A method and system for identifying ground objects in transmission line engineering survey tasks

Abstract

The present invention belongs to the field of image processing technology, and specifically relates to a method and system for object recognition for power transmission line engineering survey tasks. The method uses high-resolution oblique aerial images as a data source, and utilizes an integrated large-model algorithm annotation tool to generate a basic sample data set of typical object elements such as houses, roads, and trees. At the same time, a basic instance segmentation network (HRFI-YOLO network) suitable for object recognition in high-resolution oblique aerial images is designed, and an aerial image object recognition model is generated through training, which can extract objects such as houses, roads, and trees in remote sensing images. The present invention reduces the workload while ensuring the availability of label data and the stability of the deep learning model. By constructing a deep learning network model, the automatic extraction of objects is realized, the influence of human subjective factors is reduced, it has strong versatility, and the generalization ability of object extraction is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for identifying ground objects used in transmission line engineering survey tasks. Background Art

[0002] Deep learning technology has made significant progress in computer vision and intelligent information processing. This deep learning technology, based on image analysis and understanding, is gradually empowering artificial intelligence in multiple fields. As a new data acquisition platform, drone systems, with their flexible mobility and ease of use, can quickly and efficiently acquire image data with significant application value. In particular, drones equipped with imaging equipment can quickly capture detailed aerial photography of target areas, acquiring high-resolution images rich in spatial information.

[0003] CN202210681276.5 discloses a satellite image data processing system based on deep learning, which only sets up training classifiers and data enhancement, and does not make any improvements to the neural network. The accuracy of image processing needs to be improved. For example, CN202011475006.6 discloses a training method and device for a remote sensing image fusion model, which is mainly used for fusion processing of data images, and does not involve improvements to the neural network model. The image recognition accuracy is uncertain. Finally, CN115861824B discloses a remote sensing image recognition method based on an improved Transformer, which adopts an improved neural network ITFormer with multi-size convolution-Gaussian weighted feature extraction, and combines it with the grid method to improve the training speed and classification speed, and can improve the rate and accuracy of image recognition. However, it has certain defects, such as its high optimization difficulty, the need for large-scale computing resources when processing large-scale multi-scale data sets, and for some tasks, it requires a large amount of labeled data to achieve its best performance.

[0004] Therefore, based on existing technologies, this application provides an intelligent recognition algorithm based on drone images. This algorithm not only automatically analyzes image content and improves information extraction efficiency, but also enables intelligent perception of ground objects and landforms. This has important theoretical significance and broad application prospects for drones to better serve important fields such as power system surveys. Summary of the Invention

[0005] Purpose of the invention: In order to solve the problems existing in the above-mentioned prior art, the present invention constructs a ground object identification method for transmission line engineering survey tasks. In addition, the present invention also proposes a ground object identification system for transmission line engineering survey tasks.

[0006] Technical solution: One aspect of the present invention provides a method for identifying ground objects for power transmission line engineering survey tasks, the method comprising the following steps:

[0007] S1 uses ISAT semi-automatic annotation software to produce a high-resolution oblique aerial photography dataset;

[0008] S2 designs a deep learning-based ground object recognition HRFI-YOLO network;

[0009] The deep learning-based ground object recognition HRFI-YOLO network includes:

[0010] Backbone network: The improved Darknet-53 network is used as the backbone network, and the C2f module and CBAM module are introduced into the backbone network;

[0011] The improved Darknet-53 network is used as the backbone network, specifically including:

[0012] The first CBS module, the second CBS module, the third C2f_SE module, the fourth CBS module, the fifth C2f_SE module, the sixth CBS module, the seventh C2f_SE module, the eighth CBS module, the ninth C2f_SE module and the CBAM_SPPF module are sequentially connected in series, wherein the CBS module includes 1 convolution layer, 1 batch normalization layer and 1 SiLU activation function;

[0013] Neck network: The improved FPN-PAN network structure is used as the basic neck network structure, and the GFPN network structure idea and CSP module are integrated to construct a multi-layer feature fusion neck network;

[0014] Detection and segmentation head: includes four detection heads of different scales, including a new detection head for small ground objects;

[0015] S3 performs data augmentation on the high-resolution oblique aerial photography dataset and trains the HRFI-YOLO network on the dataset;

[0016] S4 uses the trained object recognition network model to extract the object target;

[0017] S5 interfaces the extracted results in the format required by the design platform.

