. Background Art

[0002] Insulators are mainly used to support conductors and provide electrical isolation protection in power transmission lines. However, due to long-term operation in harsh environments, insulators are prone to defects such as self-explosion, breakage, and flashover, which seriously affect the reliable operation of the power system. Since traditional manual inspection methods are difficult to operate in deep mountains and rivers, the use of computer vision algorithms on drones provides a new idea for insulator inspections in the power industry. Insulator defect detection using drones can cover areas that are difficult to detect with traditional manual inspections, providing important guarantees for the safe and stable operation of power networks.

[0003] With the great achievements of deep learning algorithms in image recognition and detection, deep learning has attracted widespread attention and in-depth research in the field of insulator defect detection. The current mainstream target detection algorithms mainly include two categories: one is the two-stage algorithm represented by Faster R-CNN, and the other is the single-stage algorithm represented by SDD and YOLO series. Zhou Ziqiang et al. improved the Faster R-CNN model through a multi-scale feature fusion strategy to achieve accurate recognition of small insulator targets. Julaiti Abilizi et al. introduced deformable convolution and balanced loss function to effectively improve the accuracy of insulator fault detection in aerial photos. Compared with the two-stage algorithm, which has complex model structure, large number of parameters and slow calculation speed, the single-stage insulator defect detection algorithm with faster reasoning speed and higher accuracy has become the main research direction. Among the existing methods, although the target detection algorithm has achieved the improvement of the accuracy of insulator defect detection, due to the small proportion of the insulator fault area in the actual drone pictures and the complex background interference, the accurate detection of multi-scale and multi-type insulator defects still remains a challenge. Summary of the invention

[0004] In view of the difficulties faced by the above research, in order to solve the problems of complex background, diverse defect types and large scale differences in insulator images in drone aerial photography, the present invention proposes an insulator defect detection method based on multi-scale features and channel perception, so as to improve the comprehensive performance of multi-scale and multi-category insulator defects in the insulator defect detection task, and selects YOLOv8-n with no anchor frame mechanism and small parameters as the baseline network.

[0005] The present invention is implemented by the following technical solutions: an insulator defect detection method based on multi-scale features and channel perception, through the following steps:

[0006] Firstly, a multi-scale feature fusion module is constructed to combine low-dimensional spatial information and high-dimensional semantic features, and to fully extract contextual information through parallel deep convolutions with different convolution kernel sizes. Secondly, an adaptive channel-aware downsampling module is designed to avoid the loss of small-scale defect information of insulators and realize feature map downsampling. At the same time, a channel attention mechanism is embedded to dynamically adjust different channels. Finally, a lightweight shared phantom convolution detection head is designed to share weight parameters to reduce the amount of model parameters. At the same time, the small-scale target detection layer is combined to improve the model's detection ability for multi-scale targets.

[0007] The details are as follows:

[0008] (1) Construct a multi-scale feature fusion module to aggregate features F2, F3, F4 or F3, F4, F5 in different dimensions, and use deep convolution with different convolution kernel sizes to capture contextual information:

[0009] F2, F3 and F4 represent low-dimensional features, current layer features and high-dimensional features respectively. High-dimensional features are used to extract abstract semantic information, while low-dimensional features are used to provide more accurate spatial information and local features. Therefore, after convolution or upsampling operations are performed on F2, F3 and F4 respectively to achieve channel and spatial dimension alignment, F2 is downsampled by convolution with a step size of 2, F3 uses convolution operations to achieve channel compression, and F4 achieves channel and spatial dimension alignment through upsampling and convolution operations, and the feature maps are obtained respectively. and The multi-scale features are aggregated by splicing in the channel dimension, and then the spliced ​​features are further extracted through convolution operation to obtain It is expressed as:

[0010]

[0011] Among them, Conv represents the convolution operation, Cat is the concatenation in the channel dimension, and finally F out A series of parallel deep convolutions are used to capture contextual information across multiple scales, and features of different receptive field sizes are integrated to obtain It is expressed as:

[0012]

[0013] Among them, DWConv k×k represents the depth convolution with different convolution kernel sizes, n represents the nth depth convolution to extract context features; at the same time, the feature F is connected using residual connection out With the processed features Add together to get the final output features Effectively retain the original information and introduce multi-scale information to improve the expressiveness of the model;

