Distribution network line icing image intelligent analysis system and method based on deep learning

Abstract

The embodiment of the invention discloses a distribution network line icing image intelligent analysis system and method based on deep learning, and the system comprises an image data collection module which obtains a large number of images related to the icing of a distribution network line; using an image labeling tool (Label Img) to carry out preset classification labeling on the icing image of the distribution line; an SK-YOLOv8 identification model network architecture is constructed; training the network architecture of the SK-YOLOv8 identification model by using the annotation data to obtain the SK-YOLOv8 identification model for identifying the icing state of the electric line; the image analysis and storage module is used for analyzing the target image, generating a report and storing the content; and the user interaction module is used for inputting an image by a user, calling the storage library to identify a target image and feeding back to the user. The system can provide a convenient image recognition tool for distribution network workers, and enables the distribution network workers to deal with a distribution network line icing scene more simply and efficiently.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing systems, and in particular to a system and method for intelligently analyzing ice-covered images of distribution network lines based on deep learning. Background Art

[0002] The distribution network is the process of transmitting electricity from power plants to end users, and the operating status of the distribution network system directly affects the reliability and quality of power supply. Distribution network equipment such as overhead lines, cables, poles, etc. are easily affected by wildfires and line icing due to being in the external environment, causing line damage and seriously endangering the safety of the distribution network. In addition, traditional distribution network management usually relies on manual inspections and manual operations, which have problems such as low inspection efficiency, high costs, and difficulty in real-time monitoring. In order to solve these problems, automatic image recognition technology came into being and has a wide range of applications.

[0003] Ice accumulation in cold seasons may cause power lines to bend, break or lose power. Early detection of ice accumulation can prompt power companies to take necessary preventive measures, such as solutions for ice-covered lines. Distribution network management based on image automatic recognition technology can improve operational efficiency, ensure power supply quality, and reduce operation and maintenance costs. It is an important direction for the development of technology in the distribution network field. Image automatic recognition technology can realize automatic identification and monitoring of equipment status, data processing of intelligent distribution network management, improve operation and maintenance efficiency and equipment safety, and improve user experience. With the continuous advancement of technology, image automatic recognition technology will play an increasingly important role in the distribution network field. Summary of the invention

[0004] The purpose of the present invention is to provide a distribution network line ice-covered image intelligent analysis system and method based on deep learning, so as to solve the problems in the prior art that line ice-covered image is difficult to identify and the maintenance cost is high, and the execution efficiency is low and there is no way to ensure accuracy.

[0005] The technical solution of the present invention is: a distribution network line ice-covered image intelligent analysis system based on deep learning, including: image data acquisition and image classification module, image annotation module, model network construction and model recognition module, image analysis and storage module; the image data acquisition and image classification module includes image acquisition and image classification, the image annotation module includes using tools to annotate images, the model network construction and model recognition module includes building an SK-YOLOv8 model, and using the model to perform target image recognition on the annotated image, and the image analysis and storage module includes analyzing the image after model recognition and generating a report for storage; the system front end is a user interaction module;

[0006] The image data acquisition module and the image classification module are used to obtain a large number of images related to ice coverage of distribution network lines and classify the images into positive samples and negative samples;

[0007] The image annotation module uses an image annotation tool (Label Img) to perform preset classification annotation on the distribution line ice-covered image;

[0008] The model network construction and model recognition module uses the labeled data to train the SK-YOLOv8 recognition model to identify the ice-covered state of the power line;

[0009] The image analysis and storage module analyzes the image after the model recognition, generates a report, and stores the content;

[0010] The user interaction module includes user input and output. The user inputs an image, calls the repository to perform target image recognition, and provides feedback to the user. The image analysis module determines whether the pixel value deviation of each sampling area of ​​the input image meets the set consistency requirements based on the user's input image;

[0011] If yes, a report is generated and the report results are fed back to the user;

[0012] If not, the user is informed that there is no valid result for the input image, and the image is stored in another storage terminal.

[0013] If the image reaches the preset number of queries, the image is input into the system backend for training and learning.

[0014] The image processing method in the above-mentioned distribution network line ice coverage image intelligent analysis system includes:

[0015] 1) The image classification module classifies images according to set rules and classifies images based on whether they contain target images;

[0016] 2) A method in which the image annotation module annotates the data based on preset classification categories.

[0017] The method for the image annotation module to annotate the data based on the preset classification category comprises the following steps:

[0018] (1) Mark the ice-covered distribution line images containing ice-covered targets as positive samples;

[0019] (2) The ice-covered distribution line images that do not contain ice-covered targets in the image are marked as negative samples.

