Novel power system path planning method based on improved DeepLabv3 +

Abstract

The invention is suitable for the technical field of power systems, and provides a novel power system path planning method based on improved DeepLabv3 +. According to the method, a DeepLabv3 + segmentation method is improved, and a remote sensing data segmentation framework based on a deep learning network is constructed. According to the framework, a power system path planning problem is converted into an image segmentation problem, and an area suitable for power system line laying is accurately divided by performing fine segmentation on a satellite remote sensing image. In the process, an attention mechanism is introduced to accurately capture a power system path, and a feature reconstruction module is designed to optimize network output. In addition, a multi-scale loss function is constructed to improve segmentation accuracy. Experimental results show that compared with a traditional method, the method has the advantages that evaluation indexes are remarkably improved, segmentation accuracy is effectively improved, continuity of segmentation results is guaranteed, and more reliable and efficient technical support is provided for path planning of a power system.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and in particular relates to a novel power system path planning method based on improved DeepLabv3+. Background Art

[0002] With the rapid development of society, electric power resources have become an indispensable element in human life and social operation, and are widely used in residential electricity, industry, transportation, communications and other fields. As a clean energy, electric power resources can effectively reduce pollution, further promote the sustainable development of global resources and the realization of my country's "dual carbon" goals. At present, the global energy structure is transforming towards diversification and sustainability. Coupled with the rapid development of smart grid technology, the complexity and scale of the power system are growing at an unprecedented rate. In this context, how to scientifically and rationally plan the path of the power system to ensure its reliability and further improve it has become a key issue that needs to be solved urgently.

[0003] In the traditional planning and construction of power systems, first of all, it is necessary to survey the target terrain area, obtain the topographic data of the area, and clarify the natural distribution characteristics such as terrain undulations and rivers and lakes. Then, the environmental influencing factors such as plants and land in the area are evaluated to reduce the negative impact on the surrounding ecological environment. On this basis, preliminary line design is carried out manually based on the acquired terrain information. Finally, the feasibility of different construction sites is evaluated through field investigations and combined with economic factors. However, this traditional method is highly dependent on the experience and judgment of experts, resulting in limited planning accuracy and efficiency, while increasing manpower and time costs. In response to the above problems, the present invention proposes a new power system path planning method based on improved DeepLabv3+. Summary of the invention

[0004] The purpose of the present invention is to provide a new power system path planning method based on improved DeepLabv3+, aiming to solve the problems raised in the above background technology.

[0005] The purpose of the present invention is achieved through the following technical solutions: A new power system path planning method based on improved DeepLabv3+ includes the following steps: Step 1: Remote sensing dataset construction and preprocessing process; According to the constraint intensity of geographic information factors in the planning area, the specific tasks of remote sensing data segmentation are clarified, and the areas within the research scope are divided into buildings, roads, vegetation, etc. Diversified satellite remote sensing image data are collected and corresponding category labels (including background, opaque water surface, buildings, low vegetation, trees and vehicles) are obtained to ensure that the data has wide applicability and representativeness. Then the labels are digitized and the categorical variables are converted into numerical types in order to generate high-quality training and test sets. The labeled data are preprocessed, including image cropping, normalization and other operations to ensure the consistency and stability of the data. And through image enhancement techniques such as rotation, flipping and scaling, the diversity of samples is increased. 85% of the data set is randomly selected as the training set, 10% as the test set, and 5% as the validation set by random sampling.

[0006] Step 2: Deep learning model and segmentation algorithm optimization; The DeepLabv3+ segmentation method is improved, and a remote sensing data segmentation framework based on a deep learning network is established to improve the performance of the model in complex scenes. The spatial attention module is introduced to focus on more important information of the image, which improves the network's perception of important information. The feature information is reconstructed through asymmetric convolution and depthwise separable convolution, and the multi-scale loss function is used to achieve high-precision segmentation of satellite remote sensing images. During the training process, the training set is input into the network for training, and the multi-scale loss function is used to constrain the network; the model parameters are also adjusted and overfitting is prevented through the validation set.

