Damaged dustproof net identification method and system, storage medium and computer equipment
Through the improved YOLOv8-seg model and edge detection technology, the problem of dustproof network identification under the wire is solved, and the accurate identification of multi-scale, multi-pose and overlapping dustproof networks is achieved, reducing the risk of safety accidents.
Patent Information
- Application Number
- CN202410931693.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-12
Smart Images

Figure CN118942025A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of power transmission channel monitoring, and in particular relates to a method, system, storage medium and computer equipment for identifying a damaged dustproof net. Background Art
[0002] In the field of power transmission channel monitoring, with the continuous advancement of artificial intelligence technology, power transmission visualization monitoring technology has become increasingly mature. At present, this technology has achieved satisfactory results in monitoring common categories such as construction machinery, fireworks, and hanging foreign objects, greatly reducing the workload of patrol personnel and meeting the needs of daily patrols. However, with the continuous advancement of technology, people's expectations for power transmission channel monitoring are also continuing to increase. In today's era where timeliness is crucial, the importance of "pre-prediction" is increasingly prominent in the three links of "pre-prediction", "in-process control" and "post-event remediation".
[0003] Foreign objects hanging on live overhead wires have become a serious hidden danger affecting the safe operation of power transmission lines. If these foreign objects are not removed in time, they may cause serious wear and tear on the wires. In particular, lightweight and flexible foreign objects such as dust screens and plastic films, once hanging on the wires and not far enough from the ground, will pose a threat of high-voltage electric shock to pedestrians below. In severe cases, these foreign objects may cause the line to short-circuit and trip, leading to large-scale power outages, and may even cause personal injury accidents caused by electric shock.
[0004] CN115240057A discloses a method for detecting images of overhead power transmission lines based on deep learning, which uses the target detection model DH-YOLOv5 to perform image detection on hanging foreign objects on the power transmission lines. However, this detection method does not take into account foreign objects such as dust nets under the power transmission lines, which are often the source of hanging foreign objects. Especially in windy weather, these foreign objects such as offline dust nets can be easily blown up by the wind and hung on the wires, thus posing a safety hazard. Therefore, accurate identification of foreign objects such as offline dust nets can detect hidden dangers that threaten power transmission lines in advance and prevent them before they happen.
[0005] The dust nets hanging on the wires show relative stability and consistency in size, shape and posture, which makes their identification process relatively easy. However, in sharp contrast, the dust nets under the wires are often affected by a variety of external environments such as wind and human activities, so they are easily folded, stretched or compressed, resulting in significant diversity and complexity in their size, shape and posture. This undoubtedly increases the difficulty of identification, making the accurate identification of the dust nets under the wires a more challenging task.
[0006] Current visualization monitoring technologies may encounter difficulties in identifying dust screens under conductors due to their multi-scale, multi-posture, overlapping, and possible damage. Specific problems include: (1) Accurate identification of dust screens: Existing monitoring technologies have difficulty accurately identifying dust screens, especially when they appear in different shapes and sizes. This requires monitoring systems to have higher flexibility and adaptability to capture the diversity and complexity of dust screens. (2) Multiple dust screen overlap and perspective differences: In actual scenarios, multiple dust screens may overlap, or the size of the dust screen in the image may vary due to different shooting perspectives. These problems pose challenges to existing monitoring technologies and may lead to a decrease in recognition accuracy. (3) Identification of damaged dust screens: Compared with intact dust screens, damaged dust screens have lower adhesion to the ground and are therefore more likely to be blown away by the wind and hang on the conductors. They are the focus of inspection personnel. For the identified dust screens, how to accurately determine whether they are intact or damaged is a technical challenge. It is difficult for existing technologies to accurately extract the shape or edge information of the dustproof net from a complex background, and thus it is impossible to perform accurate status analysis. Summary of the invention
[0007] The purpose of the present invention is to provide a method for identifying a damaged dust-proof net, aiming to solve the problems of multi-scale, multi-posture and overlap faced by existing visualization monitoring technologies when identifying dust-proof nets, improve the accurate recognition rate of dust-proof nets, effectively deal with the recognition difficulties caused by the overlap of multiple dust-proof nets and the difference in viewing angles, and accurately judge whether the identified dust-proof net is in a damaged state, thereby realizing accurate analysis of the state of the dust-proof net.