[0020] The method for the image annotation module to annotate the data based on a preset classification category includes:

[0021] (1) Obtain image training data including positive samples and negative samples;

[0022] (2) Use the image annotation tool (Label Img) to preset classification annotations for the distribution line ice-covered images.

[0023] The image training data is obtained by randomly selecting 70% of the samples as a training set, 20% of the samples as a test set, and 10% of the samples as a prediction set.

[0024] The construction method of the SK-YOLOv8 model in the above-mentioned deep learning-based distribution network line ice-covered image intelligent analysis system includes:

[0025] Methods for constructing SK mechanism modules and YOLOv8 model networks;

[0026] SK-YOLOv8 recognition model network architecture construction method.

[0027] The steps of constructing the SK mechanism module include:

[0028] The SK module is an attention mechanism based on the convolution kernel. It can adaptively adjust the size of the convolution kernel according to the size of the target to improve the perception ability. The processing of the SK module includes the following three steps:

[0029] Split: By using two convolution kernels of different sizes for Group Convolution and ReLU activation functions, and using 3x3 and 5x5 convolution kernels to process the input feature map, two different feature maps are generated to test the sensitivity of different convolution kernels to the target and improve the detection accuracy;

[0030] Fuse: According to the two different feature maps obtained above, the information flow is controlled by the gate to fuse, and the results of the two branches are fused using element-wise summation, where Fgp is the global average pooling operation and Ffc is a two-layer fully connected layer of dimensionality reduction and dimensionality increase;

[0031] Select: Use two different weight matrices to weight the results of the two branches, and then sum the weighted results to get the output vector. This module adjusts the receptive field through different convolutional layers to improve the detection ability and accuracy of multi-scale targets. The receptive field refers to the area or range of input data that a specific neuron (or pixel of the feature map) in the network can perceive. The size of the receptive field is crucial to the performance of the model, especially when dealing with tasks with spatial characteristics (such as image classification, target detection, etc.). A sufficiently large receptive field allows the network to obtain contextual information to better understand the input data.

[0032] The YOLOv8 model network construction consists of four parts: input end, backbone network, feature enhancement module, and detection module:

[0033] Input: Data enhancement and adaptive image scaling are performed. Mosaic enhancement is mainly used for data enhancement. The Mosaic enhancement part of YOLOv8 introduces the operation of Mosaic enhancement in the last 10 epochs of YOLOX, which enhances the robustness of the model. Adaptive image scaling scales the original images to the same size.

[0034] Backbone network: It consists of CBS, C2f and SPPF modules. The CBS module mainly includes Conv, BN and SiLU activation functions. This module mainly performs convolution operations on feature maps. The C2f module is the main module for extracting features. The SPPF (Spatial Pyramid Pooling Fast) module is used to extract features of different receptive fields. This structure enables the network to better adapt to targets of different scales.

[0035] Feature Enhancement Module (Neck): The feature enhancement module is mainly composed of FPN and PAN. FPN is a top-down feature pyramid network structure that generates a prediction feature map by fusing high-level and low-level feature information through upsampling. PAN (Path Aggregation Network) is an improvement on FPN. It introduces semantic information of lateral connection enhancement features, so that the bottom-up feature map can be fused with the top-down feature map.

[0036] Detection module (Detect): Using the anchor-free matching mechanism, the target center point and width and height of feature maps of different scales are regressed, which greatly reduces the calculation time. Finally, by utilizing the rich information from feature maps of different scales, the classification and location information of large, medium and small-sized targets can be accurately obtained.

[0037] The SK-YOLOv8 recognition model network architecture construction includes: the structure is composed of four parts: input end, backbone network, feature enhancement module, and detection module. In the SK-YOLOv8 network, the SK mechanism is introduced into the C2f module on the backbone network and the Neck network for improvement, while keeping the overall network structure and processing unit unchanged. Such improvement improves the feature extraction capability and adaptability of the network, while retaining the advantages of the light weight and high efficiency of the YOLOv8 network;

[0038] The SK-YOLOv8 network uses different loss functions to calculate the target score, category probability score and position loss. The specific loss functions include BECLogits, which is used to calculate the target score, BCEclsloss, which is used to calculate the category probability score, and GIoU Loss, which is used to evaluate the position loss between the predicted box and the true box. The combination of these loss functions can effectively optimize the target detection task and improve the detection performance. The loss of the target score uses the BCEclsloss loss function to calculate the category probability score, and GIoU Loss is used to evaluate the position loss between the predicted box and the true box.