[0007] Step 3: Evaluation and testing of power system path planning model; The performance of the model is evaluated on the test set using pixel accuracy, average pixel accuracy, average intersection-over-unit ratio and frequency-weighted intersection-over-unit ratio as metrics. The proposed method is compared with SegNet, UNet, DANet and DeepLabv3+ to verify its superiority. In addition, ablation experiments are performed to evaluate the impact of different modules on the model performance and ensure the reliability and effectiveness of the model.

[0008] Furthermore, the category labels include background, opaque water surface, buildings, low vegetation, trees and vehicles.

[0009] Furthermore, the network structure of the remote sensing data segmentation framework based on the deep learning network includes an encoder and a decoder, wherein the encoder is used to receive the input of the network, compress the data through transformation, and extract the feature information contained in the input data, and the decoder is used to reconstruct the output of the network and restore the detail information in the original image by reconstructing the compressed data; In the encoder, a MobileNetV2 network is first used to extract preliminary information from the input image. Then, it is input into the decoder through the deep feature branch and the shallow feature branch respectively. In the shallow feature branch, a spatial attention module is used to focus on more important information of the image. In the deep feature branch, a dilated spatial pyramid pooling module is used to capture the contextual information of the image. Then, a 1*1 convolution is used to further fuse the captured information and output the deep image of the network. In the decoder, an upsampling module is used to increase the scale of deep features to the scale of shallow features; then the shallow features are concatenated with the deep features, and a 3*3 convolution is used to fuse the shallow features with the deep features; finally, the extracted features are fully fused through the feature reconstruction module, and the output of the network is increased to the size of the original image, that is, the output of the network is obtained.

[0010] Furthermore, in the encoder: In the MobileNetV2 network, the input features first pass through a 1*1 convolution, and the Relu6 activation function is used to limit the output value between [0, 6]; then the data passes through a 3*3 depth-separable convolution, and the Relu6 activation function is used to limit the output value between [0, 6]; finally, the channel is compressed through a 1*1 convolution, and the extracted features are integrated through a fully connected layer to generate the output of the MobileNetV2 network; The Relu6 activation function is expressed as: ; in, x Represents the input of the Relu6 activation function; In the spatial attention module, the input features are firstly subjected to average pooling and maximum pooling operations to obtain two compressed feature maps; then the two compressed feature maps are channel-joined, and then the channel information is fused through a 7*7 convolution, and the weight value is obtained through the Sigmoid activation function; finally, the output features are multiplied with the input features; the operation process is expressed as: ; in, represents the output of the spatial attention module; Represents the input of the spatial attention module; Represents a 7*7 convolution operation; and Represent the average pooling operation and the maximum pooling operation respectively; Represents the Sigmoid activation function; The atrous spatial pyramid pooling module comprises the following parallel components: a 1*1 convolution, three atrous convolutions with different dilation rates and a pooling module, wherein the dilation rates of the three atrous convolutions are 6, 12 and 18 respectively; the three atrous convolutions use different dilation rates to extract feature information of different scales, obtain context information of different scales, and enable the network to segment target bodies of different sizes; the pooling module performs global average pooling on the input features, and restores the pooled features to the size of the original features through an interpolation method to obtain global feature information.

[0011] Furthermore, in the feature reconstruction module of the decoder, an asymmetric convolution strategy is first adopted to decompose the 3*3 convolution into 1*3 convolution and 3*1 convolution to keep the receptive field unchanged and reduce the amount of calculation. 3*3 convolution, 1*3 convolution and 3*1 convolution are used in combination to enrich the level of feature extraction; then a depthwise separable convolution module is used to further improve the effect of feature reconstruction. The depthwise separable convolution module includes two parts: depthwise convolution and pointwise convolution. The depthwise convolution is used to perform convolution operations on each channel, and there is no interaction between channels. The pointwise convolution is used to interact with the information between channels to generate a new feature map; finally, the upsampling module is used to restore the feature map size to the same size as the input image to obtain the final output feature of the network.