[0008] The embodiment of the present invention provides a method for identifying a damaged dust screen, comprising the following steps:
[0009] S1: Acquire patrol images under the power transmission channel;
[0010] S2: using the trained dust screen detection model to pre-label the inspection image, and obtain the transmission channel inspection image containing the dust screen target through manual review;
[0011] S3: using pixel-level annotation software, segmenting and annotating the dust-proof net target in the transmission channel inspection image containing the dust-proof net target, and dividing the annotated image data set into a training set, a validation set, and a test set according to a ratio;
[0012] S4: Taking the YOLOv8-seg model as the basic model, the deformable convolution DCNv3 is introduced into the feature extraction backbone network Backbone, and the WIoU Loss is introduced into the model to construct an improved YOLOv8-seg model;
[0013] S5: Use the data set divided in S3 to train the improved YOLOv8-seg model to obtain a trained power transmission dust protection network segmentation model;
[0014] S6: input the power transmission channel image to be identified into the trained power transmission dust-proof net segmentation model to obtain a segmentation mask of the dust-proof net, perform edge detection on the segmentation mask, and output the recognition result of the dust-proof net.
[0015] By means of the above technical solution, the present invention has at least the following advantages:
[0016] (1) The damaged dust screen identification method disclosed in the embodiment of the present invention takes the dust screen under the transmission line as the identification object, which can realize an early warning mechanism to prevent potential risks, thereby significantly reducing the incidence of safety accidents caused by foreign objects entangled under the line.
[0017] (2) Considering the characteristics of the dust net such as easy deformation and non-rigidity, the present invention selects the YOLOv8-seg model as the basic model. YOLOv8-seg introduces the Anchor-Free detection head. Unlike the traditional Anchor-Based method, the Anchor-Free detection head does not require a predefined anchor frame, but directly detects the key points of the target, such as the corners or center points of the rectangular bounding box, to achieve target positioning, so that it can more adaptively detect targets of different sizes and proportions, thereby improving the flexibility and accuracy of detection. This enables the model to better adapt to various target shapes and sizes, especially for dust nets that exhibit a variety of postures, the model shows excellent generalization capabilities.
[0018] (3) DCNv3 is the third version of the deep deformable convolutional network, which can adaptively adjust the shape of the convolution kernel, effectively capture the deformation information of the target, and enable the network to more flexibly adapt to the geometric transformation of the input data, such as scaling, rotation, and translation. The present invention adds the DCNv3 module in the feature extraction stage, which not only makes up for the shortcomings of conventional convolution in processing long-distance dependencies and adaptive spatial aggregation, but also inherits the inductive bias of convolution, so that the model can operate efficiently even with less training data and shorter training time. This method is particularly suitable for adapting to the multi-scale and multi-posture characteristics of dust-proof nets.
[0019] (4) Since the training data inevitably contains samples of low quality, geometric factors such as distance and aspect ratio will impose greater penalties on these samples, thereby affecting the generalization ability of the model. In order to solve this problem, the present invention introduces WIoU Loss. When the anchor box and the target box are highly overlapped, WIoU Loss can reduce the penalty of geometric factors, thereby improving the generalization ability of the model by reducing intervention in the training process. The introduction of WIoU Loss is precisely to deal with the situation where multiple dust nets overlap and the problem of inconsistent dust net sizes due to differences in viewing angles, thereby improving the recognition accuracy and generalization performance of the model.
[0020] (5) In the output processing stage, the present invention performs edge detection on the segmentation mask of the dustproof net. Through this step, it is possible to accurately determine whether the dustproof net is in a complete or damaged state, thereby providing a more accurate recognition result.