[0039] The calculation formula of GIoU Loss is: let p1 and p2 be the coordinates of the upper left and lower right vertices of the predicted box, and l1 and l2 be the coordinates of the upper left and lower right vertices of the real box, then:

[0040]

[0041] GIoU Loss = 1-GIoU

[0042] IoU is the ratio of the intersection and union areas of the predicted box and the true box, referred to as the intersection-union ratio. A is the predicted box, B is the true box, C is the minimum bounding box of A and B, S1 represents the area of ​​the overlapping area of ​​A and B, S2 represents the area after A and B are merged, S3 is the area of ​​the minimum rectangular box surrounding A and B, and GIoU Loss represents the measure of position loss. represents the prediction box boundary, Indicates the horizontal coordinate of the upper left corner of the prediction box, Indicates the ordinate of the upper left corner of the prediction box, Indicates the horizontal coordinate of the lower right corner of the prediction box, Indicates the ordinate of the lower right corner of the prediction box; represents the real box boundary, Indicates the horizontal coordinate of the upper left corner of the real box, Indicates the ordinate of the upper left corner of the real box, Indicates the horizontal coordinate of the lower right corner of the real box, Indicates the ordinate of the lower right corner of the real box.

[0043] GIoU Loss measures the position error between the predicted box and the true box by calculating the difference in the intersection-over-union ratio between them. The closer the value of this indicator is to 1, the higher the match between the predicted box and the true box; the farther the value is from 1, the lower the match, that is, the greater the error.

[0044] The image analysis and storage module generates a complete report for the processed and analyzed images and stores them.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention adopts the SK-YOLOv8 combined recognition model, which aims to introduce the SK mechanism into the C2f module in YOLOv8, mainly bringing the following improvements:

[0047] ① Enhanced feature expression: The SK mechanism can effectively enhance the model's ability to learn important features by adaptively selecting and combining different features. This enables the model to capture targets more accurately in complex scenarios.

[0048] ② Multi-scale feature fusion: The SK mechanism supports the effective fusion of multi-scale features, which improves the model's ability to detect objects of different sizes. This is very helpful for processing objects of diverse sizes.

[0049] ③ Reduce the number of parameters: By introducing the SK mechanism, the complexity and number of parameters of the model can be reduced while maintaining performance, thereby improving computing efficiency and adapting to a wider range of application scenarios.

[0050] ④ Improve robustness: The SK mechanism can improve the robustness of the model to noise and deformation, especially when dealing with scenes with complex backgrounds, the model's performance is more stable.

[0051] ⑤ Better generalization ability: After the introduction of the SK mechanism, the performance gap between the training data set and the test data set of YOLOv8 has narrowed, which has improved the generalization ability of the model and made it more adaptable in different domains.

[0052] With these improvements, YOLOv8 has shown higher accuracy and efficiency in target detection tasks, especially on more challenging datasets.

[0053] 2. The SK-YOLOv8 network uses three loss functions: BECLogits loss function, BCEclsloss loss function, and GIoU Loss, aiming to effectively optimize the target detection task and improve the detection performance. The combination of BECLogits, BCEclsLoss, and GIoU Loss in the SK-YOLOv8 network brings the following improvements:

[0054] ①BECLogits loss:

[0055] Better handling of imbalanced categories: BECLogits can effectively handle the class imbalance problem in target detection by combining binary cross entropy, thereby improving the performance of the model on minority class targets.

[0056] Improve classification accuracy: This loss design allows the model to more accurately determine the existence of the target and improve the overall classification performance.

[0057] ②BCEclsLoss loss:

[0058] Improved classification performance: BCEclsLoss (Binary Cross-Entropy Classification Loss) helps optimize the accuracy of multi-category target recognition, making the model more accurate in identifying each category.

[0059] Contribution of segmentation loss: Through independent category losses, the model's learning of each category can be further refined, especially in complex scenarios, the ability to distinguish similar objects is enhanced.

[0060] ③GIoU Loss:

[0061] Improve positioning accuracy: The GIoU (Generalized Intersection over Union) loss function not only focuses on the overlap between the predicted box and the true box, but also considers their relative position, which can more comprehensively evaluate the quality of the box and thus improve the accuracy of target positioning.

[0062] Solving the bounding box regression problem: GIoU can effectively solve the problem that traditional IoU is not effective when dealing with non-overlapping boxes, so that even when the initial overlap is small, the model can gradually learn accurate predictions.

[0063] ④Overall effect:

[0064] Comprehensive optimization: The combined use of these three loss functions can more comprehensively improve the classification and regression performance of the model, making SK-YOLOv8 show higher accuracy and robustness in target detection tasks.

