Detection model, method, equipment and storage medium for self-explosion defects of aerial insulators

Abstract

The present invention discloses a detection model, method, device and storage medium for aerial insulator self-explosion defects. The detection model includes a backbone network D-Darknet53, a feature mining module and a four-scale prediction layer; the backbone network D-Darknet53 is used to extract image feature information of different scales and different levels of insulator defect images; the feature mining module is used to fuse image feature information of different scales and different levels of insulator defect images to obtain an insulator defect feature image with rich feature information; the four-scale prediction layer is used to predict the four-scale insulator defect feature images after feature fusion to obtain the best prediction frame of the insulator defect. The present invention can accurately identify and locate the self-explosion defects of aerial insulator strings in overhead transmission lines, and can solve the problems of slow detection speed and low detection accuracy of insulator self-explosion defects using deep learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and relates to a detection model, method, device and storage medium for self-explosion defects of insulators in aerial photography. Background Art

[0002] With the rapid development of computer, automation and other technologies, smart grid is an inevitable trend in the development of power grid technology. As an important part of the power grid, the safe and stable operation of transmission lines plays an important role in the stable development of the national economy. Insulators are one of the common electrical equipment in transmission lines. They are in a relatively harsh natural environment for a long time and are prone to self-explosion, rust, dirt and other problems. Therefore, regular inspection of insulators is of great significance to the normal operation of the power system.

[0003] The traditional manual insulator inspection method has high risk factor and low efficiency. With the rapid development of deep learning in the computer field in recent years, the inspection method using deep learning algorithms and drones has become a mainstream trend. At present, the insulator self-explosion defect detection methods studied by scholars are mainly divided into two categories, one is the cascade network structure detection method, and the other is the single-stage network structure detection.

[0004] In the study of cascade network structure detection methods, Tao et al. proposed a new deep convolutional neural network cascade structure for insulator positioning and defect detection, transforming defect detection into a two-level target detection problem. At the same time, a method of mixing multiple data enhancement algorithms was used to address the problem of too few data sets. This method has high detection accuracy, but slow detection speed, which is difficult to meet the needs of practical applications. For details, see "Tao X, Zhang D, Wang Z, et al. Detection of power line insulator defects using aerial images analyzed with convolutional neural networks[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2018, 50(4): 1486-1498"; in the study of single-stage network structure detection methods, Gao et al. improved the training strategy of YOLOv4, used multi-stage transfer learning and cosine annealing learning rate decay method to improve the network detection performance, and combined with small target image super-resolution to generate a network to improve the network's ability to recognize small targets. Although this method has a fast detection speed, it has poor detection effect when facing insulator images with complex backgrounds and small defects. For details, see "Gao Wei, Zhou Chen, Guo Mou FA Research on insulator defect identification based on improved yolov4 and sr-gan[J] Journal of motor and control, 2021, 25(11):93-104".

[0005] From the above, we can see that the insulator defect detection methods currently studied by scholars still have shortcomings in detection accuracy or detection speed in the face of complex backgrounds and small defects, making it difficult to meet the requirements of actual drone inspections. With the continuous development of embedded device technology, the deployment of target detection algorithms in practical applications has attracted more and more attention, and algorithm lightweighting is the key to the deployment of target detection algorithms. How to effectively reduce the number of model parameters while ensuring that the network can accurately and quickly detect insulator defects with complex backgrounds and small targets is a problem that needs to be solved urgently. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a detection model for aerial insulator self-explosion defects, which can accurately identify and locate the self-explosion defects of aerial insulator strings in overhead transmission lines, and improve the detection speed and detection accuracy of insulator self-explosion defects by deep learning technology.

[0007] The second object of the present invention is to provide a method for detecting self-explosion defects of insulators by aerial photography.

[0008] A third object of the present invention is to provide an electronic device.

[0009] A fourth object of the present invention is to provide a computer storage medium.

[0010] The technical solution adopted by the present invention is a detection model for the self-explosion defect of an aerial insulator, comprising:

[0011] The backbone network D-Darknet53 is used to extract image feature information of different scales and levels of insulator defect images;

[0012] A feature mining module is used to fuse the local features of the insulator defect image with the overall feature information to obtain an insulator defect feature image with rich feature information;

[0013] The four-scale prediction layer is used to predict the four-scale insulator defect feature images after fusing feature information of different scales and levels to obtain the best prediction frame of the insulator defect.