[0021] Preferably, the embodiment of the present invention further discloses that, in S1, the patrol images acquired under the transmission channel cover images of transmission channel scenes in different regions, different pole tower locations, different seasons and different times. The present invention collects images from different pole tower locations to cover diverse geographical environment scenes. In addition, the present invention also collects images according to different seasons and time nodes to further increase the diversity of image features. Through these means, the robustness and generalization of the algorithm model are enhanced.
[0022] Preferably, the embodiment of the present invention further discloses that the pixel-level annotation software is labelme. Labelme is an open source image annotation tool used to prepare data sets for machine learning or computer vision projects. Users can use labelme to annotate images in various forms such as polygons, rectangles, circles, line segments, and points, and generate a JSON file corresponding to the image, which contains detailed annotation information.
[0023] Preferably, the embodiment of the present invention further discloses that the division ratio of the training set, the validation set and the test set is 8:1:1.
[0024] Preferably, the embodiment of the present invention also discloses that, in S4, DCNv3 is used to replace the 5-layer conventional convolution part in the YOLOv8-seg feature extraction network. In this way, combined with CSPLayer (cross-stage local network layer) and SPPF (spatial pyramid pooling fast version) in YOLOv8-seg, a DCSNet (DCNv3+CSPLayer+SPPF) feature extraction network is constructed, which is designed to improve the accuracy and efficiency of target detection and segmentation.
[0025] Preferably, the embodiment of the present invention further discloses that, in S6, edge detection uses a Canny edge detector.
[0026] Preferably, the embodiment of the present invention further discloses that, in S6, the step of performing edge detection on the segmentation mask is:
[0027] S61, applying a Gaussian filter to smooth the dust screen mask output by the power transmission dust screen segmentation model;
[0028] S62, applying a Sobel operator to the image after Gaussian filtering to calculate the gradient intensity and direction of each pixel in the image;
[0029] S63, after obtaining the gradient image, performing non-maximum suppression processing;
[0030] S64, applying double threshold processing to the image after non-maximum suppression.
[0031] Based on the same inventive concept, an embodiment of the present invention further discloses a power transmission channel dustproof net segmentation system, comprising:
[0032] An image acquisition unit, used for acquiring patrol images under the power transmission channel;
[0033] An image pre-annotation unit, used for pre-annotating the patrol image;
[0034] A segmentation and annotation and data set division unit, used to segment and annotate the dust-proof net target in the transmission channel patrol image containing the dust-proof net target, and divide the annotated image data set into a training set, a verification set and a test set according to a ratio;
[0035] The model building unit uses the YOLOv8-seg model as the base model, introduces the deformable convolution DCNv3 into the feature extraction backbone network Backbone, and introduces WIoU Loss into the model, so as to build an improved YOLOv8-seg model;
[0036] A model training unit is used to train the improved YOLOv8-seg model using the data set to obtain a trained power transmission dust protection network segmentation model;
[0037] The post-processing optimization unit is used to receive the dust-proof net segmentation mask output by the power transmission dust-proof net segmentation model, perform edge detection on it, and finally output an accurate dust-proof net recognition result.
[0038] Based on the same inventive concept, an embodiment of the present invention further discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned method for identifying a damaged dustproof net is implemented.
[0039] Based on the same inventive concept, an embodiment of the present invention further discloses a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the above-mentioned method for identifying a damaged dustproof net. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 This is a flow chart of the method for identifying a damaged dust screen in Example 1.
[0041] Figure 2 Schematic diagram of the improved YOLOv8-seg model architecture in Example 1.
[0042] Figure 3 This is a schematic diagram of the Segment Head in the improved YOLOv8-seg in Example 1.
[0043] Figure 4 This is a schematic diagram of the Mask Head in the improved YOLOv8-seg in Example 1.
[0044] Figure 5 This is the original image of the dust screen obtained by the power transmission monitoring equipment in Example 1.
