Zero sample learning-based power grid line diagram segmentation method and system

Abstract

The invention discloses a zero sample learning-based power grid line diagram segmentation method and system, and the method comprises the steps: firstly obtaining a power grid line aerial view as input data, then learning the image features of the input data, obtaining an image detection frame, updating a cross attention function through adding an adaptive feature positioning algorithm, and obtaining a self-correction model, and finally, inputting the image detection frame as a frame prompt of the self-correction model, and introducing a multi-scale reasoning mechanism to segment the image. In application, an open set target detection algorithm is combined with an everything segmentation algorithm, and an adaptive feature positioning algorithm is designed for self-correction, so that instance segmentation of a power grid line image under a zero sample is realized, the defect of low labeling precision is overcome, and meanwhile, a multi-scale reasoning mechanism is introduced, so that the accuracy of the power grid line image is improved. And the segmentation accuracy and robustness are improved. The method is low in cost, and can achieve the precise segmentation of a complex scene image.

Description

Technical Field

[0001] The present invention relates to an image processing method, belonging to the technical field of image processing, and in particular to a power grid line diagram segmentation method and system based on zero-sample learning. Background Art

[0002] With the rapid development of the power industry and the acceleration of urbanization, the safe operation of the power grid has received more and more attention from the society, and the timely and accurate detection of power grid lines is an important part of ensuring the safe operation of the power grid; the traditional power grid line detection method mainly relies on manual inspection, which is not only inefficient but also may have safety hazards. With the rapid development of drone technology, the use of drones for aerial inspection of power grid lines has become a new solution. Drone aerial photography can provide a large number of high-resolution images, providing rich data resources for the detection of power grid lines.

[0003] However, the power line images obtained by drone aerial photography usually contain a lot of background noise and complex scenes, which brings great challenges to the instance segmentation of power line. Traditional instance segmentation methods usually require a large amount of labeled data to train the model. In power lines, the detection scenes are relatively complex. It is unrealistic to obtain a large amount of labeled data for complex scenes, which requires a high cost. Therefore, a low-cost method is urgently needed to achieve accurate segmentation of complex scene images of power lines. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned defects and problems existing in the prior art and to provide a power grid line diagram segmentation method and system based on zero-sample learning which is low in cost and can realize accurate segmentation of complex scene images.

[0005] To achieve the above objectives, the technical solution of the present invention is: a power grid line diagram segmentation method based on zero-sample learning, comprising:

[0006] S1. Obtain a bird's-eye view of the power grid line as input data;

[0007] S2. Based on the open set target detection algorithm, learn the image features of the input data and obtain the image detection frame of the input data;

[0008] S3. Based on the everything segmentation algorithm, an adaptive feature localization algorithm is added to the decoder module of the everything segmentation algorithm to update the image label cross attention function of the everything segmentation algorithm to obtain a self-correcting everything segmentation model;

[0009] The updated image label cross attention function is as follows:

[0010] DCAttn(t,x)=σ(Q(t)·K(xp+Δp) T)·V(xp+Δp);

[0011] Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position;

[0012] S4. The image detection frame of the input data is used as the frame prompt of the self-correcting object segmentation model and a multi-scale reasoning mechanism is introduced to segment the bird's-eye view of the power grid line to complete the instance segmentation of the bird's-eye view of the power grid line.

[0013] The step S3 specifically refers to:

[0014] The original image label cross attention function formula of the object segmentation algorithm is as follows:

[0015] CAttn(t,x)=σ(Q(t)·K(xp) T )·V(xp);

[0016] An adaptive feature localization algorithm is added to the decoder module of the object segmentation algorithm. The adaptive feature localization algorithm first performs feature extraction, and the input feature F(I) represents the feature vector extracted from the image feature. Then, the adjustment function is defined as G(F(I), θ), and the parameter θ is adjusted according to the input feature F(I). Finally, the image label cross attention function is updated using the adjusted parameter θ′=G(F(I), θ); wherein: F(I)=xp, θ′=Δp; Δp is the offset, and its expression is as follows:

[0017] Δp=θ s (θ offset (xp));

[0018] Where: θ s is the scaling function, θ offset For offset network.