[0014] (2) For the input feature map The ACAD module downsamples it according to the downsampling scale factor. When the scale factor is 2, the feature map will be divided into 4 sub-feature maps of the same size, and the mapping slices are:

[0015]

[0016] In the formula, S represents the width or height size of the input feature map X, and the feature map is obtained by concatenating four sets of low-resolution feature maps in the channel dimension. The spatial dimension of the input image is converted into the channel dimension to achieve downsampling, which retains the complete spatial feature information of the original feature map, thus ensuring that key information is not lost during downsampling;

[0017] For the downsampled feature map Adaptive channel perception is achieved through the ECA module, so that the model focuses on channel features containing more information: First, the feature map x is averaged and pooled to obtain the aggregated features. Then use one-dimensional convolution with adaptive size to achieve cross-channel information interaction; finally use Sigmoid activation function to obtain channel weights Multiply the input feature map by the channel weight;

[0018] (3) First, the four feature layers of different scales output from the neck network are first adjusted to the same number of channels as the intermediate layer through their respective phantom convolutions; then, these feature layers are converged into two shared phantom convolution modules with a kernel size of 1 for further feature extraction; finally, the feature information extracted by the shared phantom convolution is input into the regression branch and the classification branch; in the regression branch, a convolution layer is used to predict the coordinate offset of the bounding box, and a Scale layer is used to scale the features to address the problem of inconsistent target scales detected by each detection head. The Scale layer introduces a learnable scaling factor and multiplies it with the input of any scale. As training progresses, the scaling factor will be updated according to the gradient descent algorithm to achieve scaling of the input features; the classification branch uses an independent convolution layer to predict the probability of each category; the branch structure enables the model to independently learn the localization and classification tasks, thereby effectively handling insulator defect targets of different scales.

[0019] Advantages and effects of the present invention: The method proposed in the present invention has achieved good results for insulator defects of different categories and multiple scales, with an average detection accuracy of 91.4%, which is 5.2% higher than the benchmark model, and the overall performance is better than other insulator defect detection algorithms. The improved model of the present invention learns target features more fully, improves the detection ability of occluded and small-scale targets, and effectively avoids the problems of false detection and missed detection of insulators and defective targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of insulator defect detection model based on multi-scale features and channel perception;

[0021] Figure 2 Schematic diagram of multi-scale feature fusion module;

[0022] Figure 3 Schematic diagram of adaptive channel-aware downsampling module;

[0023] Figure 4 Schematic diagram of lightweight shared phantom convolution detection head;

[0024] Figure 5 Sample diagrams for various types of insulators;

[0025] Figure 6 The PR curves before and after improvement. DETAILED DESCRIPTION

[0026] The present invention will be further described below by way of examples and drawings.

[0027] Example 1

[0028] The insulator defect detection method based on multi-scale features and channel perception goes through the following steps:

[0029] Firstly, a multi-scale feature fusion module is constructed to combine low-dimensional spatial information and high-dimensional semantic features, and to fully extract contextual information through parallel deep convolutions with different convolution kernel sizes. Secondly, an adaptive channel-aware downsampling module is designed to avoid the loss of small-scale defect information of insulators and realize feature map downsampling. At the same time, a channel attention mechanism is embedded to dynamically adjust different channels. Finally, a lightweight shared phantom convolution detection head is designed to share weight parameters to reduce the amount of model parameters. At the same time, the small-scale target detection layer is combined to improve the model's detection ability for multi-scale targets.

[0030] Specifically:

[0031] Insulator defect detection model

[0032] YOLOv8 is a fast single-stage target detection algorithm, which includes a series of target detection architectures and models, and is divided into five versions from small to large: n, s, m, l, and x by setting different network depth and width coefficients. Among them, YOLOv8n contains the least number of parameters and is easy to deploy in practical application scenarios. Therefore, in order to reduce computing resources and enhance the mobile terminal deployment of the model, the present invention is based on the YOLOv8-n network model, and proposes an insulator defect detection model based on multi-scale features and channel perception for insulator defect detection on transmission lines. Its overall structure is as follows Figure 1 As shown in the figure, it mainly consists of three parts: backbone network (Backbone), neck network (Neck) and detection layer (Prediction).