[0018] Further, including:

[0019] The step S1 specifically includes:

[0020] S11 creates a virtual environment, obtains the ISAT installation package source code and installs dependencies;

[0021] S12 downloads the pre-trained model and stores it in the weight folder directory;

[0022] S13 performs semi-automatic annotation and adjusts the target segmentation mask by prompting the region of interest or the region of no interest;

[0023] S14 ISAT uses the JSON file format to save the annotation results. The JSON file is manually parsed and converted into TXT format. The training, validation, and test data sets are split by configuring the ratio of the validation set and the test set.

[0024] Further, including:

[0025] The C2f_SE module is specifically:

[0026] The SE mechanism is introduced into the existing C2f module. After the input feature map group is processed by the C2f module, the attention weight is weighted to the features of each channel through global information embedding and adaptive correction to output the feature map group. Among them, the C2f module contains 2 convolutional layers, 3 bottleneck layers and 1 merge connection layer.

[0027] Further, including:

[0028] In the detection and segmentation head, a new detection head for small ground objects is added, so that the network has four detection heads of different scales, including:

[0029] An additional 160×160 detection head is introduced into the head network part and aligned with the multi-layer feature fusion neck network of the neck to improve the detection ability of targets of various scales.

[0030] Further, including:

[0031] The CBAM_SPPF module introduces the CBAM attention mechanism into the backbone network before the feature map group is input into the SPPF module. The SPPF module contains multiple small-size pooling layers and a merge connection layer. The CBAM contains two key modules: the channel attention module and the spatial attention module. These two modules generate attention maps in the channel and spatial directions respectively. These maps are then multiplied with the input feature map and input into the CBAM after passing through a CBS module.

[0032] Further, including:

[0033] The neck network includes: the feature map output from the CBAM_SPPF module is input into the first connection layer after passing through the third CBS module, the first connection layer also receives the feature map output from the fifth CBS module, and after being processed by the first connection layer, is input into the first CSP module;

[0034] The data of the fifth CBS module comes from the seventh C2f_SE module, and the feature map output from the seventh C2f_SE module is also output to the second connection layer. A first upsampling is further provided between the first CSP module and the second connection layer. The second connection layer receives three-party data: the first upsampling output data, the seventh C2f_SE module output data, and the feature map of the fifth C2f_SE module output data after passing through the seventh CBS module;

[0035] The second connection layer processes the data and outputs it to the second CSP module, which then undergoes a second upsampling and outputs it to the third connection layer. The third connection layer also receives data from the fifth C2f_SE module and simultaneously outputs the data to the third CSP module and the third upsampling module. The third upsampling module transmits the data to the fourth connection layer. The fourth connection layer also receives the feature map from the third C2f_SE module and transmits the output data to the fourth CSP module.

[0036] After the fourth CSP module transfers the data to the ninth CBS module, it enters the fifth connection layer. The fifth connection layer also receives data from the third CSP module. After processing by the connection layer, the data enters the fifth CSP module, and then enters the sixth connection layer after passing through the tenth CBS module. At the same time, the sixth connection layer also receives data from the second CSP module. After connection, it enters the fifth CSP module and the eleventh CBS module. Together with the data from the second CSP module after passing through the twelfth CBS module, it enters the seventh connection layer and finally enters the segmentation detection head through the sixth CSP module.

[0037] Further, including:

[0038] The detection segmentation head specifically includes: a first detection head, which receives a feature map output from the third CSP module, corresponding to a size of 20×20; a second detection head, which receives a feature map output from the fourth CSP module, corresponding to a size of 40×40; a third detection head, which receives a feature map output from the fifth CSP module, corresponding to a size of 80×80; and a fourth detection head, which receives a feature map output from the sixth CSP module.