[0019] In step S4, the introducing of a multi-scale reasoning mechanism to segment the bird's-eye view of the power grid line refers to:

[0020] Firstly, the original image is converted into different scales to obtain the image features of each scale. Then, considering the global large-scale features and local small-scale features of the image features, the image features are extracted and segmented respectively based on the everything segmentation algorithm. Subsequently, the image features of each scale are fused to obtain the fused multi-scale features. Finally, the bird's-eye view of the power grid line is segmented based on the fused multi-scale features to complete the instance segmentation of the bird's-eye view of the power grid line.

[0021] In step S1, the power grid line bird's-eye view refers to:

[0022] The bird's-eye view of power grid lines is obtained through drone aerial photography or the Internet, and the bird's-eye view of power grid lines does not contain additional label information that can be used to directly indicate specific parts or characteristics of image content.

[0023] The image features of the input data include: visual features and semantic features of brackets, towers and lines in a bird's-eye view of power grid lines;

[0024] The visual features include: color, edge, texture, shape, outline, specific pattern, line direction, connection point, density, environmental background, topography, facilities and equipment;

[0025] The semantic features include: spatial relationship, interactive relationship, line type, line status, and safety hazard.

[0026] A power grid line diagram segmentation system based on zero-shot learning, comprising:

[0027] An input data acquisition module is used to acquire a bird's-eye view of a power grid line as input data;

[0028] An image feature acquisition module is used to learn the image features of input data based on an open set target detection algorithm and obtain an image detection frame of the input data;

[0029] A self-correcting model building module is used to add an adaptive feature localization algorithm to the decoder module of the everything segmentation algorithm to update the image label cross-attention function of the everything segmentation algorithm to obtain a self-correcting everything segmentation model;

[0030] The updated image label cross attention function is as follows:

[0031] DCAttn(t,x)=σ(Q(t)·K(xp+Δp) T )·V(xp+Δp);

[0032] Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position;

[0033] The instance segmentation module is used to input the image detection box of the input data as the box prompt of the self-correcting object segmentation model and introduce a multi-scale reasoning mechanism to segment the bird's-eye view of the power grid line, thereby completing the instance segmentation of the bird's-eye view of the power grid line.

[0034] The self-correction model construction module is used to obtain a self-correction all things segmentation model according to the following method;

[0035] The original image label cross attention function formula of the object segmentation algorithm is as follows:

[0036] CAttn(t,x)=σ(Q(t)·K(xp) T )·V(xp);

[0037] An adaptive feature localization algorithm is added to the decoder module of the object segmentation algorithm. The adaptive feature localization algorithm first performs feature extraction, and the input feature F(I) represents the feature vector extracted from the image feature. Then, the adjustment function is defined as G(F(I), θ), and the parameter θ is adjusted according to the input feature F(I). Finally, the image label cross attention function is updated using the adjusted parameter θ′=G(F(I), θ); wherein: F(I)=xp, θ′=Δp; Δp is the offset, and its expression is as follows:

[0038] Δp=θ s (θ offset (xp));

[0039] Where: θ s is the scaling function, θ offset For the offset network;

[0040] Then the updated image label cross attention function is as follows:

[0041] DCAttn(t,x)=σ(Q(t)·K(xp+Δp) T )·V(xp+Δp);

[0042] Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position.

[0043] The instance segmentation module is used to perform instance segmentation according to the following method;

[0044] Firstly, the original image is converted into different scales to obtain the image features of each scale. Then, considering the global large-scale features and local small-scale features of the image features, the image features are extracted and segmented respectively based on the everything segmentation algorithm. Subsequently, the image features of each scale are fused to obtain the fused multi-scale features. Finally, the bird's-eye view of the power grid line is segmented based on the fused multi-scale features to complete the instance segmentation of the bird's-eye view of the power grid line.