[0012] Furthermore, the loss function adopts the cross entropy loss function, which is expressed as: ; in, Loss represents the cross entropy loss function; N 1 and N 2 represents the length of the image at two scales; M 1 and M 2 represents the width of the image at two scales; Indicates i Sample No. c The true labels of the categories; Indicates i Sample No. c The model predicts the probability of each category; and Represent the weights of the two loss functions respectively.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The present invention improves the DeepLabv3+ segmentation method and constructs a remote sensing data segmentation framework based on a deep learning network. The framework cleverly transforms the power system path planning problem into an image segmentation problem. By finely segmenting satellite remote sensing images, the area suitable for laying power system lines is accurately divided, laying a solid foundation for the subsequent path planning of the power system. In this process, the present invention introduces an attention mechanism, which enables the model to focus on the key information area in the image, thereby accurately capturing the power system path. At the same time, a feature reconstruction module is designed to optimize the network output to ensure that the edges of the segmentation results are clearer and continuous. In addition, a new multi-scale loss function is constructed to further optimize the features of different scales and achieve high-precision segmentation of satellite remote sensing images. Experimental results show that compared with traditional methods, the present invention has achieved significant improvements in evaluation indicators, effectively improved the accuracy of segmentation, and ensured the continuity of segmentation results, providing more reliable and efficient technical support for power system path planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the present invention.

[0015] Figure 2 Part of the data and corresponding labels of the Vaihingen dataset.

[0016] Figure 3 The network structure of the remote sensing data segmentation framework based on deep learning network.

[0017] Figure 4 This is the structural diagram of the MobileNetV2 network.

[0018] Figure 5 Structural diagram of the spatial attention module.

[0019] Figure 6 This is the structural diagram of the feature reconstruction module.

[0020] Figure 7 Visualize the segmentation results of different methods on the Vaihingen dataset. DETAILED DESCRIPTION

[0021] In order to have a clearer understanding of the technical features, purposes and beneficial effects of the present invention, the technical solution of the present invention is now described in detail below, but it should not be construed as limiting the applicable scope of the present invention.

[0022] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0023] like Figure 1As shown, it is a flow chart of a novel power system path planning method based on improved DeepLabv3+ provided by an embodiment of the present invention, and the method specifically includes the following steps: 1. Remote sensing dataset construction and preprocessing process; 1) Dataset acquisition and preliminary analysis; The experimental data comes from the Vaihingen urban dataset, which contains 33 high-resolution remote sensing images covering multiple areas, with rich ground object types and complex scene structures. The category labels provided by the dataset include background, opaque water surface, buildings, low vegetation, trees and vehicles, which provide the basis for subsequent image classification and recognition tasks.

[0024] 2) Label digital processing and visualization; In order to facilitate model learning and classification, the category labels need to be converted into corresponding digital codes. Table 1 shows the correspondence between category labels and digital codes.

[0025] Table 1 Tags and their corresponding codes

[0026] In addition, in order to intuitively understand the content and structure of the dataset, the visualization results of some images and their corresponding labels are drawn. Figure 2 As shown, each image is annotated with the corresponding category label, which helps to better understand the distribution and characteristics of the dataset.

[0027] 3) Image preprocessing; Image cropping and data set division: Since the original image size is too large, using it directly for model training will result in excessive consumption of computing resources and low training efficiency. Therefore, each original image is cropped into multiple 512*512 image blocks. This step not only reduces the image size, but also increases the diversity of the data set, which helps to improve the generalization ability of the model. After cropping, a total of 3300 image blocks are obtained. Next, in order to reasonably divide the training set, test set, and validation set, a random sampling method is adopted. Specifically, 85% (i.e., 2805 image blocks) of the data set are randomly selected as the training set for model training and learning; 10% (i.e., 330 image blocks) are used as the test set to evaluate the final performance of the model; the remaining 5% (i.e., 165 image blocks) are used as the validation set to adjust model parameters and prevent overfitting during training.

[0028] Normalization: Before the image is input into the network for training, it needs to be normalized. The purpose of normalization is to adjust the pixel value of the image to a suitable range (such as between 0 and 1) to eliminate the brightness difference and noise interference between different images, thereby improving the training efficiency and accuracy of the model. In addition to normalization, other preprocessing operations can be performed according to specific needs, such as image enhancement (such as rotation, scaling, flipping, etc.) to increase the diversity of the data.

