A real-time segmentation method and system for composite insulators based on DeepLabV3+
Through the improved DeepLabV3+ network, combined with MobileNetV2 and ECA modules, the problem of insufficient real-time and accuracy in real-time segmentation of composite insulators is solved, and efficient and accurate segmentation is achieved in complex backgrounds, and suitable for drone inspections.
Patent Information
- Application Number
- CN202210181881.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-26
- Publication Date
- 2024-12-27
- Estimated Expiration
- 2042-02-26
AI Technical Summary
The prior art has problems of insufficient real-time and accuracy in real-time segmentation of composite insulators, especially in complex contexts, where traditional methods rely on image preprocessing and manual intervention is complex, resulting in poor segmentation accuracy.
Using the real-time segmentation method of composite insulators based on DeepLabV3+, by replacing the MobileNetV2 network with a lightweight neural network, the lightweight and efficient channel attention module ECA is introduced, and the PointRend fine segmentation module is added to the output end to improve the semantic segmentation results and build an improved DeepLabV3+ network is trained and segmented in real time.
It significantly improves the real-time and accuracy of composite insulator segmentation, reduces the calculation time and number of parameters, enhances feature representation ability, improves segmentation accuracy, and realizes efficient patrol on UAV embedded equipment.
Smart Images
Figure CN114549563B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of power transmission lines, and specifically relates to a real-time segmentation method and system for composite insulators based on DeepLabV3+, which can be applied to intelligent inspection and detection by unmanned aerial vehicles in power transmission line operation and maintenance. Background Art
[0002] With the increase in the scale of use and the years of operation, composite insulators have frequently experienced defects such as abnormal heating in recent years. The further development of defects may lead to serious faults such as internal insulation breakdown and string breakage. Using infrared detection technology to detect defects in insulators is a relatively mature live detection technology. The infrared image of the composite insulator can intuitively reflect the overheating defects of the insulator, so as to detect the defects in advance and prevent their deterioration. Therefore, in order to detect overheating defects in infrared composite insulators, it is necessary to separate the composite insulators from the complex infrared background to avoid identifying the hot spots in the irrelevant background as insulator overheating defects. In addition, in the past, the inspection of power transmission lines in my country generally adopted manual inspection, but this method is inefficient and will also threaten the personal safety of the staff. In recent years, my country has increased the investment in the research and development of drone line inspection technology. At present, most of the maintenance departments of domestic power companies use drones for inspection. In order to make drone inspection more efficient and intelligent, it has become a research hotspot to carry embedded devices on drones for real-time inspection.
[0003] Traditional insulator segmentation mainly includes edge detection and template matching methods, which require manual confirmation of segmentation results and adjustment of annotation results. Although the edge detection method can be applied to many insulator shapes, it cannot be directly used as a reference for inspection personnel due to its low accuracy. The model matching rule can only be applied to one or several specific insulator shapes. If there is a complex insulator background, it cannot achieve a good accuracy.
[0004] The difficulty of real-time segmentation of composite insulators lies mainly in the fact that composite insulators have more complex edge shapes than other insulators, and the available computing resources in embedded devices are less than those in the host computer. Compared with traditional image processing methods, semantic segmentation shows greater potential in real-time segmentation of composite insulators, because deep learning performs model training on large-scale insulator images, and its segmented insulators can achieve high accuracy. However, most scholars' research uses algorithms on the host computer, which is not conducive to efficient on-site inspections. In addition, due to the complex structure of composite insulators, there is still a lot of room for improvement in accuracy. Therefore, developing a real-time segmentation method for composite insulators has high practical application value.
[0005] The Chinese patent application number CN201911104286.7 discloses a composite insulator overheating defect detection method based on instance segmentation, which uses a segmentation network to segment composite insulators. The Mask R-CNN deep learning network, which is widely used in the field of deep learning, is further improved on the basis of its original framework. The network has significant advantages in the accuracy and computational efficiency of instance segmentation. The design of the Mask R-CNN network is relatively simple. On the basis of Faster R-CNN, it adds a branch for semantic segmentation on the original two branches (classification and bounding box regression). Secondly, the ROI Pooling layer is improved, and the ROI Align method is proposed. When downsampling, the pixels are aligned to make the pixel segmentation more accurate. The patent uses the Mask R-cnn instance segmentation network to complete the segmentation of composite insulators. However, since the backbone of Mask R-cnn is the two-stage Faster R-cnn, its FPS is very low, that is, the real-time performance is very poor.