[0045] Figure 6 This is a diagram of the dust screen recognition result predicted by the model in Example 1. DETAILED DESCRIPTION
[0046] The embodiment of the present invention proposes a method for identifying a damaged dust net, which is used for identifying a damaged dust net under a specific power transmission channel. The method is based on an improved YOLOv8-seg model and aims to solve the multi-scale, multi-posture and overlapping problems in dust net identification. The method first uses a trained dust net detection model to pre-label the power transmission channel image, and then finely segment and label the dust net target after manual review. By introducing deformable convolution DCNv3 and WIoU Loss, an optimized YOLOv8-seg model is constructed to adapt to the diversity and complexity of dust nets. The model can flexibly detect targets of different sizes and shapes, and effectively deal with dust net overlap and perspective differences. Finally, the segmented dust net is edge detected to accurately determine its complete or damaged state. This method improves the recognition accuracy of the dust net, especially in multi-posture, multi-scale and overlapping situations, and provides strong support for the status monitoring of the dust net, so that potential safety hazards can be nipped in the bud. Specifically, when the dust screen is slightly damaged but has not yet developed into a high-risk hanging foreign object, this method can promptly and accurately identify the damaged dust screen, providing a reliable basis for the timely intervention and treatment of professional staff on site. This helps prevent possible serious and dangerous accidents in the future, thereby ensuring the continued stability of safe production.
[0047] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0048] like Figure 1 As shown, in Example 1, a method for identifying a damaged dust screen includes the following steps:
[0049] S1: Obtain patrol images under the power transmission channel.
[0050] The image acquisition device (hereinafter referred to as “image acquisition unit”) may be a visualization device installed on a pole tower, such as a remote monitoring camera.
[0051] In order to enhance the robustness and generalization of the algorithm model, in S1, the patrol images of the transmission channel obtained cover images of transmission channel scenes in different regions and at different tower locations, aiming to cover diverse geographical environment scenes. In addition, the patrol images of the transmission channel obtained also cover images of transmission channel scenes in different seasons and at different times, to further increase the diversity of image features.
[0052] S2: Use the trained dust net detection model to pre-label the inspection image, and select the correctly pre-labeled image through manual review to obtain the transmission channel inspection image containing the dust net target as the segmentation and labeling data.
[0053] Pre-labeling is achieved through the trained dust net detection model. In this embodiment, the detection model under the YOLOv8 architecture is used. After training on the labeled dust net detection data set, the dust net detection model is obtained, and then applied to the unlabeled data to automatically generate preliminary labeling results. The purpose of pre-labeling is to provide a basis for subsequent labeling work, thereby reducing the workload of manual labeling and improving labeling efficiency, while improving the accuracy and consistency of data labeling.
[0054] S3: Use pixel-level annotation software to segment and annotate the dustproof net targets in the transmission channel patrol image containing the dustproof net targets, label them as dustproof, and obtain a labeled dustproof net dataset; and divide the dataset according to the ratio of training set: validation set: test set = 8:1:1 to obtain the DustproofSeg-Dataset-5K+ dataset.
[0055] In this embodiment, the pixel-level annotation software is labelme. Labelme is an open source image annotation tool used to prepare data sets for machine learning or computer vision projects. Users can use labelme to annotate images in various forms such as polygons, rectangles, circles, line segments, and points, and generate a JSON file corresponding to the image, which contains detailed annotation information.
[0056] S4: Figure 2 As shown in the figure, the YOLOv8-seg model is used as the basic model, the deformable convolution DCNv3 is introduced into the feature extraction backbone network Backbone, and the WIoU Loss is introduced into the model to construct an improved YOLOv8-seg model.
[0057] Specifically, considering that the dust net is easy to deform and non-rigid, this embodiment selects the YOLOv8-seg model as the basic model. YOLOv8-seg introduces the Anchor-Free detection head. Unlike the traditional Anchor-Based method, the Anchor-Free detection head does not require a predefined anchor frame, but directly detects the key points of the target, such as the corners or center points of the rectangular bounding box, to achieve target positioning, so that it can more adaptively detect targets of different sizes and proportions, improving the flexibility and accuracy of detection. This enables the model to better adapt to various target shapes and sizes, especially for dust nets that exhibit a variety of postures, the model shows excellent generalization capabilities.