[0033] First, the Backbone part inputs three feature maps of different scales F2, F3, F4 or F3, F4, F5 into the Multi-scale Feature Fusion Module (MSFF), and uses a series of parallel deep convolutions to better realize context information extraction. Secondly, in order to prevent the loss of important feature information when the feature map is downsampled, the Adaptive Channel-Aware Downsampling Module (ACAD) is embedded in the Backbone. Then the Neck part constructs the Feature Pyramid Networks (FPN) and Path Aggregation Network (PAN) structures to effectively fuse image features at different levels. Finally, in order to maintain the model detection performance while reducing the number of model parameters, the Precision part uses the Lightweight Shared Ghost Convolutional Detection Head (LSGDH) to process feature maps at four scales of P2, P3, P4 and P5. The final prediction result includes the bounding box, category and location information of the target.

[0034] Multi-scale feature fusion module

[0035] In the insulator defect detection task, in order to cope with the challenges brought by multi-scale insulator defect targets and improve the detection effect of multi-scale insulator defects, a multi-scale feature fusion module is constructed to aggregate the features F2, F3, F4 or F3, F4, F5 in different dimensions respectively, and use deep convolution with different convolution kernel sizes to capture context information to further improve the feature extraction ability of the model. Taking the three different scale feature maps F2, F3 and F4 in the backbone network as the input of the MSSF module as an example, its structure is as follows Figure 2 shown.

[0036] F2, F3 and F4 represent low-dimensional features, current layer features and high-dimensional features respectively. High-dimensional features are used to extract abstract semantic information, while low-dimensional features are used to provide more accurate spatial information and local features. Therefore, after convolution or upsampling operations are performed on F2, F3 and F4 respectively to achieve channel and spatial dimension alignment, F2 is downsampled by convolution with a step size of 2, F3 uses convolution operations to achieve channel compression, and F4 achieves channel and spatial dimension alignment through upsampling and convolution operations, and the feature maps are obtained respectively. and The multi-scale features are aggregated by splicing in the channel dimension, and then the spliced ​​features are further extracted through convolution operation to obtain It is expressed as:

[0037]

[0038] Among them, Conv represents the convolution operation, Cat is the concatenation in the channel dimension, and finally F out A series of parallel deep convolutions are used to capture contextual information across multiple scales, and features of different receptive field sizes are integrated to obtain It is expressed as:

[0039]

[0040] Among them, DWConv k×k represents the depth convolution with different convolution kernel sizes, n represents the nth depth convolution to extract context features; at the same time, the feature F is connected using residual connection out With the processed features Add together to get the final output features The original information is effectively retained and multi-scale information is introduced to improve the expressiveness of the model.

[0041] Adaptive channel-aware downsampling module

[0042] In the insulator defect target detection task, due to the small proportion of insulator defect targets in the image and the interference of complex background, it is easy to lose important feature information after multiple convolution downsampling operations with a step size of 2, resulting in missed detection or false detection of small target insulator defect targets. In order to retain small-scale feature information in the feature extraction process of the backbone network and focus on channel features containing more information to further improve the model feature extraction capability, an adaptive channel-aware downsampling module is embedded in the backbone network. The process is as follows: Figure 3 shown.

[0043] For the input feature map The ACAD module downsamples it according to the downsampling scale factor. When the scale factor is 2, the feature map will be divided into 4 sub-feature maps of the same size, and the mapping slices are:

[0044]

[0045] In the formula, S represents the width or height size of the input feature map X, and the feature map is obtained by concatenating four sets of low-resolution feature maps in the channel dimension. The spatial dimension of the input image is converted into the channel dimension to achieve downsampling, which retains the complete spatial feature information of the original feature map, thereby ensuring that key information is not lost during downsampling.

[0046] Since the efficient channel attention mechanism (ECA) can improve the feature representation ability of the convolutional neural network without adding too many parameters, Adaptive channel perception is achieved through the ECA module, so that the model focuses on channel features containing more information: First, the feature map x is averaged and pooled to obtain the aggregated features. Then use one-dimensional convolution with adaptive size to achieve cross-channel information interaction; finally use Sigmoid activation function to obtain channel weights Multiply the input feature map by the channel weights.

[0047] Lightweight shared phantom convolutional detection head

[0048] The detection head of YOLOv8 uses the currently widely used decoupled head structure, which separates the classification and positioning tasks of the target, effectively improving the detection ability of the model. However, since each branch contains two convolutions and one convolution, this also leads to an increase in the number of parameters of the model. Therefore, based on the four detection heads, the present invention combines GhostConv in the lightweight GhostNet to construct a lightweight shared phantom convolution detection head. At the same time, according to Tian et al.’s paper, using Group Normalization (GN) normalization in the detection head can improve its positioning and classification capabilities, so GN normalization is used in the phantom convolution structure. While maintaining a high detection performance, LSGDH reduces the number of model parameters by sharing phantom convolution operations. Its structure is as follows: Figure 4 shown.