[0014] Furthermore, the backbone network D-Darknet53 consists of 5 residual modules Resdx (x=1, 2, 8, 8, 4) and downsampling convolutions. Each residual module is preceded by a downsampling convolution. The input insulator defect image is resized to 608×608 by 3×3 convolution, and then passes through downsampling convolution and residual modules Resdx (x=1, 2, 8, 8, 4) in sequence.

[0015] The residual module Resdx is used to extract the texture and edge information of the object in the insulator defect image under different receptive fields, obtain feature maps of different scales, and use deep separable convolution to reduce the number of parameters of the network model;

[0016] The residual module Resdx is composed of x residual units Resdunit;

[0017] The residual unit adopts depthwise separable convolution, which is composed of depthwise convolution with a convolution kernel size of 3×3 and point convolution of 1×1;

[0018] The depthwise separable convolution is to divide the convolution into two parts: depthwise convolution and point convolution. The depthwise convolution takes each channel as a group, and the convolution kernel performs convolution on each channel respectively, and then performs 1×1 point convolution on the output feature map. The point convolution is the ordinary convolution with a convolution kernel size of 1×1. The ratio of the depthwise separable convolution to the conventional convolution parameter is:

[0019]

[0020] Among them, D K ×D K ×1×M is the number of parameters for the depthwise convolution, D K ×D K ×1 is the convolution kernel size of the depth convolution, M is the number of convolution kernels, 1×1×M×N is the parameter amount of point convolution, 1×1×M is the convolution kernel size of the point convolution, D K ×D K ×M×N is the number of parameters of ordinary convolution, D K ×D K ×M is the size of the convolution kernel of ordinary convolution, and N is the number of convolution kernels of ordinary convolution;

[0021] The kernel size of the depthwise separable convolution is 3×3, and the value of N is much larger than D. K , so the number of parameters of depth-separable convolution is about 1 / 9 of that of conventional convolution;

[0022] The residual unit input end and the output end of the point convolution are added to each other to obtain the output of the residual unit;

[0023] The outputs of the last four residual modules Resdx (x=2, 8, 8, 4) of the five residual modules are L 1 , L 2 , L 3 , L 4 , the feature map size of the previous layer is twice that of the next layer;

[0024] The output end of the backbone network D-Darknet53 is connected to the input end of the feature mining module.

[0025] Furthermore, the feature mining module is composed of three convolutional layers, dense spatial pyramid pooling Dense-SPP and three convolutional layers in sequence;

[0026] The convolution layer is used to perform a convolution operation on the output feature map of the backbone network to improve the feature expression capability of the insulator defect feature map;

[0027] The dense spatial pyramid pooling Dense-SPP is used to perform pooling operations of different scales on the insulator defect feature map output by the convolution layer, strengthen the fusion of local features and overall features, and obtain an insulator defect feature map with rich feature information;

[0028] The input of the feature mining module is the output L of the backbone network D-Darknet 4 , L 4After three convolutions with kernel sizes of 1×1, 3×3, and 1×1, the size of the feature map is not changed, and the features are extracted to obtain the insulator defect feature map L 5 ; L 5 As the input of dense spatial pyramid pooling Dense-spp, L 5 Four parallel branches are sent in, one of which is a skip connection, and the other three are composed of cascade, maximum pooling layer and 1×1 convolution layer. The sizes of the three pooling kernels are 5×5, 9×9, and 13×13 respectively; finally, the outputs of the four branches are spliced ​​to obtain the feature L mined by Dense-SPP 6 ; L 6 Then, after three consecutive convolutions with kernel sizes of 1×1, 3×3, and 1×1, feature L is obtained. 7 , which is the final output of the feature mining module.

[0029] Furthermore, the four-scale prediction layer is composed of four prediction feature layers, and the last three feature layers are upsampled and concatenated to obtain a four-scale prediction feature map;

[0030] The upsampling operation inserts new elements between pixels using a suitable interpolation algorithm based on the original insulator defect feature map, expands the original feature map by 2 times, and then performs a splicing operation with the feature map of the previous layer;

[0031] The scales of the four-scale prediction layers are 19*19, 38*38, 76*76, and 152*152, respectively. The receptive field of the small-scale prediction layer is larger and is used to detect large targets, while the receptive field of the large-scale prediction layer is smaller and is used to detect small targets.