[0058] DCNv3 is the third version of the deep deformable convolutional network, which can adaptively adjust the shape of the convolution kernel, effectively capture the deformation information of the target, and enable the network to more flexibly adapt to the geometric transformation of the input data, such as scaling, rotation, and translation. This embodiment adds the DCNv3 module in the feature extraction stage, which not only makes up for the shortcomings of conventional convolution in processing long-distance dependencies and adaptive spatial aggregation, but also inherits the inductive bias of convolution, so that the model can operate efficiently even with less training data and shorter training time. This method is particularly suitable for adapting to the multi-scale and multi-posture characteristics of dust-proof nets.
[0059] like Figure 2 As shown in the figure, DCNv3 is used to replace the 5-layer conventional convolution part in the YOLOv8-seg feature extraction network. In this way, combined with CSPLayer (cross-stage local network layer) and SPPF (spatial pyramid pooling fast version) in YOLOv8-seg, a feature extraction network DCSNet (DCNv3+CSPLayer+SPPF) is constructed. This network is designed to improve the accuracy and efficiency of target detection and segmentation.
[0060] The convolution of DCNv3 is:
[0061]
[0062] Among them, y(p0) represents the output feature after the improved deformable convolution sampling; G represents the number of grouping aggregations; K represents the number of sampling points; for the g-th group, w g represents the position-independent mapping weight of the group; m gk represents the modulation scalar obtained by normalizing the k-th sampling point of the g-th group with the softmax function along dimension k; x g represents the input feature map of the slice; p0 and p k Represent the current and kth sampling points respectively; Δp gk is the sampling position p of the g-th group of grids k The corresponding offset.
[0063] In specific design, such as Figure 2As shown in the figure, in the structure of the backbone feature extraction unit (Backbone), the original image data first passes through the first DCNv3 layer. DCNv3 can dynamically adjust the convolution kernel according to the shape and position of the input features adaptively, so as to better capture the spatial relationship between features. This is achieved through an offset prediction network and a deformable convolution operation. The offset prediction network predicts the offset of each input position, and then the deformable convolution operation convolves the input features according to these offsets. Next, the data passes through four alternating DCNv3 and CSPLayer 2Conv layers in sequence, and each such alternation further refines and strengthens the feature representation. Among them, CSPLayer 2Conv enhances the network's learning ability and feature extraction efficiency by introducing cross-stage partial connections, while reducing the number of parameters and the amount of computation. After four rounds of DCNv3 and CSPLayer 2Conv processing, the data arrives at SPPF. At this layer, the feature map undergoes spatial pyramid pooling or similar operations to fuse features of different scales to obtain a more comprehensive feature representation.
[0064] The neck unit plays a vital role in the target detection architecture. It is mainly responsible for fusing the multi-scale feature maps from the backbone. By combining the bottom-up and top-down paths, the neck structure can effectively integrate high- and low-level features, thereby significantly enhancing the performance of target detection, especially for the recognition of small-sized targets.
[0065] In the bottom-up process, Neck first processes the lowest resolution feature map output by Backbone through a specially designed SPPF layer (which may be a specific feature aggregation layer) to generate a 20×20 feature map P1. Subsequently, the size of P1 is enlarged to 40×40 by upsampling, and concat (feature concatenation) is performed with the feature map of the same size output by the 6th layer CSPLayer2Conv in Backbone to generate a new 40×40 feature map P2. Then, P2 is upsampled to 80×80 and concat with the feature map output by the 4th layer CSPLayer 2Conv in Backbone, finally generating an 80×80 feature map N1.
[0066] In the top-down process, Neck adopts a similar strategy. Starting from the highest resolution feature map, it is first downsampled through a convolutional layer to match its size with P2, and then concat with P2 to generate a new 40×40 feature map N2. Next, N2 is convoluted downsampled to make its size consistent with P1, and concat with P1 to finally generate a 20×20 feature map N3.