[0029] 2. Deep learning model and segmentation algorithm optimization; 1) Network structure; The network structure of the remote sensing data segmentation framework based on deep learning network is as follows Figure 3 As shown in the figure, the network structure consists of two parts: the encoder and the decoder. The encoder is responsible for receiving the input of the network and compressing the data through a series of transformations to extract the rich feature information contained in the input data. The decoder is responsible for reconstructing the output of the network and restoring the detailed information in the original image by reconstructing the compressed data.

[0030] In the encoder, the input image is firstly extracted through a MobileNetV2 network. The MobileNetV2 network uses deep separable convolution to effectively reduce the number of parameters and calculations of the model. Next, it is input into the decoder through the deep feature branch and the shallow feature branch respectively. These two branches focus on the extraction of deep features and shallow features respectively. In the shallow feature branch, a spatial attention module is used to focus on more important information of the image, which helps to restore image details and texture information. In the deep feature branch, a hollow spatial pyramid pooling module (ASSP) is used to capture the contextual information of the image. ASSP uses hollow convolution to increase the receptive field of the network, which helps the network understand the overall structure of the image. Then the captured information is further fused using 1*1 convolution, and the deep image of the network is output.

[0031] In the decoder, the scale of the deep features is first increased to the scale of the shallow features through an upsampling module to ensure the unity of the two in the spatial dimension. Then the shallow features are concatenated with the deep features, and the two features are fused through a 3*3 convolution block, which can effectively integrate the shallow features' ability to focus on detail information and the deep features' ability to understand the overall structure. Finally, the extracted features are fully fused through the feature reconstruction module, and the network output is increased to the size of the original image to ensure that its resolution is consistent with the original image, and finally the network output is obtained.

[0032] 2) MobileNetV2 network; The structure of the MobileNetV2 network is as follows Figure 4As shown in the figure, the left side of the figure is a structure with a step size of 1, and the right side of the figure is a structure with a step size of 2. In this network, a 1*1 convolution is first used to increase the number of channels, thereby improving the feature extraction capability of the network. At the same time, the Relu6 activation function is used to limit the output value to between [0, 6] to prevent the gradient from disappearing or exploding, and to ensure the stability of the network during back propagation. Then, a 3*3 depthwise separable convolution is used to reduce the amount of computation while maintaining the extraction capability of the network. Similarly, this layer also uses the Relu6 activation function to limit the range of output values. Finally, a 1*1 convolution is used to compress the channel, and the extracted features are integrated through a fully connected layer. Among them, the Relu6 activation function is expressed as: ; in, x Represents the input of the Relu6 activation function; 3) Spatial attention module; The spatial attention module dynamically adjusts the weights of different spatial locations to enhance important information and suppress irrelevant information. Figure 5 The structure of the spatial attention module is shown. The input features are firstly subjected to average pooling and maximum pooling operations to obtain two compressed feature maps, and then the two compressed feature maps are channel-joined. Next, the channel information is fused through a 7*7 convolution operation, and the weight value is obtained through the Sigmoid activation function. Finally, the obtained weight value is multiplied with the input feature to obtain the output feature. The operation process can be expressed as: ; in, represents the output of the spatial attention module; Represents the input of the spatial attention module; Represents a 7*7 convolution operation; and Represent the average pooling operation and the maximum pooling operation respectively; Represents the Sigmoid activation function; 4) Atrous spatial pyramid pooling module; The dilated spatial pyramid pooling module contains the following parallel components: 1*1 convolution, three dilated convolutions with different dilation rates, and a pooling module. The dilation rates of the three dilated convolutions are 6, 12, and 18 respectively. The 1*1 convolution can ensure that the network still pays attention to local features while expanding the receptive field, ensuring the network's ability to pay attention to detail information. By using dilated convolutions with different dilation rates to extract feature information of different scales, and then obtaining contextual information of different scales, the network can segment objects of different sizes. By performing global average pooling on the input features through the pooling module and then restoring them to the size of the original features through interpolation, global feature information can be obtained, the network's receptive field is increased, and the segmentation ability is further enhanced.