[0032] The input of the four-scale prediction layer is the output L of the feature mining module. 7 The output L of the three residual modules of the backbone network i (i=1,2,3), the inputs of the 152*152, 76*76, and 38*38 branches are respectively obtained by the previous layer after 1×1 convolution and upsampling operations and L i (i=1,2,3) splicing, and then 5 convolution feature extractions to get the output A i (i=1,2,3), A i (i=1,2,3,4) After two convolution operations with kernel sizes of 3×3 and 1×1 respectively, we finally get four prediction feature maps B with different scales. i Output L of (i=1,2,3,4) 8 ;

[0033] The A 4 =L 7 ;

[0034] The predicted feature layer B with a scale of 152×152 4 , is to output the feature network A 3 After 2 times of upsampling operation, the residual module Resd2 of the backbone network outputs L 1 After concatenation and convolution, the predicted feature map B is obtained. 4 The shallow feature L 1 The location, details and deep features of A 3 The semantic information fusion of the network reduces the loss of feature information at the insulator defect during the feature extraction process, thereby improving the accuracy of network detection of insulator defects.

[0035] A method for detecting self-explosion defects of insulators by aerial photography, wherein the detection process is performed according to the following steps:

[0036] Step S1, collecting insulator defect images and constructing an insulator defect data set;

[0037] Step S2: training the MDD-YOLOv3 network to obtain an insulator defect detection model;

[0038] Step S3: Using a drone to take aerial photos and collect images to test the insulator defect detection model;

[0039] Step S4: output the detection result of the insulator defect detection model to identify and locate the insulator defect.

[0040] Furthermore, in step S1, an insulator defect data set is constructed, and the collected insulator defect images are annotated and the data is expanded;

[0041] The insulator defect image labeling uses the labeling tool LabelImg to label the insulators and insulator defects in the image, with the labeling categories being insulator and defect, and an xml file with corresponding labeling location information and category is generated for each image;

[0042] The insulator defect data set is expanded, the number of collected pictures is small, and the network training is prone to overfitting. The original pictures are flipped, randomly cropped, rotated, translated, noise-perturbed, and the brightness contrast is transformed. The expanded data set is divided into a training set and a test set in a ratio of 8:2.

[0043] Furthermore, the process of training the MDD-YOLOv3 network in step S2 is as follows:

[0044] Step S2.1: Input the insulator defect image to be detected into the backbone network D-Darknet53 to extract the feature output L 4 ;

[0045] Step S2.2, L 4As the input of the feature mining module, after three convolutional feature layers, the extracted feature L 5 Input the Dense-SPP module to obtain the feature L mined by Dense-SPP 6 , L 6 After three convolutional feature layers connected in sequence, feature L is obtained 7 , which is the final output of the feature mining module;

[0046] Step S2.3: Output L of feature mining module 7 The output L of the three residual modules of the backbone network i (i=1,2,3) as the input of the four-scale prediction layer; the three large-scale branches are respectively obtained by convolution upsampling operation and L i After the concatenation operation of (i=1,2,3), the output A is obtained through 5 convolution feature extractions. i (i=1,2,3);

[0047] Step S2.4, A i (i=1,2,3,4) After two convolution operations, four prediction feature maps B of different scales are obtained. i (i=1,2,3,4);

[0048] Step S2.5: B i The set of (i=1,2,3,4) is denoted as L 8 , for L 8 The non-maximum suppression algorithm NMS is used to remove redundant frames to obtain the insulator self-explosion defect detection results.

[0049] The specific process of step S2.5 is as follows:

[0050] First, the prediction feature layer outputs four prediction feature maps of scales 19×19, 38×38, 76×76, and 152×152;

[0051] Secondly, before training, the k-means clustering algorithm is used to cluster four types of prior frames, which are divided into four groups from small to large, and four prediction feature maps of different scales are assigned. Each prediction feature map is divided into N×N units. Three bounding boxes are predicted in each unit of each feature map. Each bounding box predicts the confidence C of the existence of the target, the coordinates of the target center and the width and height of the box, as well as the probability intersection of each target category.

[0052] The confidence calculation formula is:

[0053]

[0054] Where P r(Object) indicates whether the cell contains an object. If the target center point is in the cell, it is 1, otherwise it is 0. Indicates the ratio of the predicted box to the true box, which is used to indicate the degree of overlap between the predicted box and the true box;

[0055] Finally, if the cell contains an object, the category confidence is obtained by multiplying the probability of the target category by the confidence, selecting the prediction box with the highest category confidence and filtering it using non-maximum suppression (NMS), and then converting the center point bias value into the midpoint coordinates of the bounding box through decoding to obtain the final prediction box position and category. The calculation formula is as follows:

[0056] b x =σ(t x )+c x

[0057] b y =σ(t y )+c y

[0058]

[0059]

[0060] where b x 、b y 、b w 、b h are the center point coordinates and width and height of the prediction box respectively; t is the deviation value; c is the x 、c y P is the coordinate of the upper left point of the cell; w , P h is the width and height of the bounding box; σ is the activation function.