[0067] After such multi-scale feature fusion, Neck finally outputs three feature maps N1, N2 and N3, which correspond to different scales and receptive fields, so as to more comprehensively cover objects of different sizes. These fused feature maps will be sent to the head unit for subsequent object classification and positioning.
[0068] In the structural design of the head unit (Head), its core function is to analyze the feature maps of three different scales (N1, N2, N3). Figure 2 As shown in Figure 2, there are two parallel and independent branches on the top feature map N1. One of them is the Mask branch (such as Figure 4 As shown in Figure 2, this branch is responsible for generating a set of mask prototypes, and the number is usually set to 32 in engineering practice. The Mask branch receives input from the top-level feature map N1. The input feature map is extracted and transformed through a series of convolutional layers to generate high-quality mask prototypes. These mask prototypes represent different mask information learned by the network, and they are combined in a linear superposition manner to form the final mask representation for the current target.
[0069] Another branch is the Segment branch (such as Figure 3 As shown in the figure, similar to N1, N2 and N3 feature maps also adopt this structure. The Segment branch receives input from the feature map, then extracts features through a series of convolution operations, and finally divides into three parallel output paths, which are used to output category information, prediction box coordinates, and Mask coefficients (Mask coef). These output data are processed later to form a refined segmentation result. The whole process demonstrates the comprehensive capabilities of the head unit in feature parsing, mask generation, and category prediction, providing strong support for image segmentation tasks.
[0070] In practical applications, the collected training data often contains some low-quality samples, which may not be accurate or representative enough due to various reasons (such as incorrect labeling, poor image quality, etc.). When performing target detection, the model needs to consider the position, size, and shape of the object. Among them, distance (that is, the distance between the anchor box and the target box) and aspect ratio (that is, the ratio of the width to the height of the object) are two important geometric factors. If these geometric factors in the training data vary greatly, or the sample quality is not high, the model may suffer a large "penalty" during the learning process, that is, the performance of the model will be greatly affected, thereby reducing its ability to generalize to unseen data. In order to solve this problem, this embodiment introduces WIoU Loss, such as Figure 3 As shown in the figure. When the anchor box and the target box are highly overlapped, WIoULoss can reduce the penalty of geometric factors, thereby improving the generalization ability of the model by reducing intervention in the training process. The introduction of WIoU Loss is precisely to deal with the situation where multiple dust screens overlap and the problem of inconsistent dust screen sizes due to differences in viewing angles, thereby improving the recognition accuracy and generalization performance of the model.
[0071] The WIoU Loss is:
[0072] LwIoUv1=RRywIouLIoU
[0073]
[0074] Among them, W g Indicates the width of the minimum outer bounding box, H g represents the height of the minimum outer bounding box; x represents the width of the anchor box, and y represents the height of the anchor box; gt Indicates the width of the target box, y gt Indicates the height of the target box. In order to prevent R WIoU Produces gradients that hinder convergence, W g and H g Separated from the calculation graph, this operation is indicated in the formula by the superscript * in the denominator.
[0075] R WIoU ∈[1,e) is used to significantly enlarge the L of the normal quality anchor box IoU (Weight difference). L IoU ∈[0,1] is used to significantly reduce the R of the high-quality anchor box WIoU (false detection rate), and when the anchor box coincides with the target box, focus on the distance between the center points.
[0076] S5: Use the DustproofSeg-Dataset-5K+ dataset to train the improved YOLOv8-seg model to obtain the trained power transmission dust protection network segmentation model Shudian_DustproofSeg.
[0077] In S5, the improved YOLOv8-seg model is trained, and the specific experimental results are as follows:
[0078] The following table shows the results of the ablation experiment on the test set, which is used to show the impact of different models on the improved performance. The size of the input image is 1280*1280. The mIoU of the basic YOLOv8x-seg model is 82.3% and the ACC is 89.5%; after adding DCNv3, compared with the basic YOLOv8x-seg model, the mIoU is increased by 3.9% and the ACC is increased by 4.6%; after adding WIoULoss, compared with the basic YOLOv8x-seg model, the mIoU is increased by 3.2% and the ACC is increased by 6.1%; after adding DCNv3 and WIoU Loss at the same time, compared with the basic YOLOv8x-seg model, the mIoU is increased by 6.4% and the ACC is increased by 7.2%.