[0039] Further, including:

[0040] In step S3, data enhancement of the high-resolution oblique aerial photography dataset specifically includes:

[0041] The visibility of ground objects is enhanced by adjusting the contrast and brightness of the original aerial image. This is achieved through a weighted algorithm. The algorithm formula is as follows:

[0042] Output=Input1×α+Input2×β+γ

[0043] Among them, Input1 and Input2 both represent input images, Output represents output image, α, β and γ are parameters, α is set to 1.1, β is set to 0.1, and γ is set to 30.

[0044] On the other hand, the present invention also provides a ground feature identification system for transmission line engineering survey tasks, the system comprising:

[0045] Data acquisition module, used to produce high-resolution oblique aerial photography datasets using ISAT semi-automatic annotation software;

[0046] Network building module for designing the HRFI-YOLO network for deep learning-based object recognition;

[0047] The deep learning-based ground object recognition HRFI-YOLO network includes:

[0048] Backbone network: The improved Darknet-53 network is used as the backbone network, and the C2f module and CBAM module are introduced into the backbone network;

[0049] The improved Darknet-53 network is used as the backbone network, specifically including:

[0050] The first CBS module, the second CBS module, the third C2f_SE module, the fourth CBS module, the fifth C2f_SE module, the sixth CBS module, the seventh C2f_SE module, the eighth CBS module, the ninth C2f_SE module and the CBAM_SPPF module are sequentially connected in series, wherein the CBS module includes 1 convolution layer, 1 batch normalization layer and 1 SiLU activation function;

[0051] Neck network: The improved FPN-PAN network structure is used as the basic neck network structure, and the GFPN network structure idea and CSP module are integrated to construct a multi-layer feature fusion neck network;

[0052] Detection and segmentation head: includes four detection heads of different scales, including a new detection head for small ground objects;

[0053] The data enhancement module is used to perform data enhancement on the high-resolution oblique aerial photography dataset and train the HRFI-YOLO network on the dataset;

[0054] The training module is used to extract ground object targets using the trained ground object recognition network model;

[0055] The corresponding module is used to interface the extraction results in the format required by the design platform.

[0056] Finally, the present invention also provides a computer-readable storage medium, on which a ground object identification program is stored. When the ground object identification program is executed by a processor, the steps of the ground object identification method for transmission line engineering survey tasks as described above are implemented.

[0057] Beneficial effects:

[0058] (1) The present invention adopts a labeling tool that integrates a large model algorithm, which significantly improves the efficiency of data labeling and solves the problem of difficult labeling in the prior art;

[0059] (2) The deep learning network model designed in the present invention introduces the SE attention mechanism and the CBAM attention mechanism in the backbone network to extract more important channel and spatial information and ignore redundant information; in the neck network part, the idea of ​​cross-stage local network is used to improve the FPN-PAN network structure, which promotes the effective feature fusion of different levels; a small ground object detection head is added to enable the network to identify and detect ground objects in higher resolution feature maps, which can realize the accurate recognition and segmentation of key ground objects such as buildings, roads, rivers, vegetation, etc. in high-resolution oblique aerial images.

[0060] (3) The present invention improves the Darknet-53 network. It not only takes into account the complex and changeable scales of ground objects and the need for the algorithm to have multi-scale processing capabilities, but also optimizes the extraction and segmentation of boundaries by adding a small ground object detection head for targets with irregular image edges.

[0061] (4) The HRFI-YOLO network designed in the present invention takes high-resolution oblique aerial images as input data, thereby directly outputting the category and segmentation contour results of the ground object to be identified. The model training time is short and the generalization ability is strong.

[0062] (5) The present invention adopts a data enhancement method to better adapt to different scenarios and conditions, and further improve the robustness and generalization ability of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0064] Figure 1 This is a flow chart of a method for identifying objects in a transmission line engineering survey task according to an embodiment of the present invention;

[0065] Figure 21 is a diagram showing the overall structure of the HRFI-YOLO network in an embodiment of the present invention;

[0066] Figure 3 This is a diagram of the network structure of the C2f_SE module in an embodiment of the present invention;

[0067] Figure 4 This is a network structure diagram of the CBAM_SPPF module in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to better understand the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0069] This method reduces workload while ensuring the availability of labeled data and the stability of deep learning models. By building a deep learning network model, it achieves automatic extraction of features, reduces the influence of human subjective factors, has strong versatility, and improves the generalization ability of feature extraction.