[0033] 5) Feature reconstruction module; The feature reconstruction module is located at the end of the network and is used to reconstruct feature information and obtain the output of the network. The structure diagram of the feature reconstruction module is as follows: Figure 6 As shown. First, an asymmetric convolution strategy is used to decompose the 3*3 convolution into 1*3 convolution and 3*1 convolution, which can ensure that the receptive field remains unchanged while reducing the amount of calculation and speeding up the inference speed. By combining 3*3 convolution, 1*3 convolution and 3*1 convolution, the level of feature extraction is enriched to capture more comprehensive image features. Then, a depthwise separable convolution module is used to further improve the effect of feature reconstruction. The depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution. Through depthwise convolution, each channel is convolved, and there is no interaction between channels. Through pointwise convolution, the information between channels is interacted to generate a new feature map. The depthwise separable convolution can further reduce the amount of calculation and ensure accuracy. Finally, the upsampling module is used to restore the feature map size to the same size as the input image to obtain the final output feature of the network.

[0034] 6) Loss function; The loss function adopts the cross entropy loss function, which is expressed as: ; in, Loss represents the cross entropy loss function; N 1 and N 2 represents the length of the image at two scales; M 1 and M 2 represents the width of the image at two scales; Indicates i Sample No. c The true labels of the categories; Indicates i Sample No. cThe model predicts the probability of each category; and Respectively represent the weights of the two loss functions. In the present invention, , .

[0035] 3. Evaluation and testing of power system path planning models; In order to evaluate the segmentation effect of each method, this paper uses pixel accuracy (PA), mean pixel accuracy (MPA), mean intersection over union (MIoU) and frequency weight intersection over union (FWIoU) as segmentation metrics. PA represents the proportion of correctly labeled pixels to total pixels, expressed as: ; MPA represents the proportion of correctly classified pixels in each class, and averages them, expressed as: ; FWIoU means setting the weight according to the frequency of each category and multiplying it by the IoU of each category, and summing them up, expressed as: ; in, k represents the number of segmentation categories, Indicates that the category i Predict into categories j The number of pixels, Indicates the number of correctly predicted pixels.

[0036] 1) Model comparison; In order to verify the effectiveness of the proposed method, the proposed method is quantitatively compared with SegNet, UNet, DANet and DeepLabv3+, that is, the performance of different methods is evaluated by different evaluation indicators. Table 2 shows the comparison of the results of different methods on the Vaihingen dataset: Table 2 Comparison of results of different methods on the Vaihingen dataset

[0037] The data results in Table 2 show that the method of the present invention is superior to other methods in terms of pixel accuracy (PA), mean pixel accuracy (MPA), mean intersection over union (MioU) and frequency weighted intersection over union (FWIoU), and achieves better evaluation indicators. Specifically: The PA of the method of the present invention is 94.18%, which is 4.31% and 0.46% higher than SegNet and DeepLabv3+, respectively. MPA is 88.95%, which is 7.98% and 2.2% higher than SegNet and DeepLabv3+, respectively. MioU is 82.09%, which is 17.27% and 2.71% higher than SegNet and DeepLabv3+, respectively. FWIoU is 89.49%, which is 7.17% and 1.01% higher than SegNet and DeepLabv3+, respectively. In summary, the method of the present invention has achieved the best results in different evaluation indicators, which fully proves its effectiveness and superiority.

[0038] In the specific evaluation of each category, we compared the MioU results of different methods for trees, buildings, and low vegetation. The following is a detailed comparison: In the tree category, the MioU of the method of the present invention is 84.46%, which is 11.56% and 1.33% higher than SegNet and DeepLabv3+, respectively. In the building category, the MioU of the method of the present invention is 92.74%, which is 4.33% and 1.03% higher than SegNet and DeepLabv3+, respectively. In the low vegetation category, the MioU of the method of the present invention is 79.50%, which is 16.77% and 2.68% higher than SegNet and DeepLabv3+, respectively. Based on the above data, it can be seen that the evaluation indicators of the method of the present invention in each category have achieved the best results. This not only proves the effectiveness of the method of the present invention again, but also fully demonstrates that it has higher segmentation performance in all categories.