[0079]
[0080]
[0081] in:
[0082] mIoU (mean intersection over union): Perform inference calculations on each category data set separately, divide the calculated intersection of the predicted area and the actual area by the union of the predicted area and the actual area, and finally take the average of the results obtained for all categories.
[0083] Acc (accuracy): refers to the ratio of pixels with correct category prediction to total pixels. The higher the accuracy, the better the model quality.
[0084] S6: Figure 5 As shown in FIG, the transmission channel image to be identified is input into the trained transmission dust screen segmentation model to obtain the segmentation mask of the dust screen, and the segmentation mask is post-processed and optimized to output the recognition result of the dust screen. The output result is shown in FIG. Figure 6 As shown, the red one is a damaged dust screen and the green one is a complete dust screen.
[0085] The segmentation mask is post-processed and optimized, which may be shape analysis or edge detection. In this embodiment, edge detection is performed on the segmentation mask to accurately determine whether the dust screen is in a complete or damaged state, thereby providing a more accurate recognition result. Specifically, if an obvious break or gap is detected, it means that the dust screen may be incomplete.
[0086] In S6, the step of performing edge detection on the segmentation mask is:
[0087] S61, apply a Gaussian filter to smooth the dust screen mask output by the power transmission dust screen segmentation model. The Gaussian filter can be implemented by a convolution operation, and its kernel function is a two-dimensional Gaussian distribution function. The purpose of the Gaussian filter is to reduce noise and unnecessary details in the image and provide a clear image basis for subsequent edge detection.
[0088] S62, applying the Sobel operator to the image after Gaussian filtering to calculate the gradient strength and direction of each pixel in the image. The Sobel operator is a commonly used edge detection operator, which can be implemented by convolution operation. The Sobel operator calculates the gradient strength and direction of each pixel in the image, which helps to determine the location of potential edges.
[0089] S63, after obtaining the gradient image, non-maximum suppression processing is performed. The purpose of this step is to retain only the local maximum point in the gradient direction, thereby refining the edge, removing unnecessary edge responses, and making the edge clearer.
[0090] S64, applying double threshold processing to the image after non-maximum suppression. Setting two high and low thresholds to distinguish strong edges from weak edges. Using hysteresis threshold method or other similar methods to determine the final edge position, usually strong edges are directly accepted, while weak edges are accepted based on whether they are connected to strong edges.
[0091] The inspection personnel focus on the images of damaged dust-proof nets in the output dust-proof net recognition results to prevent accidents that threaten power transmission lines.
[0092] Those skilled in the art can understand that the method for identifying a damaged dust screen provided in the embodiment of the present invention is also applicable to identifying a plastic film under a power transmission line.
[0093] This embodiment also discloses a power transmission channel dust screen segmentation system, including:
[0094] An image acquisition unit, used for acquiring patrol images under the power transmission channel;
[0095] An image pre-annotation unit, used for pre-annotating the patrol image;
[0096] A segmentation and annotation and data set division unit, used to segment and annotate the dust-proof net target in the transmission channel patrol image containing the dust-proof net target, and divide the annotated image data set into a training set, a verification set and a test set according to a ratio;
[0097] The model building unit uses the YOLOv8-seg model as the base model, introduces the deformable convolution DCNv3 into the feature extraction backbone network Backbone, and introduces WIoU Loss into the model, so as to build an improved YOLOv8-seg model;
[0098] A model training unit is used to train the improved YOLOv8-seg model using the data set to obtain a trained power transmission dust protection network segmentation model;
[0099] The post-processing optimization unit is used to receive the dust-proof net segmentation mask output by the power transmission dust-proof net segmentation model, perform edge detection on it, and finally output an accurate dust-proof net recognition result.