[0006] The Chinese patent application number CN201910062932.1 discloses an insulator identification and positioning method based on shape features and image segmentation. The method belongs to the field of image processing technology and power grid detection. The specific steps are as follows: (1) preprocessing the original insulator image to reduce noise interference and improve image contrast; (2) introducing an adaptive particle swarm optimization algorithm and combining it with the maximum inter-class variance method to segment the insulator image; (3) forming a binary image through morphological filtering and area filtering, extracting the skeleton of the binary image, and initially positioning the insulator area according to the Hough transform result; (4) relying on the shape features of the insulator to achieve accurate positioning of the insulator. This method makes full use of the prior knowledge of the shape features of the insulator and the technical methods of image processing to achieve automatic identification and positioning of the insulator string, laying the foundation for subsequent insulator fault diagnosis. The invention is feasible and improves the accuracy and efficiency of positioning. This patent achieves good segmentation of insulators under simple backgrounds. However, this traditional method relies too much on image preprocessing, which directly affects the segmentation effect. Image preprocessing is largely done by manual selection of shape feature operators. Therefore, the feature selection process is complicated, and the segmentation accuracy is poor for the complex and changeable backgrounds during actual inspections. Summary of the invention
[0007] The object of the present invention is to provide a real-time segmentation method and system for composite insulators based on DeepLabV3+, which is conducive to improving the real-time performance and accuracy of composite insulator segmentation.
[0008] To achieve the above object, the technical solution adopted by the present invention is: a real-time segmentation method for composite insulators based on DeepLabV3+, comprising the following steps:
[0009] S1. Obtain infrared images of power equipment containing composite insulators taken during power inspection and construct an original data set;
[0010] S2, perform data augmentation on the original data set to obtain a training data set;
[0011] S3. Build an improved DeepLabV3+ network: replace the backbone network of DeepLabV3+ with a lightweight neural network MobileNetV2 to improve real-time performance, and introduce a lightweight and efficient channel attention module ECA to achieve local cross-channel interaction without dimensionality reduction. Then, add the Pointrend fine segmentation module to the output of DeepLabV3+ for post-processing to further improve the semantic segmentation results. Train the improved DeepLabV3+ network using the training data set to obtain a trained improved DeepLabV3+ network.
[0012] S4. The infrared images of the captured power equipment are processed through the trained improved DeepLabV3+ network to segment the composite insulators in real time.
[0013] Furthermore, the specific implementation method of step S1 is: obtaining infrared images of power equipment taken during the power inspection process, and screening out infrared images of power equipment containing composite insulators to form an original data set; using an image annotation tool to annotate the composite insulator core rod and umbrella disk parts in the original data set image.
[0014] Furthermore, the specific implementation method of step S2 is: perform data augmentation on the original data set to increase the data volume and form a training data set; the data augmentation method is: read the pictures in the original data set, flip the pictures horizontally, flip the pictures vertically, rotate them to a set angle or randomly crop them to obtain new pictures; and the newly obtained pictures and the pictures in the original data set together form a training data set.
[0015] Furthermore, in step S3, when constructing the improved DeepLabV3+ network, MobileNetV2 is integrated with DeepLabV3+, and the MobilenetV2 network is used as the encoder, and linear is used to prevent relu from destroying features; deep separable convolution is used to reduce the amount of calculation and parameters and increase the number of channels to obtain more features in the image.
[0016] Furthermore, the improved DeepLabV3+ network includes an encoder and a decoder, the encoder includes a dilated convolution and a spatial pyramid combined with a MobileNetV2 convolutional network, the dilated convolution extracts features calculated at arbitrary resolution in MobileNetV2, and before global pooling, uses the ratio of the spatial resolution of the input image to the spatial resolution of the final output image as the output stride, and controls the density of the encoder features according to the budget of computing resources; the spatial pyramid uses dilated convolutions with different dilation rates to perform multi-scale fusion of features calculated by MobileNetV2, enriching semantic information while balancing accuracy and running time; the decoder consists of two convolution structures with different convolution kernels, which are used to reduce the number of channels using the same convolutional network.