[0070] Research on intelligent recognition algorithms based on drone images not only automatically analyzes image content and improves information extraction efficiency, but also enables intelligent perception of terrain features and landforms. This has important theoretical implications and broad application prospects for drones to better serve important fields such as power system surveys. This invention will effectively address the key difficulties of terrain mapping in power line design, enabling automatic detection and recognition, and bringing intelligence to fields such as geotechnical engineering and power design.

[0071] like Figure 1 FIG. 1 is a flowchart of the technical solution of the present invention, which is a method for identifying objects in a transmission line engineering survey task. The method includes the following steps:

[0072] S1. Use ISAT semi-automatic annotation software to create a high-resolution oblique aerial photography dataset; ISAT is a semi-automatic segmentation and annotation software derived from segment anything.

[0073] S2. Design a deep learning-based object recognition HRFI-YOLO network;

[0074] S3, perform data augmentation on the high-resolution tilted dataset and train the HRFI-YOLO network on the dataset;

[0075] S4. Use the trained object recognition network model to extract objects such as houses, roads, and trees;

[0076] S5. Connect the extracted results in the format required by the design platform.

[0077] Wherein, step S1 includes the following steps:

[0078] S11. Create a virtual environment, obtain the ISAT installation package source code and install dependencies;

[0079] S12. Download the pre-trained model and store it in the weight folder directory;

[0080] S13. Click the toolbar to start semi-automatic annotation, use the left (or right) mouse button to highlight the region of interest (or region of no interest), and adjust the target segmentation mask;

[0081] S14. ISAT uses JSON file format to save the annotation results. Manually parse the JSON file and convert it into TXT format. Split the training, validation, and test data sets by configuring the ratio of the validation set and the test set.

[0082] like Figure 2 As shown in FIG, the overall structure diagram of the HRFI-YOLO network of the present invention is shown. The HRFI-YOLO network includes the following three parts:

[0083] The backbone network uses the improved Darknet-53 network as the backbone network, and the backbone network introduces the C2f module and CBAM module;

[0084] Neck network, adopts the improved FPN-PAN network structure as the basic neck network structure, integrates the GFPN network structure idea and the CSP module to build a multi-layer feature fusion neck network;

[0085] The CSP module includes the Darknet53 network with a CSP architecture, which serves as the backbone network for the YOLOv4 object detector. CSPNet is a cross-stage local network architecture. The addition of the CSP structure solves the problem of information duplication in the backbone network, particularly during gradient optimization of large neural networks. This significantly reduces the number of model parameters and the number of floating point operations (FLOPs), resulting in faster inference for the final model.

[0086] Detection and segmentation head: A new detection head for small ground objects is added, so that the network has 4 detection heads of different scales.

[0087] Among them, the HRFI-YOLO network uses the improved Darknet-53 network as the backbone network, and its structure can be expressed as: the first CBS module, the second CBS module, the third C2f_SE module, the fourth CBS module, the fifth C2f_SE module, the sixth CBS module, the seventh C2f_SE module, the eighth CBS module, the ninth C2f_SE module and the CBAM_SPPF module are connected in series in sequence.

[0088] The CBS module contains one convolutional layer, one batch normalization layer, and one SiLu activation function. In CBS, C stands for Conv, B stands for BatchNorm2d, and S stands for SiLu activation function. The CBS module encapsulates the combined operations of convolution, batch normalization, and activation functions.

[0089] The SiLU formula is as follows:

[0090] silu(x)=x·σ(x)

[0091] Where σ(x) represents the Sigmoid function.

[0092] like Figure 3 As shown in the figure, it is the structure diagram of the C2f_SE module of the present invention. The SE mechanism is introduced into the existing C2f module. After the input feature map group is processed by the C2f module, the attention weight is weighted to the features of each channel through global information embedding and adaptive correction to output the feature map group. Among them, the C2f module contains 2 convolutional layers, 3 Bottleneck layers and 1 merge connection layer.