[0061] In S2.5, the process of the non-maximum suppression algorithm NMS is as follows:

[0062] (1) Sort the output bounding boxes according to the corresponding categories according to the confidence scores to obtain a bounding box list;

[0063] (2) Select the bounding box with the highest confidence and add it to the final output list, and delete it from the bounding box list;

[0064] (3) Calculate the correlation between all bounding boxes and the bounding box with the highest confidence

[0065] (4) Bounding boxes larger than the threshold are removed from the list;

[0066] (5) Repeat the above process until the bounding box list is empty, and the bounding box in the output list is the final predicted box.

[0067] An electronic device, characterized in that the above method is used to detect the self-explosion defect of the insulator.

[0068] A computer storage medium stores at least one program instruction, and the at least one program instruction is loaded and executed by a processor to implement the above-mentioned insulator self-explosion defect detection method.

[0069] The beneficial effect of the embodiment of the present invention is that the conventional convolution is replaced by the deep separable convolution in the backbone network to establish a new backbone network D-Darknet53, which effectively reduces the number of network model parameters and improves the detection speed of the network while slightly reducing the detection accuracy. A dense spatial pyramid pooling module Dense-SPP is proposed to achieve the fusion of local features and global features of the feature map, enrich the expression ability of the feature map, and effectively improve the detection accuracy of the network. A four-scale prediction layer is constructed to make full use of the shallow feature information of the prediction feature layer to improve the network's detection performance for small targets. The problems of low detection accuracy, slow speed, missed detection, and wrong detection at the insulator defect during drone inspection are solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0071] Figure 1 It is a structural schematic diagram of a detection model for self-explosion defects of insulators by aerial photography according to an embodiment of the present invention.

[0072] Figure 2 It is a structural schematic diagram of a backbone network D-Darknet53 of a detection model for self-explosion defects of insulators photographed by aerial photography according to an embodiment of the present invention.

[0073] Figure 3 It is a structural schematic diagram of a dense pyramid pooling module Dense-SPP of a detection model for self-explosion defects of aerial insulators according to an embodiment of the present invention.

[0074] Figure 4 It is a structural schematic diagram of a four-scale prediction layer of a detection model for self-explosion defects of insulators photographed by aerial photography according to an embodiment of the present invention.

[0075] Figure 5It is a flow chart of a method for detecting self-explosion defects of insulators by aerial photography according to an embodiment of the present invention.

[0076] Figure 6 It is a comparison diagram of the detection effects of a method for detecting self-explosion defects of insulators by aerial photography according to an embodiment of the present invention and an original YOLOv3 algorithm. DETAILED DESCRIPTION

[0077] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] Embodiment 1,

[0079] A detection model for self-explosion defects of insulators by aerial photography, the structure of which is as follows: Figure 1 As shown, it includes the backbone network D-Darknet53, feature mining module and four-scale prediction layer connected in sequence;

[0080] The backbone network D-Darknet53 is used to extract image feature information of different scales and levels of insulator defect images.

[0081] like Figure 2 As shown in the figure, the backbone network D-Darknet53 consists of 5 residual modules Resdx (x = 1, 2, 8, 8, 4) and downsampling convolutions. Each residual module is preceded by a downsampling convolution. The input insulator defect image is resized to 608 × 608 by 3 × 3 convolution, and then passes through downsampling convolution and residual modules Resdx (x = 1, 2, 8, 8, 4) in sequence.

[0082] The residual module Resdx is used to extract the texture and edge information of objects in the insulator defect image under different receptive fields, obtain feature maps of different scales, and use deep separable convolution to reduce the number of parameters of the network model; the residual module Resdx consists of x residual units Resdunit;

[0083] The residual unit Resdunit uses depthwise separable convolution, which consists of a depthwise convolution with a kernel size of 3×3 and a point convolution of 1×1. The residual unit input is added to the output of the point convolution and the result is the output of the residual unit.