[0017] Furthermore, in step S3, an ECA module is introduced into MobileNetV2. The implementation method of ECA is as follows: first, global average pooling is performed on each feature, then adaptive Kernel operation is performed, each feature is obtained through a sigmoid function, and then forward propagation is performed to assign the importance weight of each feature to the original feature map.
[0018] Furthermore, in step S3, the segmentation result is post-processed using the PointRend fine segmentation module to make the edge of the segmentation result smoother and improve the segmentation accuracy; PointRend introduces the idea of image rendering into semantic segmentation, uses bilinear interpolation to upsample the fine-grained features output by the network, selects several most uncertain points in the upsampling result as boundary segmentation difficulty points, and combines the difficulty points directly obtained from the fine-grained features to use a multi-layer perceptron for label prediction; repeats the above upsampling-point selection-label prediction process until the required resolution is reached; and by implementing adaptive subdivision in the upsampling process to retain contour information as much as possible, boundary points are retained in the segmentation to improve segmentation accuracy.
[0019] The present invention also provides a real-time segmentation system for composite insulators based on DeepLabV3+, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above-mentioned method steps can be implemented.
[0020] Compared with the prior art, the present invention has the following beneficial effects: the present invention provides a real-time segmentation method and system for composite insulators based on DeepLabV3+. The present invention uses the MobilenetV2 network to replace the traditional Xception network in the DeepLabV3+ encoder part, which greatly reduces the calculation time and the number of parameters and improves the real-time performance of the segmentation. At the same time, the super channel attention module (ECA-Net) is introduced to further enhance the discrimination ability of the feature representation for scene segmentation. In addition, the Pontrend segmentation module is introduced to refine the edge of the insulator and improve the segmentation accuracy. By deploying the present invention on the embedded device of the drone, the efficiency and accuracy of the inspection can be improved while reducing the workload of manual inspections. It has strong practicality and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a flow chart of the method implementation of the embodiment of the present invention.
[0022] Figure 2 It is a schematic diagram of a network model in which the backbone network of DeepLabV3+ is replaced by MobileNetV2 in an embodiment of the present invention.
[0023] Figure 3 Schematic diagram of an ECA module in an embodiment of the present invention.
[0024] Figure 4 Schematic diagram of the PointRend upsampling adaptive subdivision process in an embodiment of the present invention.
[0025] Figure 5 It is a flow chart of the implementation of PointRend in an embodiment of the present invention.
[0026] Figure 6 It is a diagram showing the segmentation effect of the composite insulator by DeepLabV3+ before and after the improvement in the embodiment of the present invention. DETAILED DESCRIPTION
[0027] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0028] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.
[0029] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.
[0030] The implementation of composite insulator segmentation on an embedded platform is similar to a multi-objective optimization problem, which requires considering segmentation performance, network parameter size, and running time at the same time. Therefore, this embodiment provides a real-time composite insulator segmentation method based on DeepLabV3+ to achieve high-precision extraction of composite insulators. Figure 1 As shown, the method comprises the following steps:
[0031] S1. Obtain infrared images of power equipment containing composite insulators taken during power inspection and construct the original data set.
[0032] S2. Perform data augmentation on the original data set to obtain a training data set.
[0033] S3. Build an improved DeepLabV3+ network: replace the backbone network of DeepLabV3+ with the lightweight neural network MobileNetV2 to improve real-time performance, and introduce a lightweight and efficient channel attention module ECA to achieve local cross-channel interaction without dimensionality reduction. Then, add the Pointrend fine segmentation module to the output of DeepLabV3+ for post-processing to further improve the semantic segmentation results. Train the improved DeepLabV3+ network using the training data set to obtain a trained improved DeepLabV3+ network.
[0034] S4. The infrared images of the captured power equipment are processed through the trained improved DeepLabV3+ network to segment the composite insulators in real time.