[0065] Improve model stability: Diversified loss function designs for different problems make the model more stable during training and reduce the risk of overfitting.

[0066] Faster convergence speed: Due to the targeted optimization objectives of each loss function, the model can converge faster during training, improving training efficiency.

[0067] With these improvements, SK-YOLOv8 is able to provide better detection performance in both general and challenging scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flow chart of the intelligent analysis system of ice-covered images of distribution network lines based on deep learning of the present invention.

[0069] Figure 2 This is a schematic diagram of the SK-YOLOv8 recognition model network architecture.

[0070] Figure 3 It is a block diagram of the intelligent analysis system of ice-covered images of distribution network lines based on deep learning of the present invention. DETAILED DESCRIPTION

[0071] Those skilled in the art will appreciate that the following examples are only used to illustrate the present invention and should not be considered to limit the scope of the present invention. If no specific techniques or conditions are specified in the examples, the techniques or conditions described in the literature in the art or the product specifications are used. If the manufacturer of the product is not specified, it is a conventional product that can be purchased.

[0072] Embodiment: A distribution network line ice image intelligent analysis system based on deep learning, such as Figure 1 As shown, the system includes: an image data acquisition and image classification module, an image annotation module, a model network construction and model recognition module, and an image analysis and storage module; the image data acquisition and image classification module includes image acquisition and image classification, the image annotation module includes using tools to annotate images, the model network construction and model recognition module includes building an SK-YOLOv8 model, and using the model to perform target image recognition on the annotated images, and the image analysis and storage module includes analyzing the images after model recognition and generating reports for storage; the system front end is a user interaction module;

[0073] The image data acquisition module and the image classification module are used to obtain a large number of images related to ice coverage of distribution network lines and classify the images into positive samples and negative samples;

[0074] The image annotation module uses an image annotation tool (Label Img) to perform preset classification annotation on the distribution line ice-covered image;

[0075] The model network construction and model recognition module uses labeled data to train the SK-YOLOv8 recognition model for identifying the icing status of power lines.

[0076] The image analysis and storage module analyzes the image after the model recognition, generates a report, and stores the content;

[0077] The user interaction module includes user input and output. The user inputs an image, calls the repository to perform target image recognition, and provides feedback to the user. The image analysis module determines whether the pixel value deviation of each sampling area of ​​the input image meets the set consistency requirements based on the user's input image;

[0078] If yes, a report is generated and the report results are fed back to the user;

[0079] If not, the user is informed that there is no valid result for the input image, and the image is stored in another storage terminal.

[0080] If the image reaches the preset number of queries, the image is input into the system backend for training and learning.

[0081] Specifically:

[0082] After the acquisition module acquires the image, the image is sent to the image annotation module, and the steps include: in the image acquisition module, the distribution line ice-covered image is an image of the distribution line containing ice-covered distribution line, which can be acquired on the distribution line site through a mobile phone or other equipment or can be acquired remotely in an artificial climate laboratory;

[0083] After the distribution line ice-covered image is collected, the collected distribution line ice-covered image is subjected to pre-processing operations such as normalization, scaling, and cropping to obtain a distribution line ice-covered image with all pixels of 640×640;

[0084] In the data preprocessing stage, the distribution line ice-covered image is enhanced by at least one of translation, rotation, mirroring or scaling;

[0085] Icing targets include iced insulators or iced conductors. The main purpose is to identify iced insulators in iced distribution lines. The specific steps include:

[0086] The icing target is the iced insulator;

[0087] The ice-covered images of distribution lines containing insulators are positive samples;

[0088] The ice-covered distribution line images without insulators are negative samples.

[0089] Annotate the image in the image annotation module;

[0090] After obtaining the training data including positive samples and negative samples, use the image annotation tool (Label Img) to perform preset classification annotation on the distribution line ice-covered images:

[0091] The image annotation tool can be a QT5 graphical interface annotation tool written in Python (the Qt5 graphical interface annotation tool written in Python is a common implementation method in the prior art. Qt is a cross-platform graphical user interface development framework, and Python, as an efficient and easy-to-use programming language, combined with Qt5 can quickly develop powerful desktop applications).

[0092] Optionally, annotation may be performed based on the acquired training data, when the training data is only an image of ice-covered distribution lines;

[0093] Optionally, the distribution line ice-covered image may be labeled based on whether it contains insulators or conductors;

[0094] Obtaining separate category labels of the distribution line ice-covered image, wherein the separate category labels include two separate category labels of ice-covered insulators and ice-covered conductors;

[0095] When marking, it is also necessary to mark the area containing the ice-covered target in the distribution line ice-covered image;

[0096] The real frame coordinates of the ice-covered target are obtained so that the target candidate frame predicted later can be compared with the real frame.