[0084] The outputs of the last four residual modules Resdx (x=2, 8, 8, 4) in the five residual modules are L 1 , L 2 , L3 , L 4 , the feature map size of the previous layer is twice that of the next layer; the output end of the backbone network D-Darknet53 is connected to the input end of the feature mining module.

[0085] The feature mining module consists of three convolutional layers, dense spatial pyramid pooling Dense-SPP and three convolutional layers, which are used to fuse the local features and overall feature information of the insulator defect image to obtain an insulator defect feature image with rich feature information;

[0086] like Figure 3 As shown,

[0087] Dense spatial pyramid pooling (Dense-SPP) is used to perform pooling operations of different scales on the insulator defect feature map output by the convolution layer, strengthen the fusion of local features and overall features, and obtain an insulator defect feature map with rich feature information;

[0088] The input of the feature mining module is the output L of the backbone network D-Darknet 4 , L 4 After three convolutions with kernel sizes of 1×1, 3×3, and 1×1, the size of the feature map is not changed, and the features are extracted to obtain the insulator defect feature map L 5 ; L 5 As the input of dense spatial pyramid pooling Dense-spp, L 5 Four parallel branches are sent in, one of which is a skip connection, and the other three are composed of cascade, maximum pooling layer and 1×1 convolution layer. The sizes of the three pooling kernels are 5×5, 9×9, and 13×13 respectively; finally, the outputs of the four branches are spliced ​​to obtain the feature L mined by Dense-SPP 6 ; L 6 Then, after three consecutive convolutions with kernel sizes of 1×1, 3×3, and 1×1, feature L is obtained. 7 , which is the final output of the feature mining module.

[0089] The four-scale prediction layer is used to predict the four-scale insulator defect feature images after fusing feature information of different scales and levels to obtain the best prediction frame of the insulator defect.

[0090] like Figure 4 As shown,

[0091] The four-scale prediction layer is composed of four prediction feature layers. The last three feature layers are upsampled and spliced ​​to obtain a four-scale prediction feature map. The upsampling operation uses a suitable interpolation algorithm to insert new elements between pixels on the basis of the original insulator defect feature map, and expands the original feature map by 2 times, so as to perform a splicing operation with the feature map of the previous layer.

[0092] The scales of the four-scale prediction layers are 19*19, 38*38, 76*76, and 152*152 respectively. The receptive field of the small-scale prediction layer is larger and is used to detect large targets. The receptive field of the large-scale prediction layer is smaller and is used to detect small targets.

[0093] The input of the four-scale prediction layer is the output L of the feature mining module. 7 The output L of the three residual modules of the backbone network i (i=1,2,3), the inputs of the 152*152, 76*76, and 38*38 branches are respectively obtained by the previous layer after 1×1 convolution and upsampling operations and L i (i=1,2,3) splicing, and then 5 convolution feature extractions to get the output A i (i=1,2,3), A i (i=1,2,3,4) After two convolution operations with kernel sizes of 3×3 and 1×1 respectively, we finally get four prediction feature maps B with different scales. i Output L of (i=1,2,3,4) 8 ; A 4 =L 7 ;

[0094] The prediction feature layer B with a scale of 152×152 4 , is to output the feature network A 3 After 2 times of upsampling operation, the residual module Resd2 of the backbone network outputs L 1 After concatenation and convolution, the predicted feature map B is obtained. 4 The shallow feature L 1 The location, details and deep features of A 3 The semantic information fusion of the network reduces the loss of feature information at the insulator defect during the feature extraction process, thereby improving the accuracy of network detection of insulator defects.

[0095] Embodiment 2,

[0096] A method for detecting self-explosion defects of insulators by aerial photography, wherein the detection steps are as follows: Figure 5 As shown,

[0097] Step S1, collecting insulator defect images and constructing an insulator defect data set;

[0098] Insulator defect image labeling: use the labeling tool LabelImg to label the insulators and insulator defects in the image. The labeling categories are insulator and defect. Each image generates an xml file with the corresponding labeling location information and category.

[0099] The insulator defect data set is expanded, the number of collected pictures is small, and the network training is prone to overfitting. The original pictures are flipped, randomly cropped, rotated, translated, noise-perturbed, and the brightness contrast is transformed. The expanded data set is divided into a training set and a test set in a ratio of 8:2.