[0049] First, the four feature layers of different scales output from the neck network are first adjusted to the number of channels of the intermediate layer through their respective phantom convolutions; then, these feature layers are converged into two shared phantom convolution modules with a convolution kernel size of for further feature extraction; finally, the feature information extracted by the shared phantom convolution is input into the regression branch and the classification branch; in the regression branch, a convolution layer is used to predict the coordinate offset of the bounding box, and the Scale layer is used to scale the features to cope with the problem of inconsistent target scales detected by each detection head. The Scale layer introduces a learnable scaling factor and multiplies it with the input of any scale. As the training progresses, the scaling factor will be updated according to the gradient descent algorithm to achieve scaling of the input features; the classification branch uses an independent convolution layer to predict the probability of each category; the branch structure enables the model to independently learn positioning and classification tasks, thereby effectively processing insulator defect targets of different scales.

[0050] Experimental results and analysis

[0051] 1. Dataset establishment and experimental environment

[0052] The data set used in the present invention consists of two parts, with a total of 2650 insulator defect samples. One part is the insulator defect image dataset (IDID) downloaded from the public network, which contains 1600 images of insulator damage and flashover. The other part comes from the insulator defect dataset taken by a power grid company, which contains 525 insulator defect images. Since this insulator defect dataset mainly consists of insulator self-explosion defects and the number of self-defect samples is small, in order to improve the training effect of the network model and avoid overfitting, this dataset is flipped upside down or left to right for image enhancement operations, and finally 1050 insulator defect images are obtained. The various types of insulator defect samples and the number of samples are as follows: Figure 5 As shown and Table 1.

[0053] Table 1 Proportion of each sample

[0054]

[0055] All images were annotated by LabelImg software, and the label categories included normal insulator, damage, flashover, and explosion defects. During the training and evaluation of the model, the dataset was randomly divided into training, validation, and test sets in a ratio of 6:2:2. All experiments were based on the Pytorch deep learning framework under the Ubuntu operating system and trained on a GPU configured as RTX 3060 (12GB). The network model was trained by stochastic gradient descent (SGD) with a momentum term of 0.937, an initial learning rate of 0.01, an input image size of 640×640, the number of training rounds was set to 300, and the training batch was 16.

[0056] 2. Evaluation indicators

[0057] In order to comprehensively evaluate the comprehensive performance of the algorithm of the present invention, the present invention uses average precision (AP), mean average precision (mAP), model size (Size), floating point operations (FLOPs) and model parameters (Parameters, Params) to comprehensively evaluate the performance indicators of the model. The detection capability of the model is verified by AP and mAP, and the comprehensive complexity of the model is analyzed based on the Size, FLOPs and Params indicators.

[0058]

[0059] The calculation formulas for AP and mAP are shown in formulas (4) and (5), respectively, where P is the precision, R is the recall, AP represents the area under the curve with precision as the ordinate and recall as the abscissa, mAP is the average of the AP values ​​of all categories, and c is the number of categories.

[0060] 3. Experimental comparison

[0061] Ablation experiment

[0062] In order to further verify the effectiveness of the proposed method in the insulator defect detection task, the present invention sets up four groups of ablation experiments to analyze the impact of different improvement methods on the model performance. The ablation experiment results are shown in Table 2. The precision-recall rate (PR) curve of the baseline model is shown in Figure 6 As shown in (a), the PR curve of the improved model is Figure 6 (b) as shown.

[0063] Table 2 Ablation experiment

[0064]

[0065] As can be seen from Table 2, the mAP of the model increased by 1.5% after the MSFF module was adopted, indicating that the MSFF module can integrate multi-level feature information and comprehensively extract contextual information through multiple parallel deep convolution operations. Subsequently, the ACAD module was introduced into the backbone network to reduce the loss of information during downsampling and guide the model to focus on key channel information. The mAP of the model increased by 1.0%. After two improvement methods, the mAP of the model increased from 86.2% to 88.7%. However, this also led to an increase of 0.49M in the number of model parameters. Therefore, under the condition of adding a small-scale target detection layer, LSGDH embeds shared phantom convolutions in the detection head to reduce the number of model parameters while enhancing the model's detection performance for multi-scale targets, further improving the model's mAP to 91.4%, and the number of parameters decreased by 0.15M compared to the baseline model. The above experiments verify the effectiveness of the various improvement methods of the YOLOv8-n network of the present invention, and the number of parameters has also decreased, which meets the lightweight deployment conditions of the mobile terminal for drone power inspection. And by Figure 6 It can be seen that the improved model has improved detection accuracy in each category compared with the baseline model, and mAP has increased by 5.2%. In general, the improved model achieves better detection accuracy compared with the baseline model.