[0100] This embodiment further discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned method for identifying a damaged dustproof net is implemented.
[0101] This embodiment further discloses a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the above-mentioned method for identifying a damaged dustproof net.
[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for identifying a damaged dust screen, characterized in that: The following steps are involved: S1: Acquire patrol images under the power transmission channel; S2: using the trained dust screen detection model to pre-label the inspection image, and obtain the transmission channel inspection image containing the dust screen target through manual review; S3: using pixel-level annotation software, segmenting and annotating the dust-proof net target in the transmission channel inspection image containing the dust-proof net target, and dividing the annotated image data set into a training set, a validation set, and a test set according to a ratio; S4: Taking the YOLOv8-seg model as the basic model, the deformable convolution DCNv3 is introduced into the feature extraction backbone network Backbone, and the WIoU Loss is introduced into the model to construct an improved YOLOv8-seg model; S5: Use the data set divided in S3 to train the improved YOLOv8-seg model to obtain a trained power transmission dust protection network segmentation model; S6: input the power transmission channel image to be identified into the trained power transmission dust-proof net segmentation model to obtain a segmentation mask of the dust-proof net, perform edge detection on the segmentation mask, and output the recognition result of the dust-proof net.
2. The method for identifying a damaged dust screen according to claim 1, characterized in that: In S1, the patrol images acquired under the transmission channel cover images of transmission channel scenes in different regions, different tower locations, different seasons and different times.
3. The method for identifying a damaged dust screen according to claim 1, characterized in that: The pixel-level labeling software is labelme.
4. The method for identifying a damaged dust screen according to claim 1, characterized in that: In S4, DCNv3 is used to replace the 5-layer conventional convolution part in the YOLOv8-seg feature extraction network.
5. The method for identifying a damaged dust screen according to claim 1, characterized in that: In S6, the Canny edge detector is used for edge detection.
6. The method for identifying a damaged dust screen under a specific power transmission channel according to claim 5, characterized in that: In S6, the step of performing edge detection on the segmentation mask is: S61, applying a Gaussian filter to smooth the dust screen mask output by the power transmission dust screen segmentation model; S62, applying a Sobel operator to the image after Gaussian filtering to calculate the gradient intensity and direction of each pixel in the image; S63, after obtaining the gradient image, performing non-maximum suppression processing; S64, applying double threshold processing to the image after non-maximum suppression.
7. A dust-proof screen segmentation system for a power transmission channel, characterized in that: include: An image acquisition unit, used for acquiring patrol images under the power transmission channel; An image pre-annotation unit, used for pre-annotating the patrol image; A segmentation and annotation and data set division unit, used to segment and annotate the dust-proof net target in the transmission channel patrol image containing the dust-proof net target, and divide the annotated image data set into a training set, a verification set and a test set according to a ratio; The model building unit uses the YOLOv8-seg model as the base model, introduces the deformable convolution DCNv3 into the feature extraction backbone network Backbone, and introduces WIoU Loss into the model, so as to build an improved YOLOv8-seg model; A model training unit is used to train the improved YOLOv8-seg model using the data set to obtain a trained power transmission dust protection network segmentation model; The post-processing optimization unit is used to receive the dust-proof net segmentation mask output by the power transmission dust-proof net segmentation model, perform edge detection on it, and finally output an accurate dust-proof net recognition result.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the damaged dustproof screen identification method according to any one of claims 1 to 6 is implemented.
9. A computer device, characterized in that: The invention comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the method for identifying a damaged dustproof net according to any one of claims 1 to 6.
Citation Information
Patent Citations
Construction site safety protection net detection method and system
CN114758260A
Unmanned aerial vehicle degraded insulator detection method and system based on target identification
CN116665080A
Pipeline defect detecting and tracking method and device
CN116703826A
Lithium battery surface defect detection method based on fusion target detection technology
CN116740048A
Marine ship identification method based on deep learning
CN118053102A
Cited By
Power transmission channel hidden danger target detection method and device, electronic equipment and storage medium
CN121190742A