[0100] Step S2: training the MDD-YOLOv3 network to obtain an insulator defect detection model;

[0101] First, the insulator defect image to be detected is input into the backbone network D-Darknet53 to extract the feature output L 4 ;

[0102] Then, L 4 As the input of the feature mining module, after three convolutional feature layers, the extracted feature L 5 Input the Dense-SPP module to obtain the feature L mined by Dense-SPP 6 , L 6 After three convolutional feature layers connected in sequence, feature L is obtained 7 , which is the final output of the feature mining module;

[0103] Next, the output L of the feature mining module 7 The output L of the three residual modules of the backbone network i (i=1,2,3) as the input of the four-scale prediction layer; the three large-scale branches are respectively obtained by convolution upsampling operation and L i After the concatenation operation of (i=1,2,3), the output A is obtained through 5 convolution feature extractions. i (i=1,2,3);

[0104] Next, A i (i=1,2,3,4) After two convolution operations, four prediction feature maps B of different scales are obtained. i (i=1,2,3,4);

[0105] Finally, B i The set of (i=1,2,3,4) is denoted as L 8 , for L 8 The non-maximum suppression algorithm NMS is used to remove redundant frames to obtain the insulator self-explosion defect detection results.

[0106] Step S3: Using a drone to take aerial photos and collect images to test the insulator defect detection model;

[0107] Step S4: output the detection result of the insulator defect detection model to identify and locate the insulator defect.

[0108] In order to evaluate the effectiveness of the improved method described in this embodiment, the present invention conducts ablation experiments on the insulator defect data set, as shown in Table 1.

[0109] Table 1 Ablation experiment results

[0110]

[0111] From the data in Table 1, we can see that Experiment 1 is a YOLOv3 network without the improvement strategy, with a model parameter of 23.47M and an mAP of 92.4% for insulator defect detection. Experiment 4 is an MDD-YOLOv3 network with the above three improvement strategies, with a model parameter of 13.61M, which is 42% lower than that of YOLOv3, and an mAP value of 3.7% higher than that of YOLOv3. Experiments 1 and 4 prove that MDD-YOLOv3 not only reduces the model parameters and meets the real-time requirements in practical applications, but also improves the detection accuracy of the model.

[0112] Experiments 2, 3, and 4 verified that the three improvement strategies are effective. In Experiment 2, D-Darknet53 uses deep separable convolution instead of ordinary convolution in the original YOLOv3 backbone network. As can be seen from the table, the mAP value of the network is reduced by 0.6% compared with the original network, and the model parameters of the network are reduced by 44%. At the expense of a smaller detection accuracy, the network model parameters are greatly reduced, which verifies that this strategy has a good effect. Experiment 3 adds a Dense-SPP module on the basis of Experiment 2. The number of parameters of the network model is slightly increased, and the mAP value is increased by 2.3% compared with Experiment 2. This effect is mainly due to the fact that the Dense-SPP module makes full use of the multi-scale information of the same feature map and enriches the semantic information of the feature map. Experiment 4 adds a four-scale prediction layer on the basis of Experiment 3. On the basis of the three prediction feature layers of the original network, a prediction feature layer with a scale of 152×152 is added to improve the network's detection accuracy for small targets. It can be seen that the mAP has increased by 2.0% compared with Experiment 3. The above can verify that the three improvement strategies proposed in this paper have good detection effects.

[0113] In order to verify the detection performance comparison of the MDD-YOLOv3 algorithm proposed in this embodiment with other mainstream target detection algorithms, this embodiment selects the current mainstream target detection network Faster R-CNN algorithm (Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks [J]. Advances in neural information processing systems, 2015, 28: 91-99); SSD algorithm (Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector [C] / / European conference on computer vision. Springer, Cham, 2016: 21-37.); RetinaNet algorithm (Lin TY, Goyal P, Girshick R, et al. Focal loss for dense object detection [C] / / Proceedings of the IEEE international conference on computer A comparative experiment was conducted on the insulator defect dataset (vision2017:2980-2988). The experimental results are shown in Table 2.

[0114] Table 2 Comparative experiments of different target detection networks

[0115]

[0116] As shown in Table 2, compared with the classic target detection algorithms Faster R-CNN, SSD, and RetinaNet, MDD-YOLOv3 has the highest detection speed and average accuracy. Compared with the traditional single-stage target detection networks SSD, RetinaNet, and YOLOv3, the two-stage target detection network Faster R-CNN with Resnet50+FPN as the backbone network has a detection accuracy of 94.2%, but its detection speed is 6fps, which is difficult to meet the actual application requirements. Among the single-stage target detection networks, RetinaNet with Resnet50+FPN as the backbone network has the highest detection accuracy and a detection speed of 11fps, which is slower than the two detection networks SSD and YOLOv3. It can be seen that MDD-YOLOv3 improves the detection speed of the network while improving the detection accuracy of the network, which can meet the real-time detection requirements of insulator defects.