[0045] The input data acquired by the input data acquisition module includes:

[0046] The bird's-eye view of power grid lines is obtained through drone aerial photography or the Internet, and the bird's-eye view of power grid lines does not contain additional label information that can be used to directly indicate specific parts or characteristics of image content.

[0047] The image features of the input data acquired by the image feature acquisition module include:

[0048] The visual and semantic features of supports, towers and lines in the bird's-eye view of power grid lines;

[0049] The visual features include: color, edge, texture, shape, outline, specific pattern, line direction, connection point, density, environmental background, topography, facilities and equipment;

[0050] The semantic features include: spatial relationship, interactive relationship, line type, line status, and safety hazard.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The invention discloses a method and system for segmenting a power line diagram based on zero-sample learning. The method first obtains a bird's-eye view of a power line as input data, then learns the image features of the input data based on an open set target detection algorithm, obtains an image detection frame of the input data, then adds an adaptive feature positioning algorithm to its decoder module based on an all-things segmentation algorithm to update the image label cross-attention function, obtains a self-correcting all-things segmentation model, and finally inputs the image detection frame of the input data as a frame prompt of the self-correcting all-things segmentation model and introduces a multi-scale reasoning mechanism to segment the power line bird's-eye view, thereby completing the instance segmentation of the power line bird's-eye view; in the application, the design combines the open set target detection algorithm with the all-things segmentation algorithm, and designs an adaptive feature positioning algorithm to self-correct the all-things segmentation algorithm, thereby realizing the instance segmentation of the power line image under zero samples, and solving the defect that the automatic labeling accuracy of the detection frame of the open set target detection algorithm is not high, and at the same time, by introducing a multi-scale reasoning mechanism, the accuracy and robustness of the complex scene segmentation of the power line are further improved. Therefore, the invention is not only low in cost, but also can realize the accurate segmentation of complex scene images. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the method steps in the present invention.

[0054] Figure 2 It is a schematic diagram of the system structure in the present invention.

[0055] Figure 3 It is a schematic diagram of the device structure in the present invention.

[0056] In the figure: input data acquisition module 1, image feature acquisition module 2, self-correction model construction module 3, instance segmentation module 4, processor 5, memory 6, computer program code 61. DETAILED DESCRIPTION

[0057] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0058] Embodiment 1:

[0059] See also Figure 1 , a power grid line diagram segmentation method based on zero-shot learning, comprising:

[0060] S1. Obtain a bird's-eye view of the power grid line as input data;

[0061] Furthermore, the power grid line bird's-eye view refers to: a power grid line bird's-eye view obtained by drone aerial photography or the Internet, and the power grid line bird's-eye view does not contain additional label information that can be used to directly indicate a specific part or feature of the image content;

[0062] S2, based on the open set target detection algorithm (Grounding Dino), learn the image features of the input data and obtain the image detection frame of the input data;

[0063] Furthermore, the open set object detection algorithm is pre-trained on large-scale unlabeled data through self-supervised learning, which enables the algorithm to effectively learn a universal representation of images without the need for additional training steps.

[0064] The large-scale unlabeled data refers to a large amount of image data that has not been labeled or classified. It can be images of power grid lines taken by drones or images from other sources. The key is that these images have no additional label information to directly indicate the specific parts or characteristics of the image content. These large-scale unlabeled data are used for pre-training through self-supervised learning methods. The purpose is to enable the open-set target detection algorithm to learn to extract common feature representations from images, so that the open-set target detection algorithm can understand and process image data without explicit labels, providing a basis for subsequent specific tasks.

[0065] The basic principle of self-supervised learning is as follows:

[0066] Generative tasks: The algorithm learns by performing some kind of prediction task on the input grid line image data, for example: predicting a part of the image or a certain attribute.

[0067] Intrinsic Labels: These prediction tasks do not require external labels, instead, they use the data itself as “labels”, for example, by occluding part of the image and letting the algorithm predict what is occluded.