[0097] After the data is pre-classified and labeled, a standard data set is obtained, and 70% of the samples are randomly selected from the standard data set as a training set, 20% of the samples are used as a test set, and 10% of the samples are used as a prediction set;

[0098] The steps of constructing the SK mechanism module include: SK is an attention mechanism for convolution kernels. According to the different sensitivity of targets of different sizes to convolution kernels of different sizes, SK dynamically generates different convolution kernels for different images and tries to find the convolution kernel of the most suitable size; therefore, SK has the characteristics of target scale adaptation and has better effect on multi-scale target detection than other attention mechanism modules. The processing of the SK module mainly includes the following three steps:

[0099] Split: By using two convolution kernels of different sizes to perform operations such as Group Convolution and ReLU activation function, and using 3x3 and 5x5 convolution kernels to process the input feature map at the same time, two different feature maps are generated to test the sensitivity of different convolution kernels to the target and improve the detection accuracy;

[0100] Fuse: The two different feature maps obtained above are fused by controlling the information flow through the gate. The results of the two branches are fused using element-wise summation, where Fgp is the global average pooling operation and Ffc is a two-layer fully connected layer that reduces and then increases the dimension.

[0101] Select: Use two different weight matrices to weight the results of the two branches, and then sum the weighted results to get the output vector. This module adjusts the receptive field through different convolutional layers to improve the detection ability and accuracy of multi-scale objects.

[0102] The specific steps of the YOLOv8 model network architecture include:

[0103] Input: Data enhancement and adaptive image scaling are performed. Mosaic enhancement is mainly used for data enhancement. The Mosaic enhancement part of YOLOv8 introduces the operation of using Mosaic enhancement in the last 10 epochs in YOLOX, which enhances the robustness of the model. Adaptive image scaling scales the original images to the same size, simplifies the model complexity, and can effectively improve the accuracy.

[0104] Backbone network: It is composed of CBS, C2f and SPPF modules. CBS mainly includes Conv, BN and SiLU activation functions. This module mainly performs convolution operations on feature maps. The new C2f module of the YOLO series is the main module for extracting features. SPPF (Spatial Pyramid Pooling Fast) is used to extract features of different receptive fields. This structure enables the network to better adapt to targets of different scales.

[0105] Feature Enhancement Module (Neck): The FPN-PAN structure is mainly composed of FPN and PAN. FPN is a top-down feature pyramid network structure that generates a prediction feature map by fusing high-level and low-level feature information through upsampling. PAN (Path Aggregation Network) is an improvement on FPN, which introduces the semantic information of the lateral connection enhancement feature, so that the bottom-up feature map can be fused with the top-down feature map;

[0106] Detection module (Detect): uses an anchor-free matching mechanism. There is no need to preset anchors. It only needs to regress the target center point and width and height of feature maps of different scales, which greatly reduces the calculation time. Finally, by utilizing the rich information from feature maps of different scales, it can accurately obtain the classification and location information of large, medium and small-sized target objects.

[0107] The specific steps of constructing the SK-YOLOv8 recognition model network architecture include:

[0108] The structure is still composed of four parts: input end, backbone network, feature enhancement module, and detection module. Among them, the input end and output end remain unchanged, and the SK mechanism is added after the C2f module on the backbone network and Neck network. Other processing units and network structures remain unchanged.

[0109] Therefore, the overall structure of the improved SK-YOLOv8 network is basically the same as the original YOLOv8 network. The improvement is the introduction of the SK mechanism after the C2f module, which improves the overall feature extraction and adaptability of the network. The SK-YOLOv8 network not only maintains the advantages of the YOLOv8 network in terms of lightness and fast calculation, but also has the multi-scale target adaptability that other one-stage target detection algorithms do not have.