[0035] The specific implementation method of step S1 is as follows: first, an infrared image of the power equipment taken during the power inspection is obtained. There are various types of power equipment, and the present invention only selects the infrared images of the power equipment containing composite insulators to form an original data set. Then, an image annotation tool is used to annotate the core rod and the umbrella disk of the composite insulator in the original data set.
[0036] The specific implementation method of step S2 is: perform data augmentation on the original data set to increase the data volume and form a training data set. In order to avoid overfitting of the model and improve the generalization ability of the model, the original data set is augmented by a data augmentation algorithm. The method of data augmentation is: read a set number of pictures in the original data set, and perform horizontal flipping, vertical flipping, rotating a set angle or random cropping on the pictures to obtain new pictures; the newly obtained pictures and the pictures in the original data set together form a training data set.
[0037] In step S3, when constructing the improved DeepLabV3+ network, MobileNetV2 is integrated with DeepLabV3+, and the MobilenetV2 network is used as the encoder. Linear is used to prevent relu from destroying features. Depthwise separable convolution is used to reduce the amount of computation and parameters and increase the number of channels to obtain more features in the image.
[0038] The underlying network of DeepLabV3+ is highly adaptable. In order to pursue segmentation accuracy, researchers have integrated ResNet into it. This type of model has high classification accuracy, but the model depth continues to deepen, which leads to increased model complexity. Complex segmentation models are restricted by hardware memory and have high requirements for mobile or embedded devices, and cannot meet the low-latency and high-response rate segmentation requirements in specific scenarios. In order to meet the requirements of real-time segmentation, the first step in constructing an improved DeepLabV3+ network in the present invention is to integrate the lightweight neural network MobileNetV2 with it. A lightweight MobilenetV2 network is introduced as an encoder, and linear is used to prevent relu from destroying features; deep separable convolution is used to reduce the amount of computation and parameters and increase the number of channels to obtain more features in the image. Improved network models such as Figure 2 shown.
[0039] The improved DeepLabV3+ network includes an encoder and a decoder (structure C), and the encoder includes structure A and structure B.
[0040] Structure A represents a dilated convolution, which extracts features calculated at any resolution in MobileNetV2. Before global pooling, the ratio of the input image spatial resolution to the final output image spatial resolution is used as the output stride to control the density of encoder features according to the budget of computing resources. For semantic segmentation tasks, after removing the span in the last one or two blocks, an output with a stride of 16 is used for more intensive feature extraction. Because when the Decoder output stride is 9, the segmentation performance is improved compared to the output stride of 18, but the computational complexity is increased. Therefore, in the present invention, the encoder uses an output stride of 18, which has the advantage of balancing segmentation accuracy and speed.
[0041] Structure B represents a spatial pyramid combined with the MobileNetV2 convolutional network. This structure uses dilated convolutions with different dilation rates to fuse the features calculated by MobileNetV2 at multiple scales, enriching the semantic information while effectively balancing accuracy and running time.
[0042] Structure C consists of two convolution structures with different convolution kernels. Its function is to use the same convolution network to reduce the number of channels and solve the problem of difficult training caused by low-level features containing a large number of channels. In the case of limited GPU resources, the output stride of the model is set to 4, and the density feature map is appropriately traded off to simplify the decoder.
[0043] In this embodiment, a lightweight and efficient channel attention (ECA) module is introduced into MobileNetV2 to improve network performance. ECA-Net mainly makes some improvements to the SE-Net module, and proposes a local cross-channel interaction strategy (ECA module) without dimensionality reduction and a method for adaptively selecting the size of a one-dimensional convolution kernel, thereby achieving performance improvement. This module only adds a small amount of parameters, but can obtain significant performance gains. Avoiding dimensionality reduction is very important for learning channel attention, and appropriate cross-channel interaction can significantly reduce the complexity of the model while maintaining performance. Therefore, through a local cross-channel interaction strategy without dimensionality reduction, it can be effectively implemented through one-dimensional convolution. Experiments have shown that the use of the ECA module in the present invention is efficient and feasible.
[0044] Efficient Channel Attention Module (ECA) Figure 3 As shown in the figure, the implementation process is as follows: first perform global average pooling on each feature, then perform adaptive Kernel operation, then obtain each feature through the sigmoid function, and then perform forward propagation to assign the importance weight of each feature to the original feature map.