[0117] like Figure 6 As shown, in order to intuitively demonstrate the detection effect of the improved network, this embodiment tests the YOLOv3 and MDD-YOLOv3 networks. Figure 6 The detection results of YOLOv3 and MDD-YOLOv3 are given, where ① and ② are two transverse insulators that are normal and have self-explosion defects, respectively, and ③ is a longitudinal insulator with self-explosion defects. It can be seen that the detection confidence of YOLOv3 for the two transverse insulators is less than 0.9, and the detection confidence of the longitudinal insulator is only 0.37. There is also a missed detection problem. YOLOv3 did not detect the self-explosion defect in ②, and the detection confidence of the self-explosion defect in ③ was only 0.64. The detection confidence of MDD-YOLOv3 for the three insulators is 0.91, 0.97 and 0.97 respectively. It not only accurately detects the self-explosion defects in ② and ③, but also the detection confidence reaches 0.96 and 0.92 respectively. Compared with YOLOv3, the detection confidence of MDD-YOLOv3 for the longitudinal insulators and their self-explosion defects, which are more difficult to detect, is increased by 162.16% and 43.75% respectively. In summary, it can be verified that the improvement strategy mentioned in this embodiment can not only detect insulators of various angles with high precision, but also effectively avoid the problem of missed detection of self-explosion defects.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A detection model for self-explosion defects of insulators by aerial photography, characterized in that: include: Backbone network D-Darknet53, feature mining module and four-scale prediction layer; The backbone network D-Darknet53 is used to extract image feature information of different scales and levels of insulator defect images; The backbone network D-Darknet53 consists of 5 residual modules Resdx (x=1, 2, 8, 8, 4) and downsampling convolutions. Each residual module is preceded by a downsampling convolution. The input insulator defect image is resized to 608×608 by 3×3 convolution, and then passes through downsampling convolution and residual modules Resdx (x=1, 2, 8, 8, 4) in sequence. The residual module Resdx is used to extract the texture and edge information of the object in the insulator defect image under different receptive fields, obtain feature maps of different scales, and use depth-separable convolution to reduce the number of parameters of the network model; the residual module Resdx is composed of x residual units Resd unit; The residual unit adopts depthwise separable convolution, which is composed of depthwise convolution with a convolution kernel size of 3×3 and point convolution of 1×1; the input end of the residual unit is added to the output end of the point convolution as the output of the residual unit; The outputs of the last four residual modules Resdx (x=2, 8, 8, 4) of the five residual modules are L1, L2, L3, and L4 respectively, and the size of the feature map of the previous layer is twice that of the next layer; The output end of the backbone network D-Darknet53 is connected to the input end of the feature mining module; The feature mining module is used to fuse the local features of the insulator defect image with the overall feature information to obtain an insulator defect feature image with rich feature information; The four-scale prediction layer is used to predict the four-scale insulator defect feature images after fusion of feature information of different scales and different levels to obtain the best prediction frame of the insulator defect; The four-scale prediction layer is composed of four prediction feature layers, and the last three feature layers are upsampled and spliced ​​to obtain a four-scale prediction feature map; The upsampling operation inserts new elements between pixels using a suitable interpolation algorithm based on the original insulator defect feature map, expands the original feature map by 2 times, and then performs a splicing operation with the feature map of the previous layer; The scales of the four-scale prediction layers are 19*19, 38*38, 76*76, and 152*152, respectively. The receptive field of the small-scale prediction layer is larger and is used to detect large targets, while the receptive field of the large-scale prediction layer is smaller and is used to detect small targets. The input of the four-scale prediction layer is the output L7 of the feature mining module and the output L of the three residual modules of the backbone network. i (i=1,2,3), the inputs of the 152*152, 76*76, and 38*38 branches are respectively obtained by the previous layer after 1×1 convolution and upsampling operations and L i (i=1,2,3) splicing, and then 5 convolution feature extractions to get the output A i (i=1,2,3), A i (i=1,2,3,4) After two convolution operations with kernel sizes of 3×3 and 1×1 respectively, we finally get four prediction feature maps B with different scales. i Output L8 of (i=1,2,3,4); A4=L7; The predicted feature layer B4 with a scale of 152×152 is obtained by performing a 2-fold upsampling operation on the feature network output A3, concatenating the convolution operation with the residual module Resd2 output L1 of the backbone network; the predicted feature map B4 fuses the position and detail information of the shallow feature L1 with the semantic information of the deep feature A3, thereby reducing the loss of feature information at the insulator defect during the feature extraction process, thereby improving the accuracy of the network in detecting insulator defects.