[0066] Comparative experiment of multi-scale feature fusion module

[0067] In order to further verify the influence of the convolution kernel size of the parallel deep convolution in the MSFF module on the model performance, the present invention sets four groups of deep convolutions with different convolution kernel sizes in MFFM and conducts comparative analysis. Table 3 shows the comparison results of the performance of each group.

[0068] Table 3 Comparative experiment of multi-scale feature fusion module

[0069]

[0070] According to the results in Table 3, after four sets of parallel deep convolutions with different convolution kernel sizes are successively embedded in MSSF, compared with the baseline network model, the mAP is improved by 1.1%, 1.5%, 1.3% and 1.8%, reaching 87.1%, 87.7%, 87.5% and 88% respectively, indicating that MSFF can capture contextual information across multiple scales through a series of parallel deep convolutions to improve the model detection accuracy. When the convolution kernel sizes are 7, 9, 11 and 13, the average accuracy of the model is the highest. However, using larger convolution kernels will increase the number of parameters and floating-point operations of the model. Therefore, considering the model performance comprehensively, deep convolutions with convolution kernel sizes of 3, 5, 7 and 9 are selected to achieve a balance between model detection accuracy and complexity.

[0071] Attention mechanism comparison experiment

[0072] In order to study the influence of different attention mechanisms on the model performance, the present invention introduces four attention mechanisms in ACAD respectively, and the comparison results are shown in Table 4. As can be seen from Table 4, after embedding the squeeze and excitation attention (SE), coordinate attention (CA), convolutional block attention module (CBAM), simple attention module (SimAM), efficient multi-scale attention (EMA) and ECA module, the mAP of the model is increased by 0.1%, 1.1%, 1.5%, 1.7%, respectively, indicating that the introduction of the attention mechanism can improve the feature expression ability of the model and thus improve the average detection accuracy of insulator defects. Compared with the increase in the number of model parameters caused by embedding other attention mechanisms, the ECA attention mechanism improves the model detection accuracy to 88.4% without increasing the number and size of model parameters. Therefore, the ECA attention mechanism guides the model to pay attention to important channel feature information to improve the detection accuracy of the model.

[0073] Table 4 Comparative experiment of attention mechanism

[0074]

[0075] Comparative experiments on different network models

[0076] In order to verify the detection performance of the model of the present invention for insulator defects, the model of the present invention is compared with the current mainstream target algorithms YOLOv3-tiny, YOLOv5-s, YOLOv7-tiny, YOLOv8-n, YOLOv8-s and YOLOv10-s based on the data set of the present invention in the same experimental environment. The experimental results are shown in Table 5.

[0077] Table 5 Comparative experiments of different network models

[0078]

[0079] As can be seen from Table 5, the detection accuracy of the proposed model for normal insulators and insulator defects has been greatly improved, among which the average detection accuracy for damage, flashover and self-explosion defects is the highest, which is 93.5%, 88.0% and 92.3% respectively. The overall average detection accuracy is the highest, which is 91.4%. Compared with other network models, the proposed model of the present invention achieves the best results in terms of parameter quantity and model size. When the parameter quantity is 33%, 40.7%, 47.5%, 25.7% and 35.6 of YOLOv3-tiny, YOLOv5-s, YOLOv7-tiny, YOLOv8s and YOLOv10-s, the mAP is increased by 12.2%, 3.6%, 3.2%, 0.8% and 3.8 respectively. In terms of floating point operations, the FLOPs of the proposed model is 13.9G, which is much better than YOLOv5-s, YOLOv8-s and YOLOv10-s. It can be seen that the improved model has more comprehensive performance in the insulator defect detection task.

[0080] 4. Detection effect analysis

[0081] In order to further compare the detection performance of the improved network model in the insulator defect detection task, five insulator images under complex backgrounds were selected from the test set for visual comparison.

