Multi-scale feature enhancement fusion insulator small target defect detection network

Abstract

The invention relates to a multi-scale feature enhancement fusion insulator small target defect detection network, which is characterized in that a multi-scale feature enhancement module is designed in the network, the regional information of a defect target is enhanced, and the effective extraction of the detection network on bottom layer features is realized; a dynamic feature fusion module is designed, and efficient fusion of insulator defect features of different scales is achieved; in consideration of factors of relatively small positions and shapes of insulator defects, a Focaler-IoU loss function is applied to the head of the network, so that the capability of the network for positioning and detecting the insulator defects is enhanced.

Description

Technical Field

[0001] The present invention relates to a method for detecting defects of transmission line insulators, and particularly to a multi-scale feature enhanced fusion insulator small target defect detection network. Background Art

[0002] With the continuous improvement of the voltage level and power capacity of transmission lines, higher requirements are imposed on the reliable operation of transmission lines. Insulators, as important equipment for the safe operation of transmission lines, play a crucial role. However, due to internal and atmospheric overvoltages and the influence of the external environment, insulators often exhibit phenomena such as flashover and breakage, seriously endangering the safe operation of the power grid. Timely detection of insulator faults is of great significance for grid safety. UAV inspection has successfully replaced manual inspection, posing higher accuracy requirements for insulator target detection technology.

[0003] The insulator image detection technology based on deep learning is a current research hotspot. Deep learning methods are currently divided into two-stage detection networks and single-stage detection networks. Among them, Faster R-CNN (faster region-based convolutional neural network)[1] and Mask R-CNN are typical two-stage detection networks, and SSD (single shot multibox detector)[2] and YOLO (you only look once). Although the above research has solved problems such as low accuracy of insulator defect detection to a certain extent, there are still missed detections and false detections, and the detection accuracy in complex backgrounds and with less original data is not high. Summary of the Invention

[0004] Aiming at the problems of missed detections and false detections of small target defects of insulators in complex backgrounds, the present invention proposes a multi-scale feature enhanced fusion insulator small target defect detection network. A multi-scale feature enhancement module is designed to enhance the regional information of defect targets and effectively extract the underlying features of the detection network; a dynamic feature fusion module is designed to achieve efficient fusion of insulator defect features at different scales; the Focaler-IoU loss function is applied at the network head to enhance the network's ability to locate and detect insulator defects.

[0005] The technical solution of the present invention is: a multi-scale feature enhanced fusion insulator small target defect detection network.

[0006] The beneficial effects of the present invention are as follows: The present invention is a multi-scale feature enhanced fusion insulator small target defect detection network, which can effectively improve the detection ability of small target insulator defects and solve the problems of missed detection and false detection to a certain extent. A multi-scale feature enhancement module is designed to enhance the regional information of defect targets and effectively extract the underlying features of the detection network. A dynamic feature fusion module is designed to achieve efficient fusion of insulator defect features at different scales. The Focaler-IoU loss function is applied at the network head to enhance the network's positioning and detection ability for insulator defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a schematic diagram of the MSFE multi-scale feature enhancement module of the present invention.

[0008] Figure 2 It is a schematic diagram of the DFF dynamic feature fusion module of the present invention.

[0009] Figure 3 It is a network framework diagram of the multi-scale feature enhanced fusion insulator small target defect detection network of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0010] The network framework diagram of the multi-scale feature enhanced fusion insulator small target defect detection network is as follows:

[0011] 1) Set the network training cycle, divide the insulator samples into a training set and a test set, and input the training sample set into the detection network for training.

[0012] 2) Through the backbone part of the network, use the Focus slicing module, and then extract the optimized scale information of the sample features through the CSP, CBL, and MSFE structures. Finally, input the overall information into the Neck network through the SPP pooling operation.

[0013] 3) Enter the feature fusion network. The feature map passes through the MSFE feature extraction network, calculates the fusion features and outputs them to the CBL layer. The CBL layer further calculates the output result of the features, performs upsampling, and then inputs it into the DFF dynamic feature fusion module and outputs it to the next layer.

[0014] 4) Calculate the training weights through three detection heads with different scales in the Head layer and use the EIoU loss function.

[0015] 5) Repeat the operations in steps 1) to 4) until the training cycle is satisfied, so as to obtain the optimal training weights. Apply this weight to the test set, and it can accurately identify the insulator defect targets in the test set.