[0045] After channel-level global average pooling without dimensionality reduction, ECA captures local cross-channel interaction information by considering each channel and its k neighbors. Practice has shown that this method guarantees model efficiency and computational effect. ECA can be effectively implemented by fast 1D convolution of size k, where the convolution kernel size is k, representing the coverage of local cross-channel interactions, i.e., how many neighbors near this channel participate in the attention prediction of this channel. Compared with the backbone model, the deep CNN with ECA module introduces few additional parameters and almost negligible computation, while bringing performance gain. The ECA module learns effective channel attention by avoiding reducing the channel dimension, while capturing cross-channel interaction information in an extremely lightweight way.
[0046] The ECA module first squeezes the feature map obtained by convolution to obtain channel-level global features, then performs an Excitation operation on the global features to learn the relationship between channels and obtain the weights of different channels. Finally, it multiplies the original feature map to obtain the final features. This attention mechanism allows the model to pay more attention to the channel features with the largest amount of information and suppress those unimportant channel features.
[0047] The original image is output as a predicted image by DeepLabV3+ after being improved by the above method. Since DeepLabV3+ uses a lightweight MobilenetV2 network model as an encoder, the calculation time and the number of parameters are greatly reduced, but the image edge segmentation effect is not good, resulting in a decrease in segmentation accuracy. Therefore, the present invention uses the PointRend fine segmentation module to post-process the segmentation results to make the edges of the segmentation results smoother and improve the segmentation accuracy.
[0048] PointRend introduces the idea of image rendering into semantic segmentation, uses bilinear interpolation to upsample the fine-grained features output by the network, selects several most uncertain points in the upsampling results as difficult points for boundary segmentation, and uses a multi-layer perceptron to predict labels in combination with the difficult points directly obtained from the fine-grained features; repeats the above upsampling-point selection-label prediction process until the required resolution is reached; and achieves adaptive subdivision in the upsampling process to retain contour information as much as possible, retains boundary points in segmentation, and improves segmentation accuracy.
[0049] PointRend introduces the idea of image rendering into semantic segmentation, uses bilinear interpolation to upsample the fine-grained features output by the CNN network, selects several most uncertain points (prediction probability close to 0.5) from the upsampling results as difficult points for boundary segmentation, and uses a multi-layer perceptron (MLP) for label prediction in combination with the difficult points directly obtained from the CNN fine-grained features. Repeat the above upsampling-point selection-label prediction process until the required resolution is reached. By implementing adaptive subdivision in the upsampling process, contour information can be retained as much as possible, boundary points can be retained in the segmentation, and segmentation accuracy can be improved. PointRend upsampling adaptive subdivision process is as follows: Figure 4 shown.
[0050] like Figure 5 As shown, the implementation process of PointRend is:
[0051] a. Perform unstable sorting on the points on the feature map and select N points. N is 8096 in the code.
[0052] b. The features of N points corresponding to the first layer of MobilenetV2 are proposed.
[0053] For example, if the backbone network used is MobilenetV2, let's take it as an example. This network outputs c1, c2, c3, c4. Among them, c1 is the featuremap (1 / 4) at a higher resolution, and c4 is the final featuremap (1 / 16). The corresponding features of the above N points on these two images are extracted.
[0054] c. Glue the features of the corresponding positions of N points together. This is implemented by the torch.cat function. For example, if the feature of C1 is [1, 19, 8096] and the feature of C2 is [1, 1256 8096], then the result is [1, 275, 8096].
[0055] d. Use MLP for segmentation prediction.
[0056] The present invention also provides a real-time segmentation system for composite insulators based on DeepLabV3+, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the above-mentioned method steps can be implemented.
[0057] MIOU (mean pixel intersection over union) and ACC (pixel accuracy) are the two most important evaluation indicators for semantic segmentation. Through experimental verification, after the improvement, MIOU increased from 60% to 82%, which is a significant improvement. The ACC also increased from 95% to 97%. The segmentation effect of DeepLabV3+ on composite insulators before and after the improvement is shown in the figure. Figure 5 shown.
[0058] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0059] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0060] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0061] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0062] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.