[0093] The feature map group output by the C2f module is used as the input feature map group X of the second part of the C2f_SE module, through F tr Operation generates feature map U; through F sq The operation performs global average pooling on the feature map to generate a 1*1*C vector z c , so that each channel is represented by a numerical value; F ex The operation consists of two fully connected layers, and the weight W generates the required weight information, where W is obtained through learning and is used to show the feature correlation required for modeling. ex Operate on the vector z obtained in the previous step c Processing is performed to obtain the channel weight value s, where different values ​​in s represent the weight information of different channels; finally, the weight vector s generated previously is assigned to the feature map U to obtain the feature map Its size is exactly the same as the feature map U, and the SE attention mechanism does not change the size of the feature map. The specific process formula is expressed as follows:

[0094] u c =F tr (x c )

[0095]

[0096] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0097]

[0098] The c2f module, short for Convolution to Feature, combines a convolutional neural network (CNN) with a fully connected neural network (FNN). This module effectively extracts image features and has achieved promising results in various fields.

[0099] CBAM (Convolutional Block Attention Module) is a lightweight convolutional attention module that combines channel and spatial attention mechanism modules.

[0100] like Figure 4 Figure 2 shows the structure of the CBAM_SPPF module of the present invention. The CBAM attention mechanism is introduced before the feature map group is input into the SPPF module. The SPPF module includes multiple small-size pooling layers and a merge connection layer. If the feature maps connected to the CBAM are large, this can lead to a significant increase in parameters and computational cost. Furthermore, the large number of parameters can increase the difficulty of model training.

[0101] To address this issue, the present invention introduces a CBAM module into the backbone network, aiming to highlight key information when extracting backbone features, rather than introducing the CBAM module in the neck network. Furthermore, the input feature map of the CBAM module is only 20×20, a 32-fold reduction compared to the full 640×640 image. This means that the computational complexity of CBAM is relatively small, without significantly impacting the overall computational burden. This design ensures the feasibility and effectiveness of CBAM in the network while maintaining efficient computation. It helps preserve key information while maintaining computational efficiency, thereby improving model performance and generalization. CBAM comprises two key modules: the channel attention module and the spatial attention module. These modules generate attention maps in the channel and spatial directions, respectively, which are then multiplied with the input feature map to facilitate adaptive feature refinement. As a result, the CBAM method highlights features that are significant in the channel and spatial directions, while suppressing the influence of redundant features.

[0102] Input features F∈R C*H*W , channel attention module M c ∈R C*1*1 , multiply the convolution result by the original image, take the channel attention output as input, and perform the two-dimensional convolution M of the spatial attention module s ∈R 1*H*W , and then multiply the output result by the original image. The specific process formula is expressed as follows:

[0103]

[0104] The channel attention module performs global maximum pooling and average pooling operations on feature maps across different channels, followed by element-wise summation and Sigmoid activation. This process helps capture important feature information in each channel and is normalized using the Sigmoid activation function to generate a channel attention map.

[0105] On the other hand, the spatial attention module performs global max pooling and average pooling on the pixel values ​​at the same position on different feature maps, then concatenates the two feature maps together, followed by a Conv2d operation and a Sigmoid activation. Through this process, the spatial attention module can capture the important correlations between pixels at the same position on different feature maps and generate a spatial attention map.

[0106] The HRFI-YOLO network's neck network uses an improved FPN-PAN network structure. The FPN leverages the feature representation structure of different dimensions of images at the same scale, from bottom to top, within a conventional CNN model. This method effectively generates multi-dimensional feature representations for a single image view. This effectively empowers the conventional CNN model, generating more expressive feature maps for subsequent computer vision tasks. Essentially, it strengthens the feature representation of the backbone CNN network.

[0107] PAN, or PANet, adds a downsampling fusion feature pyramid to the upsampling fusion feature pyramid of FPN. The original PAN uses a shortcut operation for downsampling fusion, rather than the concat operation used in YOLOV4. This paper improves PANet by eliminating downsampling and using only upsampling to fuse feature maps from bottom to top, ensuring accurate preservation of spatial information.