[0082] It can be seen from the first and second rows that under complex backgrounds, the benchmark model does not learn enough about the features of insulator targets and self-explosion defects, resulting in false detections, and missed detections of occluded insulators and insulator self-explosion defects; in the third row, due to the lack of extraction of contextual information of insulator features by the benchmark model, the occluded target is mistakenly identified as a self-explosion defect. Since the flashover and damage defect areas in the fourth and fifth row images are small, the benchmark model has weak detection capabilities for small-scale flashover and damage defects, resulting in missed detections by the benchmark model. The improved model of the present invention learns target features more fully, improves the detection capabilities of occluded and small-scale targets, and effectively avoids the problems of false detection and missed detection of insulators and defective targets.

Claims

1. An insulator defect detection method based on multi-scale features and channel perception, characterized in that: The following steps are involved: Firstly, a multi-scale feature fusion module is constructed to combine low-dimensional spatial information and high-dimensional semantic features, and to fully extract contextual information through parallel deep convolutions with different convolution kernel sizes. Secondly, an adaptive channel-aware downsampling module is designed to avoid the loss of small-scale defect information of insulators and realize feature map downsampling. At the same time, a channel attention mechanism is embedded to dynamically adjust different channels. Finally, a lightweight shared phantom convolution detection head is designed to share weight parameters to reduce the amount of model parameters. At the same time, the small-scale target detection layer is combined to improve the model's detection ability for multi-scale targets.

2. The insulator defect detection method based on multi-scale features and channel perception according to claim 1 is characterized in that: The specific steps are as follows: (1) Construct a multi-scale feature fusion module to aggregate features F2, F3, F4 or F3, F4, F5 in different dimensions, and use deep convolution with different convolution kernel sizes to capture contextual information: F2, F3 and F4 represent low-dimensional features, current layer features and high-dimensional features respectively. High-dimensional features are used to extract abstract semantic information, while low-dimensional features are used to provide more accurate spatial information and local features. Therefore, after convolution or upsampling operations are performed on F2, F3 and F4 respectively to achieve channel and spatial dimension alignment, F2 is downsampled by convolution with a step size of 2, F3 uses convolution operations to achieve channel compression, and F4 achieves channel and spatial dimension alignment through upsampling and convolution operations, and the feature maps are obtained respectively. and The multi-scale features are aggregated by splicing in the channel dimension, and then the spliced ​​features are further extracted through convolution operation to obtain It is expressed as: Among them, Conv represents the convolution operation, Cat is the concatenation in the channel dimension, and finally F out A series of parallel deep convolutions are used to capture contextual information across multiple scales, and features of different receptive field sizes are integrated to obtain It is expressed as: Among them, DWConv k×k represents the depth convolution with different convolution kernel sizes, n represents the nth depth convolution to extract context features; at the same time, the feature F is connected using residual connection out With the processed features Add together to get the final output features (2) For the input feature map The ACAD module downsamples it according to the downsampling scale factor. When the scale factor is 2, the feature map will be divided into 4 sub-feature maps of the same size, and the mapping slices are: In the formula, S represents the width or height size of the input feature map X, and the feature map is obtained by concatenating four sets of low-resolution feature maps in the channel dimension. For the downsampled feature map Adaptive channel perception is achieved through the ECA module, so that the model focuses on channel features containing more information: First, the feature map x is averaged and pooled to obtain the aggregated features. Then use one-dimensional convolution with adaptive size to achieve cross-channel information interaction; finally use Sigmoid activation function to obtain channel weights Multiply the input feature map by the channel weight; (3) First, the four feature layers of different scales output from the neck network are first adjusted to the number of channels of the intermediate layer through their respective phantom convolutions; then, these feature layers are converged into two shared phantom convolution modules with a convolution kernel size of 1 for further feature extraction; finally, the feature information extracted by the shared phantom convolution is input into the regression branch and the classification branch; in the regression branch, a convolution layer is used to predict the coordinate offset of the bounding box, and the Scale layer is used to scale the features to cope with the problem of inconsistent target scales detected by each detection head. The Scale layer introduces a learnable scaling factor and multiplies it with the input of any scale. As training progresses, the scaling factor will be updated according to the gradient descent algorithm to achieve scaling of the input features; the classification branch uses an independent convolution layer to predict the probability of each category.

Images