[0016] The construction of the network module structure is specifically elaborated as follows:

[0017] (1) In the network framework diagram of the multi-scale feature enhanced fusion insulator small target defect detection network designed by the present invention, the structural and functional analysis of the Focus, CSP1, CBL, and SPP modules in the backbone network is as follows.

[0018] Focus slices the input insulator image. First, the picture is transformed into a feature map of 320×320×12, then through a 3×3 convolution operation, the output channels are 32, and finally it becomes a feature map of 320×320×32, so that there is no information loss in the subsequent downsampling; the CBL structure consists of Conv + Bn + Leaky_relu activation function, which performs feature extraction operations on the feature map; the CSP1_X structure consists of two branches. One branch is composed of CBL + Resunit residual components + Conv, where Resunit is X residual components. For example, CSP1_1 indicates that it is composed of 1 residual component, and the other branch directly performs Conv operations. The two branches are then connected by Concat and output after BN + Leaky relu + CBL. CSP1_X can avoid the vanishing gradient caused by the deepening of the number of layers and extract finer-grained features; the SPP structure performs max pooling operations to solve the problem of inconsistent input image sizes.

[0019] (2) For the multi-scale feature enhanced fusion insulator small target defect detection network designed by the present invention, first, the MSFE multi-scale feature enhancement module is analyzed.

[0020] Figure 1 The schematic diagram of the MSFE multi-scale feature enhancement module is shown. Traditional feature extraction methods are prone to losing important information at small target defects of insulators and have poor feature extraction capabilities. MSFE makes full use of the two key attributes of locality and sparsity in the attention matrix, enabling the network to better capture the detailed information of small target defects and fully enhancing the image block information. Its calculation process is as follows:

[0021] First, the feature vector z is input into two branches respectively. In one branch, first through the CBL convolutional block, and then different feature channels of the output feature vector are input into the sliding window head along the linear layer, and self-attention operations are performed on the feature map at different dilation rates (g = 1, 2, 3) according to the matrix formed by the color window and the gray window. Finally, different features are connected and sent to the linear layer. The calculation formula is as follows:, and the specific calculation method is as follows:

[0022] (1)

[0023] Among them, SWDA is the sliding window expansion attention operation, which is used to query the sparse keys in the window. Q i, Ki ,V i is the feature map slice of the input i-th head, g i is the dilation rate of the i-th head. Finally, the result a i is connected to the input for feature aggregation in the linear layer, and the output result passes through the Fast Fourier Convolution Layer (FFC). Its use of the Fourier transform has translational invariance and computational efficiency, and can more efficiently capture the detailed features of defective targets.

[0024] Another branch is to avoid the overfitting problem caused by the deepening of the network layers. The feature vector z is directly input into the FFC. Finally, the results of the two branches are concatenated and input into the next layer through the BN, LR, and CBL convolution blocks.

[0025] (3) In the multi-scale feature enhanced fusion insulator small target defect detection network designed by the present invention, the upsample in the Neck layer is used to upsample the size of the feature map, increase the spatial resolution of the feature map, improve the accuracy of target detection, and is applied after the CBL operation.

[0026] (4) In the multi-scale feature enhanced fusion insulator small target defect detection network designed by the present invention, analyze the DFF dynamic feature fusion module:

[0027] The Neck network needs to fuse the low-level and high-level feature maps extracted from the backbone. However, the basic Concat connection operation only stacks the feature maps along the channel dimension and then uses standard convolution for channel output, without involving learned connections. However, there are multiple feature scale maps in the backbone layer. If the same weight is given to each feature for fusion, the larger-scale features will mask the smaller-scale features, and important target information will be lost.

[0028] The DFF dynamic feature fusion module is as Figure 2 shown, where X is the low-level feature map, Y is the high-level feature map from the upper layer. X After adding Y and W , it is input into the weight function calculation network G ( Z ) and the local channel context L ( Z ) are calculated through two network paths respectively to calculate the global channel context G ( Z ). Among them, in existing experiments, the Spatial Depth Convolution (SPD-Conv) shows superior performance in the small target object detection task and is used as the context aggregator. C×1×1is obtained through global average pooling, SPD-Conv, and ReLU activation layers; L ( Z ) Without global pooling, it has the same shape as the input features and can retain and highlight details in low-level features. The final weight function M ( Z ) is calculated, and the calculation formula is as follows:

[0029] (2)

[0030] where σ represents the Sigmoid normalization function that maps values to the range of 0 - 1; ⊕ represents broadcast addition that combines G(Z) and L(Z), and its calculation is as follows:

[0031] (3)

[0032] For the multi-scale feature enhanced fusion insulator small target defect detection network designed by the present invention, the specific calculation steps of the Focaler-IoU loss function are as follows:

[0033] The Focaler-IoU loss function can improve the detection performance of insulator defects by focusing on different regression samples. Focaler-IoU reconstructs the IoU loss through a linear interval mapping method, and the calculation formula is as follows:

[0034] (4)

[0035] where IoU focaler is the reconstructed Focaler-IoU, IoU is the original IoU value, [d, u] ∈ [0, 1], and by adjusting the values of d and u, IoU focaler can focus on different insulator defect samples. The loss is defined as follows:

[0036] (5)

[0037] The present invention designs a multi-scale feature enhanced fusion insulator small target defect detection network. It designs a multi-scale feature enhancement module to enhance the regional information of defect targets and effectively extract the underlying features of the detection network; designs a dynamic feature fusion module to achieve efficient fusion of insulator defect features at different scales; applies the Focaler-IoU loss function at the network head to enhance the network's ability to locate and detect insulator defects. The multi-scale feature enhanced fusion insulator small target defect detection network is as Figure 3 shown.

[0038] The instance dataset was verified according to the multi-scale feature enhanced fusion insulator small target defect detection network. The dataset consists of 3,000 insulator defect images from the IEEE insulator dataset, the Chinese Transmission Line Insulator Dataset (CPLID), and those collected from the Internet. However, the protection scope of the present invention is not limited to the above instance sets. The specific training parameter settings of the embodiments are shown in the following table.

[0039] Configuration Name Version Parameter GPU NVIDIAGeForce RTX 2080Ti CPU Intel(R) Xeon(R) Gold 6226R CPU @ 2.90GHz×2 CUDA 10.2 CuDNN 7.6.5

Claims

1. A multi-scale feature enhancement fusion insulator small target defect detection network, specifically including the following steps: 1) Setting the network training cycle, dividing the insulator samples into training sets and test sets, and inputting the training sample set into the detection network for training; 2) Through the backbone part of the network, using the Focus slicing module, CSP, CBL, MSFE structure to extract the scale information of sample feature optimization; finally, the overall information is input into the Neck network through the SPP pooling operation. 3) Entering the feature fusion network, the feature map passes through the MSFE feature extraction network, calculates the fusion feature and outputs it to the CBL layer. The CBL layer further calculates the result of the feature output, upsamples it, and then inputs it into the DFF dynamic feature fusion module and outputs it to the next layer. 4) The training weight is calculated by the three different scale detection heads in the Head layer and the Focaler-IoU loss function. 5) Repeat steps 1) to 4) until the training cycle is met, thereby obtaining the optimal training weight. Applying this weight to the test set can accurately identify the insulator defect targets of the test set.

2. The multi-scale feature enhanced fusion insulator small target defect detection network according to claim 1 is characterized in that: The step 2) MSFE multi-scale feature enhancement module improves the network's attention to the regional information of the defect target, and can focus on the two key attributes of locality and sparsity in the attention matrix, so that the network can better capture the detailed information of small defect targets.

3. The multi-scale feature enhanced fusion insulator small target defect detection network according to claim 1 is characterized in that: The step 3) can fully fuse the insulator defect target features extracted by the upper and lower network layers according to the designed DFF dynamic feature fusion module, and also has less calculation and parameter amount.

4. The multi-scale feature enhanced fusion insulator small target defect detection network according to claim 1 is characterized in that: The step 4) is solved by using the Focaler-IoU loss function, which takes into account the location of the insulator defect and the small shape. The original CioU loss function cannot meet the detection requirements. In order to enable the model to learn the target box features more comprehensively, the Focaler-IoU loss function is introduced.

5. The multi-scale feature enhanced fusion insulator small target defect detection network according to claim 1 is characterized in that: A multi-scale feature enhancement module is designed to enhance the regional information of defective targets and realize the effective extraction of underlying features by the detection network; a dynamic feature fusion module is designed to realize the efficient fusion of insulator defect features of different scales; considering the location and small shape of insulator defects, the Focaler-IoU loss function is applied to the network head to enhance the network's positioning and detection capabilities for insulator defects. On the basis of being able to accurately identify insulator equipment, the model's feature extraction and detection capabilities for small targets are improved, which to a certain extent makes up for the small application range of some detection models and reduces the missed detection and false detection of small targets with unclear insulator features.

Images