[0068] Application in open set target detection algorithm:

[0069] First, feature extraction is performed on the power line images, which usually involves identifying and encoding important visual patterns and structures in the images;

[0070] Next, the algorithm learns a general representation of the image by solving automatically generated tasks such as reconstruction, prediction, or other forms of internal tasks. In this process, the algorithm learns the ability to capture the intrinsic characteristics and patterns of the image rather than focusing on a specific prediction task.

[0071] Finally, its understanding of image representation is further refined through repeated iterations and optimization processes.

[0072] The image features of the input data include: visual features and semantic features of brackets, towers and lines in a bird's-eye view of power grid lines;

[0073] The visual features include: color, edge, texture, shape, outline, specific pattern, line direction, connection point, density, environmental background, topography, facilities and equipment;

[0074] The semantic features include: spatial relationship, interactive relationship, line type, line status, and safety hazard.

[0075] These image features help identify the basic shape, structure, layout, relative position, specific object categories, scene types, and dynamic environments of different objects and scenes, which means that the algorithm can capture features that are widely applicable to a variety of different images, rather than being limited to specific categories or scenes.

[0076] S3. Based on the Segment Anything SAM algorithm, an adaptive feature localization algorithm is added to the decoder module of the Segment Anything SAM algorithm to update its image tag cross-attention function to obtain a self-correcting SAM model.

[0077] Furthermore, the original image labeling cross attention formula of the object segmentation algorithm is as follows:

[0078] CAttn(t,x)=σ(Q(t)·K(xp) T )·V(xp);

[0079] An adaptive feature localization plugin (AFLP) is added to the decoder module of the object segmentation algorithm. The adaptive feature localization plugin (AFLP) is used to offset the network θ. offset Learn the offset Δp of the feature sampling position from the image feature xp to predict the offset of the feature sampling position;

[0080] The expression of the offset is as follows:

[0081] Δp=θ s (θ offset (xp));

[0082] Where: θ s is the scaling function;

[0083] Then the updated feature sampling position is p+Δp, where: p is the original image feature space sampling position;

[0084] Then the updated image label cross attention function is as follows:

[0085] DCAttn(t,x)=σ(Q(t)·K(xp+Δp) T )·V(xp+Δp);

[0086] Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position.

[0087] The step of updating the image label cross attention function of the adaptive feature localization algorithm includes:

[0088] (1) Feature extraction:

[0089] Let the input feature F(I) represent the feature vector extracted from the image features;

[0090] (2) Parameter adjustment:

[0091] Define the adjustment function as G(F(I), θ), and adjust the parameter θ of the cross attention function according to the input feature F(I);

[0092] (3) Update the image label cross attention function:

[0093] Update the image label cross-attention function using the updated parameters θ′=G(F(I),θ); where: F(I)=xp, θ′=Δp.

[0094] After adding AFLP, the decoder part of the object segmentation algorithm needs to be retrained. By adjusting the sampling position of image features, the object segmentation algorithm can be helped to focus on the target area more accurately when processing low-quality cues, thereby improving the accuracy and stability of segmentation.

[0095] Let the segmentation result be S(I), and minimize the segmentation error L by fine-tuning the decoder module in the object segmentation algorithm. seg =|S(I)-GT|2, where: GT is the actual instance segmentation label.

[0096] S4. The image detection frame of the input data is used as the frame prompt of the self-correcting object segmentation model, and a multi-scale reasoning mechanism is introduced to segment the bird's-eye view of the power grid line, thereby completing the instance segmentation of the bird's-eye view of the power grid line.