[0110] The SK-YOLOv8 network uses different loss functions to calculate the target score, category probability score, and position loss. The specific loss functions include the BECLogits loss function for calculating the target score, the BCEclsloss loss function for calculating the category probability score, and the GIoU Loss for evaluating the position loss between the predicted box and the true box. The combination of these loss functions can effectively optimize the target detection task and improve the detection performance. For the loss of the target score, the BCEclsloss loss function is used to calculate the category probability score, and the GIoU Loss is also used to evaluate the position loss between the predicted box and the true box. The calculation formula of GIoU Loss is, let p1 and p2 be the coordinates of the upper left and lower right vertices of the predicted box, l1 and l2 are the coordinates of the upper left and lower right vertices of the true box, then:

[0111]

[0112] GIoU Loss = 1-GIoU

[0113] IoU is the ratio of the intersection and union areas of the predicted box and the true box, referred to as the intersection-union ratio. A is the predicted box, B is the true box, C is the minimum bounding box of A and B, S1 represents the area of ​​the overlapping area of ​​A and B, S2 represents the area after A and B are merged, S3 is the area of ​​the minimum rectangular box surrounding A and B, and GIoU Loss represents the measure of position loss. represents the prediction box boundary, Indicates the horizontal coordinate of the upper left corner of the prediction box, Indicates the ordinate of the upper left corner of the prediction box, Indicates the horizontal coordinate of the lower right corner of the prediction box, Indicates the ordinate of the lower right corner of the prediction box; represents the real box boundary, Indicates the horizontal coordinate of the upper left corner of the real box, Indicates the ordinate of the upper left corner of the real box, Indicates the horizontal coordinate of the lower right corner of the real box, Indicates the ordinate of the lower right corner of the real box.

[0114] GIoU Loss measures the position error between the predicted box and the true box by calculating the difference in the intersection-over-union ratio between them. The closer the value of this indicator is to 1, the higher the match between the predicted box and the true box; and the farther the value is from 1, the lower the match, that is, the greater the error. Compared with the traditional IoU, GIoU solves the problem of gradient calculation and gives a meaningful negative value when the predicted box does not overlap with the true box at all.

[0115] The image analysis and storage module generates a complete report for the processed and analyzed images. The specific steps are as follows:

[0116] The image analysis module is used to post-process images that have been trained, tested, and predicted by the model. The specific steps include:

[0117] Performing feature analysis and recognition on the image, that is, extracting and describing features in the image, such as color features, texture features, shape features, etc.;

[0118] Automatically identify and classify objects in the image, that is, detect and identify objects in the image, such as iced insulators, iced conductors, and un-iced conductors.

[0119] The storage module is used to efficiently manage and retrieve large amounts of image data; it includes image storage and management, such as the design and implementation of image databases, image retrieval and indexing technology, etc.

[0120] User interaction interface, where the user inputs an image, calls the repository to perform target image recognition, and provides feedback to the user;

[0121] According to the user's input and the image analysis module, it is determined whether the pixel value deviation of each sampling area of ​​the input image meets the set consistency requirement;

[0122] If yes, a report is generated and the report results are fed back to the user;

[0123] If not, the user is informed that there is no valid result for the input image, and the image is stored in another storage end;

[0124] If the image reaches the preset number of queries, the image is input into the system backend for training and learning.

[0125] The design of the user interaction interface includes the following specific steps:

[0126] Interaction design, including the design of the way users interact with the system, such as buttons, menus, forms, navigation, etc., to ensure that users can operate and navigate easily;

[0127] Visual design, including visual presentation of information and data, such as charts, maps, animations, etc., so that users can understand and analyze data more intuitively;

[0128] User feedback mechanism: This includes designing feedback mechanisms for user interaction with the system, such as prompts, error prompts, success prompts, etc., so that users can understand whether their operations are successful;

[0129] Multi-platform adaptation: Considering that users may use the system on different devices, it is necessary to design a user interface that is adapted to multiple platforms, such as PC, mobile, and tablet computers.

[0130] User experience design: involves considering the user's psychological and behavioral characteristics and designing the user interface to improve user satisfaction and user experience.

[0131] In addition, to achieve the above purpose, the present invention also proposes a distribution network line ice image intelligent analysis system based on deep learning, such as Figure 3 As shown, the system includes a front end and a back end. The system back end includes: an image data acquisition and image classification module, which includes image acquisition and image classification; an image annotation module includes using tools to annotate images; a model network construction and model recognition module includes building an SK-YOLOv8 model and using the model to perform target image recognition on annotated images; an image analysis and storage module includes analyzing the image after model recognition and generating a report for storage; the system front end interface is a user module, which includes user input and output.

[0132] 1. Image acquisition and classification module:

[0133] Input: distribution network line ice coverage image;

[0134] Annotation module: The annotation tool marks whether the distribution network line ice-covered image contains ice-covered targets;

[0135] Classification module: classify according to the annotation results;

[0136] Output: classified distribution network line ice coverage image.

[0137] 2. Model building and identification module:

[0138] Input: output of the first step;

[0139] Model construction: Establish SK-YOLOv8 recognition model and train the model;

[0140] Identification: Use the SK-YOLOv8 recognition model to identify ice-covered images of distribution network lines and test the model performance;

[0141] Output: loss function results, model performance data.