[0108] The present invention first introduced FPN to solve the problem of hierarchical feature fusion in convolutional neural networks, and has been proven to effectively enhance the capabilities of deep learning models in processing detection tasks, especially in detecting objects of different scales. PANet is used to strengthen feature propagation and encourage information reuse, thereby improving the representation ability of the feature pyramid. FPN first extracts feature maps from the convolutional neural network, and then uses upsampling and coarse-grained feature maps to achieve feature map fusion in a top-down manner. In contrast, PANet fuses feature maps from bottom to top to ensure that spatial information is accurately preserved. However, the combination of FPN and PANet can only support top-down and bottom-up feature fusion. The structures of BiFPN, AFPN and GFPN are designed to quickly integrate features into various levels, and improve the effect of feature fusion by adding more levels to meet the feature fusion requirements at different levels.

[0109] This paper modifies the FPN-PANet architecture in YOLOv8, achieving multi-level feature fusion across different layers by strengthening the network's multi-path fusion. Using the CSP module, this paper adds skip connections, simultaneously sharing dense information across non-adjacent levels of varying spatial scales and latent semantics. This enables the model to balance processing of high-level semantic information with low-level spatial information in the neck network.

[0110] Specifically, the neck network includes: the feature map output from the CBAM_SPPF module is input into the first connection layer after passing through the third CBS module, the first connection layer also receives the feature map output from the fifth CBS module, and after being processed by the first connection layer, it is input into the first CSP module;

[0111] The data of the fifth CBS module comes from the seventh C2f_SE module, and the feature map output from the seventh C2f_SE module is also output to the second connection layer. A first upsampling is further provided between the first CSP module and the second connection layer. The second connection layer receives three-party data: the first upsampling output data, the seventh C2f_SE module output data, and the feature map of the fifth C2f_SE module output data after passing through the seventh CBS module;

[0112] The second connection layer processes the data and outputs it to the second CSP module, which then undergoes a second upsampling and outputs it to the third connection layer. The third connection layer also receives data from the fifth C2f_SE module and simultaneously outputs the data to the third CSP module and the third upsampling module. The third upsampling module transmits the data to the fourth connection layer. The fourth connection layer also receives the feature map from the third C2f_SE module and transmits the output data to the fourth CSP module.

[0113] After the fourth CSP module transfers the data to the ninth CBS module, it enters the fifth connection layer. The fifth connection layer also receives data from the third CSP module. After processing by the connection layer, the data enters the fifth CSP module, and then enters the sixth connection layer after passing through the tenth CBS module. At the same time, the sixth connection layer also receives data from the second CSP module. After connection, it enters the fifth CSP module and the eleventh CBS module. Together with the data from the second CSP module after passing through the twelfth CBS module, it enters the seventh connection layer and finally enters the segmentation detection head through the sixth CSP module.

[0114] The HRFI-YOLO network adds a detection head for small ground objects. The original YOLO series algorithm has three detection heads, with heights of 20×20, 40×40, and 80×80 widths, respectively. By comparison, these heads still cannot meet the detection and segmentation requirements for object recognition in high-resolution oblique aerial imagery, resulting in suboptimal detection accuracy for objects larger than the original scale. An additional 160×160 detection and segmentation head is introduced in the head, aligned with the new feature fusion network structure in the neck, to improve the detection capabilities of objects of various scales. A new scale detection and segmentation head is added as the fourth detection head, building on the original 80×80 detection scale of YOLOv8. It fuses the shallow information extracted from the input image by the first C2f module and combines it with other feature fusion networks to detect objects of a wider range of scales.

[0115] Specifically, the specific detection segmentation head specifically includes: a first detection head, which receives the feature map output from the third CSP module, corresponding to a size of 20×20; a second detection head, which receives the feature map output from the fourth CSP module, corresponding to a size of 40×40; a third detection head, which receives the feature map output from the fifth CSP module, corresponding to a size of 80×80; and a fourth detection head, which receives the feature map output from the sixth CSP module.

[0116] In step S3, data enhancement is performed on the dataset, including:

[0117] The visibility of ground objects is enhanced by adjusting the contrast and brightness of the original aerial image. This is achieved through a weighted algorithm. The algorithm formula is as follows:

[0118] Output=Input1×α+Input2×β+γ

[0119] Where Input represents the input image, Output represents the output image, α, β, and γ are parameters, α is set to 1.1, β is set to 0.1, and γ is set to 30.