2. The detection model of the self-explosion defect of an insulator by aerial photography according to claim 1 is characterized in that: The feature mining module is composed of three convolutional layers, dense spatial pyramid pooling Dense-SPP and three convolutional layers in sequence; The convolution layer is used to perform a convolution operation on the output feature map of the backbone network to improve the feature expression capability of the insulator defect feature map; The dense spatial pyramid pooling Dense-SPP is used to perform pooling operations of different scales on the insulator defect feature map output by the convolution layer, strengthen the fusion of local features and overall features, and obtain an insulator defect feature map with rich feature information; The input of the feature mining module is the output L4 of the backbone network D-Darknet. L4 is convolved with three convolution kernels of size 1×1, 3×3, and 1×1 respectively, without changing the size of the feature map, and the features are extracted to obtain the insulator defect feature map L5; L5 is used as the input of dense spatial pyramid pooling Dense-spp, and L5 is sent to four parallel branches, one of which is a skip connection, and the other three are composed of cascade, maximum pooling layer and 1×1 convolution layer, and the sizes of the three pooling kernels are 5×5, 9×9, and 13×13 respectively; finally, the outputs of the four branches are spliced ​​to obtain the feature L6 mined by Dense-SPP; L6 is then convolved with three convolution kernels of size 1×1, 3×3, and 1×1 respectively connected in sequence to obtain the feature L7, which is the final output of the feature mining module.

3. A method for detecting self-explosion defects of insulators by aerial photography, characterized in that: The detection process is carried out in the following steps: Step S1, collecting insulator defect images and constructing an insulator defect data set; Step S2, training the MDD-YOLOv3 network to obtain an insulator defect detection model; the process of training the MDD-YOLOv3 network in step S2 is as follows: Step S2.1, input the insulator defect image to be detected into the backbone network D-Darknet53 to extract the feature output L4; Step S2.2, L4 is used as the input of the feature mining module. After passing through three convolutional feature layers, the extracted feature L5 is input into the Dense-SPP module to obtain the feature L6 mined by the Dense-SPP. L6 passes through three sequentially connected convolutional feature layers to obtain feature L7, which is the final output of the feature mining module; Step S2.3: The output L7 of the feature mining module and the output L of the three residual modules of the backbone network i (i=1,2,3) as the input of the four-scale prediction layer; the three large-scale branches are respectively obtained by convolution upsampling operation and L i After the concatenation operation of (i=1,2,3), the output A is obtained through 5 convolution feature extractions. i (i=1,2,3); Step S2.4, A i (i=1,2,3,4) After two convolution operations, four prediction feature maps B of different scales are obtained. i (i=1,2,3,4); Step S2.5: B i The set of (i=1,2,3,4) is recorded as L8. After using the non-maximum suppression algorithm NMS on L8 and removing the redundant frames, the insulator self-explosion defect detection result is obtained; Step S3: Using a drone to take aerial photos and collect images to test the insulator defect detection model; Step S4: output the detection result of the insulator defect detection model to identify and locate the insulator defect.

4. The method for detecting the self-explosion defect of an insulator by aerial photography according to claim 3 is characterized in that: In the step S1, an insulator defect data set is constructed to annotate the collected insulator defect images and expand the data; The insulator defect image labeling uses the labeling tool LabelImg to label the insulators and insulator defects in the image, with the labeling categories being insulator and defect, and an xml file with corresponding labeling location information and category is generated for each image; The insulator defect data set is expanded, the number of collected pictures is small, and the network training is prone to overfitting. The original pictures are flipped, randomly cropped, rotated, translated, noise-perturbed, and the brightness contrast is transformed. The expanded data set is divided into a training set and a test set in a ratio of 8:

2.

5. An electronic device, characterized in that: The method described in any one of claims 3 to 4 is used to detect the self-explosion defect of the insulator.

6. A computer storage medium, characterized in that: At least one program instruction is stored in the storage medium, and the at least one program instruction is loaded and executed by the processor to implement the method for detecting the self-explosion defect of an insulator as described in any one of claims 3-4.

Images