[0142] 3. Image analysis and storage module:

[0143] Input: The output of the second step and the corresponding image;

[0144] Analyze model results: analyze images and corresponding image recognition results;

[0145] Analyze user images: Judgment module: According to the user input and the image analysis and storage module, determine whether the user input contains the target image. If so, generate a corresponding report and feedback to the user; if not, inform the user that there is no valid result. At the same time, store the image, and determine whether to input the query image into the system for processing based on whether the input image reaches the preset query number, analyze the user input image and the corresponding image analysis results, and generate a report;

[0146] Storage: Store the results and the corresponding image recognition results;

[0147] Output: recognition results and image analysis report.

[0148] 4. User interaction interface:

[0149] Input: image input by the user;

[0150] Call the image analysis and storage module;

[0151] Output: The result after the third step analysis.

Claims

1. A distribution network line ice image intelligent analysis system based on deep learning, characterized by: The system backend includes: image data acquisition and image classification module, image annotation module, model network construction and model recognition module, image analysis and storage module; the image data acquisition and image classification module includes image acquisition and image classification, the image annotation module includes using tools to annotate images, the model network construction and model recognition module includes building an SK-YOLOv8 model, and using the model to identify the target image of the annotated image, and the image analysis and storage module includes analyzing the image after model recognition and generating a report for storage; the system frontend is a user interaction module; The image data acquisition module and the image classification module are used to obtain a large number of images related to ice coverage of distribution network lines and classify the images into positive samples and negative samples; The image annotation module uses an image annotation tool (Label Img) to perform preset classification annotation on the distribution line ice-covered image; The model network construction and model recognition module uses the labeled data to train the SK-YOLOv8 recognition model to identify the ice-covered state of the power line; The image analysis and storage module analyzes the image after the model recognition, generates a report, and stores the content; The user interaction module includes user input and output. The user inputs an image, calls the repository to perform target image recognition, and provides feedback to the user. The image analysis module determines whether the pixel value deviation of each sampling area of ​​the input image meets the set consistency requirements based on the user's input image; If yes, a report is generated and the report results are fed back to the user; If not, the user is informed that there is no valid result for the input image, and the image is stored in another storage end. If the image reaches the preset number of queries, the image is input into the system backend for training and learning.

2. The image processing method in the distribution network line ice-covered image intelligent analysis system based on deep learning according to claim 1 is characterized in that: include 1) The image classification module classifies images according to set rules and classifies images based on whether they contain target images; 2) A method in which the image annotation module annotates the data based on preset classification categories.

3. The image processing method in the distribution network line ice-covered image intelligent analysis system based on deep learning according to claim 2 is characterized in that: The method for the image annotation module to annotate the data based on the preset classification category comprises the following steps: (1) Mark the ice-covered distribution line images containing ice-covered targets as positive samples; (2) The ice-covered distribution line images that do not contain ice-covered targets in the image are marked as negative samples.

4. The image processing method in the distribution network line ice-covered image intelligent analysis system based on deep learning according to claim 2 is characterized in that: The method for the image annotation module to annotate the data based on a preset classification category includes: (1) acquiring image training data including positive samples and negative samples; (2) Use the image annotation tool (Label Img) to preset classification annotations for the distribution line ice-covered images.

5. The image processing method in the distribution network line ice-covered image intelligent analysis system based on deep learning according to claim 4 is characterized in that: The image training data is obtained by randomly selecting 70% of the samples as a training set, 20% of the samples as a test set, and 10% of the samples as a prediction set.

6. The method for constructing the SK-YOLOv8 model in the deep learning-based distribution network line ice image intelligent analysis system according to claim 1, characterized in that: include Methods for constructing SK mechanism modules and YOLOv8 model networks; SK-YOLOv8 recognition model network architecture construction method.

7. The method for constructing the SK-YOLOv8 model in the distribution network line ice image intelligent analysis system based on deep learning according to claim 6 is characterized in that: The steps of constructing the SK mechanism module include: The SK module is an attention mechanism based on the convolution kernel. It can adaptively adjust the size of the convolution kernel according to the size of the target to improve the perception ability. The processing of the SK module includes the following three steps: Split: By using two convolution kernels of different sizes for Group Convolution and ReLU activation functions, and using 3x3 and 5x5 convolution kernels to process the input feature map, two different feature maps are generated to test the sensitivity of different convolution kernels to the target and improve the detection accuracy; Fuse: According to the two different feature maps obtained above, the information flow is controlled by the gate to fuse, and the results of the two branches are fused using element-wise summation, where Fgp is the global average pooling operation and Ffc is a two-layer fully connected layer of dimensionality reduction and dimensionality increase; Select: Use two different weight matrices to weight the results of the two branches, and then sum the weighted results to get the output vector. This module adjusts the receptive field through different convolutional layers to improve the detection capability and accuracy of multi-scale targets.