[0097] Furthermore, the original image is first converted into different scales to obtain image features of each scale, and then the global large-scale features and local small-scale features of the image features are considered, and the image features are extracted and segmented based on the all-things segmentation algorithm, and then the image features of each scale are fused to obtain fused multi-scale features, and finally the bird's-eye view of the power grid line is segmented based on the fused multi-scale features to complete the instance segmentation of the bird's-eye view of the power grid line;

[0098] Assume that the segmentation results of different scales are S1(I), S2(I), ...S n (1), the final segmentation result is the fusion of these segmentation results, namely:

[0099]

[0100] Multi-scale reasoning enables the model to better adapt to objects of different sizes and shapes, and also improves the model's adaptability to image scale changes and different resolutions, which is especially important for drone images taken from various angles and distances.

[0101] Embodiment 2:

[0102] See also Figure 2 , a power grid line diagram segmentation system based on zero-shot learning, comprising:

[0103] An input data acquisition module 1 is used to acquire a bird's-eye view of a power grid line as input data;

[0104] Furthermore, the input data acquired by the input data acquisition module 1 includes:

[0105] Acquire a bird's-eye view of power grid lines through drone aerial photography or the Internet, wherein the bird's-eye view of power grid lines does not contain additional label information that can be used to directly indicate specific parts or characteristics of image content;

[0106] Image feature acquisition module 2, used to learn the image features of input data based on an open set target detection algorithm and obtain an image detection frame of the input data;

[0107] Furthermore, the image features of the input data acquired by the image feature acquisition module 2 include:

[0108] The visual and semantic features of supports, towers and lines in the bird's-eye view of power grid lines;

[0109] The visual features include: color, edge, texture, shape, outline, specific pattern, line direction, connection point, density, environmental background, topography, facilities and equipment;

[0110] The semantic features include: spatial relationship, interactive relationship, line type, line status, and safety hazard;

[0111] A self-correction model construction module 3 is used to update the image label cross-attention function of the all-things segmentation algorithm based on the all-things segmentation algorithm by adding an adaptive feature localization algorithm in its decoder module to obtain a self-correction all-things segmentation model;

[0112] The updated image label cross attention function is as follows:

[0113] DCAttn(t,x)=σ(Q(t)·K(xp+Δp) T )·V(xp+Δp);

[0114] Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position;

[0115] Furthermore, the self-correction model construction module 3 is used to obtain a self-correction all-things segmentation model according to the following method;

[0116] The original image label cross attention function formula of the object segmentation algorithm is as follows:

[0117] CAttn(t,x)=σ(Q(t)·K(xp) T )·V(xp);

[0118] An adaptive feature localization algorithm is added to the decoder module of the object segmentation algorithm. The adaptive feature localization algorithm first performs feature extraction, and the input feature F(I) represents the feature vector extracted from the image feature. Then, the adjustment function is defined as G(F(I), θ), and the parameter θ is adjusted according to the input feature F(I). Finally, the image label cross attention function is updated using the adjusted parameter θ′=G(F(I), θ); wherein: F(I)=xp, θ′=Δp; Δp is the offset, and its expression is as follows:

[0119] Δp=θ s (θ offset (xp));

[0120] Where: θ s is the scaling function, θ offset For the offset network;

[0121] Then the updated image label cross attention function is as follows:

[0122] DCAttn(t,x)=σ(Q(t)·K(xp+Δp) T )·V(xp+Δp);

[0123] Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position.

[0124] The instance segmentation module 4 is used to input the image detection frame of the input data as a frame prompt of the self-correcting object segmentation model and introduce a multi-scale reasoning mechanism to segment the power grid line bird's-eye view, thereby completing the instance segmentation of the power grid line bird's-eye view.

[0125] Furthermore, the instance segmentation module 4 is used to perform instance segmentation according to the following method;

[0126] Firstly, the original image is converted into different scales to obtain the image features of each scale. Then, considering the global large-scale features and local small-scale features of the image features, the image features are extracted and segmented respectively based on the everything segmentation algorithm. Subsequently, the image features of each scale are fused to obtain the fused multi-scale features. Finally, the bird's-eye view of the power grid line is segmented based on the fused multi-scale features to complete the instance segmentation of the bird's-eye view of the power grid line.