[0120] In step S3, the HRFI-YOLO network is trained on the dataset, including:

[0121] The hardware environment for training the network is AMD Ryzen 7 5800H 16G + NVIDIA RTX 3090 24G, and the operating system is UBUNTU 20.04.

[0122] Set the number of training rounds to 1000, the number of images per batch to 4, the number of data loading worker threads to 4, turn off mosaic enhancement for the last 10 rounds of training, enable automatic mixed precision training, set the initial learning rate to 0.01, set the final learning rate to 0.01, set the SGD momentum to 0.937, and set the optimizer weight decay to 0.0005. After the above training, a model with better test results can be obtained.

[0123] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0124] Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if such changes and modifications of the embodiments of the present invention fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for identifying ground objects in a power transmission line engineering survey task, characterized in that: The method comprises the following steps: S1 uses ISAT semi-automatic annotation software to produce a high-resolution oblique aerial photography dataset; S2 designs a ground object recognition instance segmentation network based on deep learning, denoted as HRFI-YOLO network; Among them, the HRFI-YOLO network for object recognition based on deep learning includes: Backbone network: The improved Darknet-53 network is used as the backbone network, and the C2f module and CBAM module are introduced, specifically including: The first CBS module, the second CBS module, the third C2f_SE module, the fourth CBS module, the fifth C2f_SE module, the sixth CBS module, the seventh C2f_SE module, the eighth CBS module, the ninth C2f_SE module and the CBAM_SPPF module are connected in series in sequence. The CBAM_SPPF module introduces the CBAM attention mechanism in the backbone network before the feature map group is input into the SPPF module. The SPPF module includes multiple small-size pooling layers and a merge connection layer. The CBAM includes two key modules: a channel attention module and a spatial attention module. Neck network: The improved FPN-PAN network structure is used as the basic neck network structure, and the GFPN network structure idea and CSP module are integrated to construct a multi-layer feature fusion neck network; Detection and segmentation head: includes four detection heads of different scales, and a new detection head for small ground objects; S3 performs data augmentation on the high-resolution oblique aerial photography dataset and trains the HRFI-YOLO network on the dataset; S4 uses the trained object recognition network model to extract the object target; S5 interfaces the extracted results in the format required by the design platform; The neck network includes: the feature map output from the CBAM_SPPF module is input into the first connection layer after passing through the third CBS module, the first connection layer also receives the feature map output from the fifth CBS module, and after being processed by the first connection layer, is input into the first CSP module; The data of the fifth CBS module comes from the seventh C2f_SE module, and the feature map output from the seventh C2f_SE module is also output to the second connection layer. A first upsampling is further provided between the first CSP module and the second connection layer. The second connection layer receives three-party data: the first upsampling output data, the seventh C2f_SE module output data, and the feature map of the fifth C2f_SE module output data after passing through the seventh CBS module; The second connection layer processes the data and outputs it to the second CSP module, which then undergoes a second upsampling and outputs it to the third connection layer. The third connection layer also receives data from the fifth C2f_SE module and simultaneously outputs the data to the third CSP module and the third upsampling module. The third upsampling module transmits the data to the fourth connection layer. The fourth connection layer also receives the feature map from the third C2f_SE module and transmits the output data to the fourth CSP module. After the fourth CSP module transfers the data to the ninth CBS module, it enters the fifth connection layer. The fifth connection layer also receives data from the third CSP module. After processing by the connection layer, the data enters the fifth CSP module, and then enters the sixth connection layer after passing through the tenth CBS module. At the same time, the sixth connection layer also receives data from the second CSP module. After connection, it enters the fifth CSP module and the eleventh CBS module. Together with the data from the second CSP module after passing through the twelfth CBS module, it enters the seventh connection layer and finally enters the detection and segmentation head through the sixth CSP module.

2. The method for identifying ground objects in a power transmission line engineering survey task according to claim 1, characterized in that: The step S1 specifically includes: S11 creates a virtual environment, obtains the ISAT installation package source code and installs dependencies; S12 downloads the pre-trained model and stores it in the weight folder directory; S13 performs semi-automatic annotation and adjusts the target segmentation mask by prompting the region of interest or the region of no interest; S14 ISAT uses the JSON file format to save the annotation results. The JSON file is manually parsed and converted into TXT format. The training, validation, and test data sets are split by configuring the ratio of the validation set and the test set.