Claims
1. A real-time segmentation method for composite insulators based on DeepLabV3+, characterized in that: The following steps are involved: S1. Obtain infrared images of power equipment containing composite insulators taken during power inspection and construct an original data set; S2, perform data augmentation on the original data set to obtain a training data set; S3. Build an improved DeepLabV3+ network: replace the backbone network of DeepLabV3+ with a lightweight neural network MobileNetV2 to improve real-time performance, and introduce a lightweight and efficient channel attention module ECA to achieve local cross-channel interaction without dimensionality reduction. Then, add the Pointrend fine segmentation module to the output of DeepLabV3+ for post-processing to further improve the semantic segmentation results. Train the improved DeepLabV3+ network using the training data set to obtain a trained improved DeepLabV3+ network. S4. The infrared images of the captured power equipment are processed through the trained improved DeepLabV3+ network to segment the composite insulators in real time.
2. The real-time segmentation method for composite insulators based on DeepLabV3+ according to claim 1, characterized in that: The specific implementation method of step S1 is: obtaining infrared images of power equipment taken during the power inspection process, and screening out infrared images of power equipment containing composite insulators to form an original data set; using an image annotation tool to annotate the composite insulator core rod and umbrella disk parts in the original data set image.
3. The real-time segmentation method for composite insulators based on DeepLabV3+ according to claim 1 is characterized in that: The specific implementation method of step S2 is: perform data augmentation on the original data set to increase the data volume and form a training data set; the data augmentation method is: read the pictures in the original data set, flip the pictures horizontally, flip the pictures vertically, rotate them to a set angle or randomly crop them to obtain new pictures; and the newly obtained pictures and the pictures in the original data set together form a training data set.
4. The real-time segmentation method for composite insulators based on DeepLabV3+ according to claim 1 is characterized in that: In the step S3, when constructing the improved DeepLabV3+ network, MobileNetV2 is integrated with DeepLabV3+, the MobilenetV2 network is used as the encoder, and linear is used to prevent relu from destroying features; Use depth-wise separable convolution to reduce the amount of computation and parameters and increase the number of channels to obtain more features in the image.
5. The real-time segmentation method for composite insulators based on DeepLabV3+ according to claim 4 is characterized in that: The improved DeepLabV3+ network includes an encoder and a decoder, wherein the encoder includes a dilated convolution and a spatial pyramid combined with a MobileNetV2 convolutional network, wherein the dilated convolution extracts features calculated at any resolution in MobileNetV2, and before global pooling, uses the ratio of the input image spatial resolution to the final output image spatial resolution as the output stride, and controls the density of encoder features according to the budget of computing resources; The spatial pyramid uses dilated convolutions with different dilation rates to perform multi-scale fusion of features calculated by MobileNetV2, enriching semantic information while balancing accuracy and running time; the decoder consists of two convolution structures with different convolution kernels, which are used to reduce the number of channels using the same convolutional network.
6. The real-time segmentation method for composite insulators based on DeepLabV3+ according to claim 1, characterized in that: In the step S3, the ECA module is introduced into MobileNetV2. The implementation method of ECA is as follows: firstly, global average pooling is performed on each feature, then adaptive Kernel operation is performed, each feature is obtained through the sigmoid function, and then forward propagation is performed to assign the importance weight of each feature to the original feature map.
7. The real-time segmentation method for composite insulators based on DeepLabV3+ according to claim 1, characterized in that: In the step S3, the segmentation result is post-processed by using the PointRend fine segmentation module to make the edge of the segmentation result smoother and improve the segmentation accuracy; PointRend introduces the idea of image rendering into semantic segmentation, uses bilinear interpolation to upsample the fine-grained features output by the network, selects several most uncertain points in the upsampled results as boundary segmentation difficulty points, and uses a multi-layer perceptron to perform label prediction in combination with the difficulty points directly obtained from the fine-grained features; repeats the above upsampling-point selection-label prediction process until the required resolution is reached; and by implementing adaptive subdivision in the upsampling process to retain contour information as much as possible, retains boundary points in the segmentation, and improves the segmentation accuracy.
8. A real-time segmentation system for composite insulators based on DeepLabV3+, characterized in that: The method comprises a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps as described in any one of claims 1 to 7 can be implemented.