[0039] In order to compare the effects of each method more intuitively, Figure 7The segmentation visualization results of different methods on the Vaihingen dataset are shown for qualitative analysis. The first column in the figure is the original image input to the network, the second column is the prediction result of the SegNet method, the third column is the prediction result of the UNet method, the fourth column is the prediction result of the DANet method, the fifth column is the prediction result of the DeepLabv3+ method, the sixth column is the prediction result of the method of the present invention, and the seventh column is the label image. By observing these visualization results, it can be found that: in the first row of images, the edge contour of the building displayed by the method of the present invention is clearer and closer to the real shape, while other methods have different degrees of information loss. For the gaps between adjacent trees, only the method of the present invention can accurately identify them. In the second row of images, the SegNet and DeepLabv3+ methods mistakenly identify the background as a vehicle. The UNet and DANet methods have segmentation errors and incoherence problems, respectively. In the third row of images, both the SegNet and UNet methods have problems of incoherence and information loss, the DANet method mistakenly identifies the background as a vehicle, and the DeepLabv3+ method cannot accurately identify the tiny background. In the fourth row of images, the other four methods all have segmentation errors, and the SegNet, UNet, and DANet methods cannot recognize tiny information. In contrast, the method of the present invention not only accurately recognizes tiny information, but also its segmentation results are more accurate, clear, and coherent, closer to the true shape, and even performs well in recognizing tiny gaps between adjacent information. In summary, the method of the present invention shows obvious advantages in visualization results, and its segmentation performance is more superior, and it can more accurately capture and restore the true shape.

[0040] 2) Ablation experiments; In order to verify the effectiveness of each module in the method of the present invention, we conducted an ablation experiment on the Vaihingen dataset. The experiment used the DeepLabv3+ method as the basic model and compared the results after adding each module. The experiment used PA, MPA, MioU and FWIoU as evaluation indicators. The comparison results are shown in Table 3: Table 3 Comparison of ablation experiment results on the Vaihingen dataset

[0041] According to the data results in Table 3, the following conclusions can be drawn: From the comparison between Experiment 1 and Experiment 2, it can be seen that after adding the attention module, the performance of the network has been improved, and each evaluation index has shown better results. From the comparison between Experiment 1 and Experiment 3, it can be seen that after adding the feature reconstruction module, PA, MPA, MioU and FWIoU have all been significantly improved, which shows the effectiveness of the module. From the comparison between Experiment 4 and Experiment 5, it can be seen that after using the new loss function, each evaluation index has been improved, and the accuracy of segmentation has been further improved, indicating that the segmentation effect of the method of the present invention is stronger than that of the DeepLabv3+ method. In summary, the results of the ablation experiment fully verify the effectiveness of each module in the method of the present invention. Under the joint action of each module, the segmentation performance of the network is significantly improved.

[0042] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A new power system path planning method based on improved DeepLabv3+, characterized in that: The following steps are involved: Step 1: Remote sensing dataset construction and preprocessing process; First, satellite remote sensing image data is collected and the corresponding category labels are obtained; then the labels are digitized and converted into corresponding digital codes; then the labeled image data is preprocessed, including image cropping, normalization and image enhancement; Step 2: Deep learning model and segmentation algorithm optimization; Improve the DeepLabv3+ segmentation method and establish a remote sensing data segmentation framework based on deep learning networks; The spatial attention module is introduced to improve the network's ability to perceive important information, feature information is reconstructed through asymmetric convolution and depthwise separable convolution, and multi-scale loss function is used to achieve accurate segmentation of satellite remote sensing images. Step 3: Evaluation and testing of power system path planning model; Pixel accuracy, average pixel accuracy, average intersection-over-unit ratio and frequency-weighted intersection-over-unit ratio are used as segmentation metrics to evaluate the performance of the model on the test set. By implementing ablation experiments, the impact of different modules on the model performance is evaluated to ensure the reliability and effectiveness of the model.

2. The novel power system path planning method based on improved DeepLabv3+ according to claim 1 is characterized in that: The class labels include background, opaque water surfaces, buildings, low vegetation, trees, and vehicles.