8. The method for constructing the SK-YOLOv8 model in the distribution network line ice image intelligent analysis system based on deep learning according to claim 6 is characterized in that: The YOLOv8 model network construction consists of four parts: input end, backbone network, feature enhancement module, and detection module composition: Input: Data enhancement and adaptive image scaling are performed. Mosaic enhancement is mainly used for data enhancement. The Mosaic enhancement part of YOLOv8 introduces the operation of Mosaic enhancement in the last 10 epochs of YOLOX, which enhances the robustness of the model. Adaptive image scaling scales the original images to the same size. Backbone network: It consists of CBS, C2f and SPPF modules. The CBS module mainly includes Conv, BN and SiLU activation functions. This module mainly performs convolution operations on feature maps. The C2f module is the main module for extracting features. The SPPF (SpatialPyramid Pooling Fast) module is used to extract features of different receptive fields. This structure enables the network to better adapt to targets of different scales. Feature Enhancement Module (Neck): The feature enhancement module is mainly composed of FPN and PAN. FPN is a top-down feature pyramid network structure that generates a prediction feature map by fusing high-level and low-level feature information through upsampling. PAN (Path Aggregation Network) is an improvement on FPN. It introduces semantic information of lateral connection enhancement features, so that the bottom-up feature map can be fused with the top-down feature map. Detection module (Detect): Using the anchor-free matching mechanism, by regressing the target center point and width and height of feature maps of different scales, the calculation time is greatly reduced. Finally, by utilizing the rich information from feature maps of different scales, the classification and location information of large, medium and small-sized target objects can be accurately obtained.

9. The method for constructing the SK-YOLOv8 model in the distribution network line ice image intelligent analysis system based on deep learning according to claim 6, characterized in that: The SK-YOLOv8 recognition model network architecture construction includes: the structure is composed of four parts: input end, backbone network, feature enhancement module, and detection module. In the SK-YOLOv8 network, the SK mechanism is introduced into the C2f module on the backbone network and the Neck network for improvement, while keeping the overall network structure and processing unit unchanged. Such improvement improves the feature extraction capability and adaptability of the network, while retaining the advantages of the light weight and high efficiency of the YOLOv8 network. The SK-YOLOv8 network uses different loss functions to calculate the target score, category probability score and position loss. The specific loss functions include BECLogits, which is used to calculate the target score, BCEclsloss, which is used to calculate the category probability score, and GIoU Loss, which is used to evaluate the position loss between the predicted box and the true box. The combination of these loss functions can effectively optimize the target detection task and improve the detection performance. The loss of the target score uses the BCEclsloss loss function to calculate the category probability score, and GIoU Loss is used to evaluate the position loss between the predicted box and the true box. The calculation formula of GIoU Loss is: let p1 and p2 be the coordinates of the upper left and lower right vertices of the predicted box, and l1 and l2 be the coordinates of the upper left and lower right vertices of the real box, then: GIoU Loss = 1-GIoU IoU is the ratio of the intersection and union areas of the predicted box and the true box, referred to as the intersection-union ratio. A is the predicted box, B is the true box, C is the minimum bounding box of A and B, S1 represents the area of ​​the overlapping area of ​​A and B, S2 represents the area after A and B are merged, S3 is the area of ​​the minimum rectangular box surrounding A and B, and GIoU Loss represents the measure of position loss. represents the prediction box boundary, Indicates the horizontal coordinate of the upper left corner of the prediction box, Indicates the ordinate of the upper left corner of the prediction box, Indicates the horizontal coordinate of the lower right corner of the prediction box, Indicates the ordinate of the lower right corner of the prediction box; represents the real box boundary, Indicates the horizontal coordinate of the upper left corner of the real box, Indicates the ordinate of the upper left corner of the real box, Indicates the horizontal coordinate of the lower right corner of the real box, Indicates the ordinate of the lower right corner of the real box. GIoU Loss measures the position error between the predicted box and the true box by calculating the difference in the intersection-over-union ratio between them. The closer the value of this indicator is to 1, the higher the match between the predicted box and the true box; the farther the value is from 1, the lower the match, that is, the greater the error.

Images