[0127] Embodiment 3:

[0128] See also Figure 3 , a power grid line diagram segmentation device based on zero-sample learning, the device comprising a processor 5 and a memory 6;

[0129] The memory 6 is used to store the computer program code 61 and transmit the computer program code 61 to the processor 5;

[0130] The processor 5 is used to execute the power grid line diagram segmentation method based on zero-sample learning according to the instructions in the computer program code 61.

[0131] Generally speaking, the computer instructions for implementing the method of the present invention may be carried by any combination of one or more computer-readable storage media. Non-transitory computer-readable storage media may include any computer-readable media, except for the signal itself that is temporarily propagating.

[0132] The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EKROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, device, or device.

[0133] Computer program code for performing the operation of the present invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, SMalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages, in particular, Python suitable for neural network computing and platform frameworks based on TensorFlow, PyTorch, etc. can be used. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer or to an external computer (for example, using an Internet service provider to connect via the Internet) through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0134] The above-mentioned device and non-temporary computer-readable storage medium can refer to the specific description of the power grid line diagram segmentation method based on zero-sample learning and its beneficial effects, which will not be repeated here.

[0135] Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A power grid line diagram segmentation method based on zero-shot learning, characterized in that: include: S1. Obtain a bird's-eye view of the power grid line as input data; S2. Based on the open set target detection algorithm, learn the image features of the input data and obtain the image detection frame of the input data; S3. Based on the everything segmentation algorithm, an adaptive feature localization algorithm is added to the decoder module of the everything segmentation algorithm to update the image label cross attention function of the everything segmentation algorithm to obtain a self-correcting everything segmentation model; The updated image label cross attention function is as follows: DCAttn(t,x)=σ(Q(t)·K(xp+Δp) T )·V(xp+Δp); Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position; S4. The image detection frame of the input data is used as the frame prompt of the self-correcting object segmentation model and a multi-scale reasoning mechanism is introduced to segment the bird's-eye view of the power grid line to complete the instance segmentation of the bird's-eye view of the power grid line.

2. The method for segmenting a power grid diagram based on zero-sample learning according to claim 1, characterized in that: The step S3 specifically refers to: The original image label cross attention function formula of the object segmentation algorithm is as follows: DCAttn(t,x)=σ(Q(t)·K(xp) T )·V(xp); An adaptive feature localization algorithm is added to the decoder module of the object segmentation algorithm. The adaptive feature localization algorithm first performs feature extraction, and the input feature F(I) represents the feature vector extracted from the image feature. Then, the adjustment function is defined as G(F(I), θ), and the parameter θ is adjusted according to the input feature F(I). Finally, the image label cross attention function is updated using the adjusted parameter θ′=G(F(I), θ); wherein: F(I)=xp, θ′=Δp; Δp is the offset, and its expression is as follows: Δp=θ s (i offset (xp)); Where: θ s is the scaling function, θ offset For offset network.

3. The method for segmenting a power grid diagram based on zero-sample learning according to claim 1, characterized in that: In step S4, the introducing of a multi-scale reasoning mechanism to segment the bird's-eye view of the power grid line refers to: Firstly, the original image is converted into different scales to obtain the image features of each scale. Then, considering the global large-scale features and local small-scale features of the image features, the image features are extracted and segmented respectively based on the everything segmentation algorithm. Subsequently, the image features of each scale are fused to obtain the fused multi-scale features. Finally, the bird's-eye view of the power grid line is segmented based on the fused multi-scale features to complete the instance segmentation of the bird's-eye view of the power grid line.

4. The method for segmenting a power grid diagram based on zero-sample learning according to claim 1, characterized in that: In step S1, the power grid line bird's-eye view refers to: The bird's-eye view of power grid lines is obtained through drone aerial photography or the Internet, and the bird's-eye view of power grid lines does not contain additional label information that can be used to directly indicate specific parts or characteristics of image content.