3. The novel power system path planning method based on improved DeepLabv3+ according to claim 1 is characterized in that: The network structure of the remote sensing data segmentation framework based on the deep learning network includes an encoder and a decoder. The encoder is used to receive the input of the network, compress the data through transformation, and extract the feature information contained in the input data. The decoder is used to reconstruct the output of the network and restore the detail information in the original image by reconstructing the compressed data. In the encoder, a MobileNetV2 network is first used to extract preliminary information from the input image. Then, it is input into the decoder through the deep feature branch and the shallow feature branch respectively. In the shallow feature branch, a spatial attention module is used to focus on more important information of the image. In the deep feature branch, a dilated spatial pyramid pooling module is used to capture the contextual information of the image. Then, a 1*1 convolution is used to further fuse the captured information and output the deep image of the network. In the decoder, an upsampling module is used to increase the scale of deep features to the scale of shallow features; then the shallow features are concatenated with the deep features, and a 3*3 convolution is used to fuse the shallow features with the deep features; finally, the extracted features are fully fused through the feature reconstruction module, and the output of the network is increased to the size of the original image, that is, the output of the network is obtained.

4. The novel power system path planning method based on improved DeepLabv3+ according to claim 3 is characterized in that: In the encoder: In the MobileNetV2 network, the input features first pass through a 1*1 convolution, and the Relu6 activation function is used to limit the output value between [0, 6]; then the data passes through a 3*3 depth-separable convolution, and the Relu6 activation function is used to limit the output value between [0, 6]; finally, the channel is compressed through a 1*1 convolution, and the extracted features are integrated through a fully connected layer to generate the output of the MobileNetV2 network; The Relu6 activation function is expressed as: ； in, x Represents the input of the Relu6 activation function; In the spatial attention module, the input features are firstly subjected to average pooling and maximum pooling operations to obtain two compressed feature maps; then the two compressed feature maps are channel-joined, and then the channel information is fused through a 7*7 convolution, and the weight value is obtained through the Sigmoid activation function; finally, the output features are multiplied with the input features; the operation process is expressed as: ； in, represents the output of the spatial attention module; Represents the input of the spatial attention module; Represents a 7*7 convolution operation; and Represent the average pooling operation and the maximum pooling operation respectively; Represents the Sigmoid activation function; The atrous spatial pyramid pooling module comprises the following parallel components: a 1*1 convolution, three atrous convolutions with different dilation rates and a pooling module, wherein the dilation rates of the three atrous convolutions are 6, 12 and 18 respectively; the three atrous convolutions use different dilation rates to extract feature information of different scales, obtain context information of different scales, and enable the network to segment target bodies of different sizes; the pooling module performs global average pooling on the input features, and restores the pooled features to the size of the original features through an interpolation method to obtain global feature information.

5. The novel power system path planning method based on improved DeepLabv3+ according to claim 3 is characterized in that: In the feature reconstruction module of the decoder, an asymmetric convolution strategy is firstly adopted to decompose the 3*3 convolution into 1*3 convolution and 3*1 convolution to keep the receptive field unchanged and reduce the amount of calculation. 3*3 convolution, 1*3 convolution and 3*1 convolution are used in combination to enrich the level of feature extraction; then a depthwise separable convolution module is used to further improve the effect of feature reconstruction. The depthwise separable convolution module includes two parts: depthwise convolution and pointwise convolution. The depthwise convolution is used to perform convolution operations on each channel without interaction between channels. The pointwise convolution is used to interact the information between channels to generate a new feature map; finally, the upsampling module is used to restore the feature map size to the same size as the input image to obtain the final output feature of the network.

6. The novel power system path planning method based on improved DeepLabv3+ according to claim 1 is characterized in that: The loss function adopts the cross entropy loss function, which is expressed as: ； in, Loss represents the cross entropy loss function; N 1 and N 2 represents the length of the image at two scales; M 1 and M 2 represents the width of the image at two scales; Indicates i Sample No. c The true labels of the categories; Indicates i Sample No. c The model predicts the probability of each category; and Represent the weights of the two loss functions respectively.

Images