3. The method for identifying ground objects in power transmission line engineering survey tasks according to claim 1, characterized in that: The C2f_SE module is specifically: The SE mechanism is introduced into the existing C2f module. After the input feature map group is processed by the C2f module, the attention weight is weighted to the features of each channel through global information embedding and adaptive correction to output the feature map group. Among them, the C2f module contains 2 convolutional layers, 3 bottleneck layers and 1 merge connection layer.

4. The method for identifying ground objects in a power transmission line engineering survey task according to claim 1, characterized in that: In the detection and segmentation head, a new detection head for small ground objects is added, so that the network has four detection heads of different scales, including: An additional 160×160 detection head is introduced into the head network part and aligned with the multi-layer feature fusion neck network of the neck to improve the detection ability of targets of various scales.

5. A ground feature recognition system for power transmission line engineering survey tasks, characterized in that: The system includes: Data acquisition module, used to produce high-resolution oblique aerial photography datasets using ISAT semi-automatic annotation software; Network building module for designing the HRFI-YOLO network for deep learning-based object recognition; The deep learning-based ground object recognition HRFI-YOLO network includes: Backbone network: The improved Darknet-53 network is used as the backbone network, and the C2f module and CBAM module are introduced, specifically including: The first CBS module, the second CBS module, the third C2f_SE module, the fourth CBS module, the fifth C2f_SE module, the sixth CBS module, the seventh C2f_SE module, the eighth CBS module, the ninth C2f_SE module and the CBAM_SPPF module are connected in series in sequence. The CBAM_SPPF module introduces the CBAM attention mechanism in the backbone network before the feature map group is input into the SPPF module. The SPPF module includes multiple small-size pooling layers and a merge connection layer. The CBAM includes two key modules: a channel attention module and a spatial attention module. Neck network: The improved FPN-PAN network structure is used as the basic neck network structure, and the GFPN network structure idea and CSP module are integrated to construct a multi-layer feature fusion neck network; Detection and segmentation head: includes four detection heads of different scales, and a new detection head for small ground objects; The data enhancement module is used to perform data enhancement on the high-resolution oblique aerial photography dataset and train the HRFI-YOLO network on the dataset; The training module is used to extract ground object targets using the trained ground object recognition network model; The corresponding module is used to interface the extraction results with the format required by the design platform; The neck network includes: the feature map output from the CBAM_SPPF module is input into the first connection layer after passing through the third CBS module, the first connection layer also receives the feature map output from the fifth CBS module, and after being processed by the first connection layer, is input into the first CSP module; The data of the fifth CBS module comes from the seventh C2f_SE module, and the feature map output from the seventh C2f_SE module is also output to the second connection layer. A first upsampling is further provided between the first CSP module and the second connection layer. The second connection layer receives three-party data: the first upsampling output data, the seventh C2f_SE module output data, and the feature map of the fifth C2f_SE module output data after passing through the seventh CBS module; The second connection layer processes the data and outputs it to the second CSP module, which then undergoes a second upsampling and outputs it to the third connection layer. The third connection layer also receives data from the fifth C2f_SE module and simultaneously outputs the data to the third CSP module and the third upsampling module. The third upsampling module transmits the data to the fourth connection layer. The fourth connection layer also receives the feature map from the third C2f_SE module and transmits the output data to the fourth CSP module. After the fourth CSP module transfers the data to the ninth CBS module, it enters the fifth connection layer. The fifth connection layer also receives data from the third CSP module. After processing by the connection layer, the data enters the fifth CSP module, and then enters the sixth connection layer after passing through the tenth CBS module. At the same time, the sixth connection layer also receives data from the second CSP module. After connection, it enters the fifth CSP module and the eleventh CBS module. Together with the data from the second CSP module after passing through the twelfth CBS module, it enters the seventh connection layer and finally enters the detection and segmentation head through the sixth CSP module.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a ground object recognition program, which, when executed by a processor, implements the steps of the ground object recognition method for power transmission line engineering survey tasks according to any one of claims 1 to 4.

Images