5. The method for segmenting a power grid line diagram based on zero-sample learning according to claim 1, characterized in that: The image features of the input data include: visual features and semantic features of brackets, towers and lines in a bird's-eye view of power grid lines; The visual features include: color, edge, texture, shape, outline, specific pattern, line direction, connection point, density, environmental background, topography, facilities and equipment; The semantic features include: spatial relationship, interactive relationship, line type, line status, and safety hazard.

6. A power grid line diagram segmentation system based on zero-sample learning, characterized in that: include: An input data acquisition module (1) is used to acquire a bird's-eye view of a power grid line as input data; An image feature acquisition module (2) is used to learn the image features of the input data based on an open set target detection algorithm to obtain an image detection frame of the input data; A self-correction model construction module (3), which is used to add an adaptive feature localization algorithm to the decoder module of the object segmentation algorithm to update the image label cross-attention function of the object segmentation algorithm to obtain a self-correction object segmentation model; The updated image label cross attention function is as follows: DCAttn(t,n)=σ(Q(t)·K(xp+Δp) T )·V(xp+Δp); Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position; The instance segmentation module (4) is used to input the image detection frame of the input data as a frame prompt of the self-correcting object segmentation model and introduce a multi-scale reasoning mechanism to segment the power grid line bird's-eye view image, thereby completing the instance segmentation of the power grid line bird's-eye view image.

7. The power grid line diagram segmentation system based on zero-sample learning according to claim 6, characterized in that: The self-correction model construction module (3) is used to obtain a self-correction all-things segmentation model according to the following method; The original image label cross attention function formula of the object segmentation algorithm is as follows: CAttn(t,x)=σ(Q(t)·K(xp) T )·V(xp); An adaptive feature localization algorithm is added to the decoder module of the object segmentation algorithm. The adaptive feature localization algorithm first performs feature extraction, and the input feature F(I) represents the feature vector extracted from the image feature. Then, the adjustment function is defined as G(F(I), θ), and the parameter θ is adjusted according to the input feature F(I). Finally, the image label cross attention function is updated using the adjusted parameter θ′=G(F(I), θ); wherein: F(I)=xp, θ′=Δp; Δp is the offset, and its expression is as follows: Δp=θ s (i offset (xp)); Where: θ s is the scaling function, θ offset For the offset network; Then the updated image label cross attention function is as follows: DCAttn(t,x)=σ(Q(t)·K(xp+Δp) T )·V(xp+Δp); Where: t is the feature of the prompt mark, xp is the image feature, Q, K, V are the embedding projection functions of the query, key and value respectively, σ is the normalized exponential function, and xp+Δp is the image feature obtained by resampling at the updated sampling position.

8. The power grid line diagram segmentation system based on zero-sample learning according to claim 6, characterized in that: The instance segmentation module (4) is used to perform instance segmentation according to the following method; Firstly, the original image is converted into different scales to obtain the image features of each scale. Then, considering the global large-scale features and local small-scale features of the image features, the image features are extracted and segmented respectively based on the everything segmentation algorithm. Subsequently, the image features of each scale are fused to obtain the fused multi-scale features. Finally, the bird's-eye view of the power grid line is segmented based on the fused multi-scale features to complete the instance segmentation of the bird's-eye view of the power grid line.

9. The power grid line diagram segmentation system based on zero-sample learning according to claim 6, characterized in that: The input data acquired by the input data acquisition module (1) includes: The bird's-eye view of power grid lines is obtained through drone aerial photography or the Internet, and the bird's-eye view of power grid lines does not contain additional label information that can be used to directly indicate specific parts or characteristics of image content.

10. The power grid line diagram segmentation system based on zero-sample learning according to claim 6, characterized in that: The image features of the input data acquired by the image feature acquisition module (2) include: The visual and semantic features of supports, towers and lines in the bird's-eye view of power grid lines; The visual features include: color, edge, texture, shape, outline, specific pattern, line direction, connection point, density, environmental background, topography, facilities and equipment; The semantic features include: spatial relationship, interactive relationship, line type, line status, and safety hazard.

Images