A method for detecting hanging objects in high-voltage power grid based on deep learning

Abstract

The present disclosure discloses a method for detecting hanging objects in a high-voltage power grid based on deep learning, comprising the following steps: collecting label images of hanging objects in a high-voltage power grid to be detected; inputting the label images of hanging objects in a high-voltage power grid to be detected into a high-voltage power grid hanging object image recognition model for detection, so as to determine the position of the hanging objects in the high-voltage power grid. The present disclosure also provides a device for detecting hanging objects in a high-voltage power grid based on deep learning, comprising: a collection unit for collecting label images of hanging objects in a high-voltage power grid to be detected; a detection unit for inputting the label images of hanging objects in a high-voltage power grid to be detected into a high-voltage power grid hanging object image recognition model for detection, so as to determine the position of the hanging objects in the high-voltage power grid.

Description

Technical Field

[0001] The present disclosure belongs to the field of image processing technology, and in particular relates to a method for detecting hanging objects in a high-voltage power grid based on deep learning. Background Art

[0002] In order to ensure the safe and stable operation of power grid equipment, transmission lines must be frequently inspected and cleaned for floating objects. Since strong winds may cause foreign objects to hang on, wires to dance, and cause damage to equipment such as tower collapse and wire breakage, it is necessary to use telescopes, drones, etc. to conduct detailed inspections of lines and equipment, master each operating line, check whether there are hanging objects on equipment and contacts, and discover hidden dangers in time. Due to the problems of high-voltage transmission lines crossing a wide area and long lines, manual inspections cannot eliminate hidden dangers in time. With the development of deep learning, the use of deep learning-based target detection plus drone intelligent inspection has emerged. Using drones to shoot transmission lines can quickly complete information collection on the power grid, and safety hazards can be promptly identified by analyzing the collected images. However, there are some problems with using drones for inspections and taking photos, and using target detection models to detect floating objects on the power grid, such as low recognition accuracy, low recall rate, and low detection efficiency. Summary of the invention

[0003] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a method for detecting hanging objects in a high-voltage power grid based on deep learning, which can improve the detection accuracy of the position of hanging objects in a high-voltage power grid image and effectively help maintenance personnel to efficiently inspect the high-voltage power grid.

[0004] To achieve the above objectives, the present disclosure provides the following technical solutions:

[0005] A method for detecting hanging objects in a high-voltage power grid based on deep learning comprises the following steps:

[0006] S100: Collecting label images of hanging objects in the high-voltage power grid to be tested;

[0007] S200: Inputting the label image of the high-voltage power grid hanging object to be tested into a high-voltage power grid hanging object image recognition model for detection to determine the position of the hanging object in the high-voltage power grid.

[0008] Preferably, the high-voltage power grid hanging object image recognition model includes:

[0009] A backbone network, used for extracting and fusing features of the label image of the hanging object in the high-voltage power grid to be tested, so as to generate a feature map;

[0010] A detection component is used to detect the feature map generated by the backbone network.

[0011] Preferably, the backbone network includes:

[0012] A feature extraction network, the feature extraction network comprising first to fifth feature extraction layers, for extracting features from the label image of the high-voltage power grid hanging object to be tested;

[0013] A feature fusion network, the feature fusion network includes first to fourth feature fusion layers, which are used to fuse the features of the extracted label images of the hanging objects in the high-voltage power grid to be tested to form a feature map.

[0014] Preferably, the detection component comprises:

[0015] A dimensionality scaling feature fusion module, used to scale down the feature map output by the feature fusion network;

[0016] A detection-side attention module, used for performing feature enhancement processing on the feature map scaled down by the dimension scaling feature fusion module;

[0017] A region of interest pooling module, used for amplifying the feature map processed by the detection end attention module through interpolation calculation to obtain a feature amplification map;

[0018] The non-maximum suppression module is used to merge the overlapping feature magnification images output by the region of interest pooling module to output a prediction box.

[0019] Preferably, the detection end attention module includes a convolutional layer, a BN layer, a RULE layer, a transposed convolutional layer, a BN layer, a transposed convolutional layer and a parallel attention layer connected in sequence.

[0020] Preferably, the training process of the high-voltage power grid hanging object image recognition model includes the following steps:

[0021] S1000: Use the COCO2017 dataset to pre-train the high-voltage power grid hanging object image recognition model;

[0022] S2000: Divide the constructed high-voltage power grid hanging object label image dataset into a training set and a test set, input the training set into the pre-trained high-voltage power grid hanging object image recognition model to train the model, and after the training is completed, input the test set into the trained high-voltage power grid hanging object image recognition model to test the model. If the model accuracy is greater than 90% and the recall rate is greater than 90% during the test, the test passes, and a trained high-voltage power grid hanging object image recognition model is obtained.

[0023] The present disclosure also provides a high-voltage power grid hanging object detection device based on deep learning, comprising:

[0024] A collection unit, used to collect label images of hanging objects on the high-voltage power grid to be tested;

[0025] The detection unit is used to input the label image of the high-voltage power grid hanging object to be detected into the high-voltage power grid hanging object image recognition model for detection to determine the position of the hanging object in the high-voltage power grid.

[0026] Preferably, the detection unit comprises:

[0027] A feature extraction subunit is used to extract features from the label image of the hanging object on the high-voltage power grid to be tested;

[0028] The feature fusion subunit is used to fuse the image features extracted by the feature extraction subunit to generate a feature map.

[0029] Preferably, the detection unit further comprises:

[0030] A feature map scaling subunit, used to reduce the scale of the feature map output by the feature fusion subunit;

[0031] A feature map enhancement subunit, used for performing feature enhancement processing on the feature map scaled down by the feature map scaling subunit;

[0032] A region of interest pooling subunit, used for amplifying the feature map processed by the feature map enhancement subunit through interpolation calculation to obtain a feature amplified map;

[0033] The non-maximum suppression subunit is used to merge the overlapping feature magnification images output by the region of interest pooling subunit to output a prediction box.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. The high-voltage power grid hanging object image detection model proposed in this disclosure can effectively detect foreign object information on the power grid and improve the text label detection capability by introducing the use of attention mechanism and interpolation method to detect hanging objects on the power grid. The use of weighted cross entropy loss can make the model output feature vector have a better training effect. The use of attention mechanism can make up for the problem of insufficient hanging object detection capability in general target detection models, making the target position of hanging objects in the power grid more accurate, improving detection capability and reducing the hidden dangers of power grid contact;

[0036] 2. The present invention adopts the ResNet+FPN structure as the backbone network, which can improve the detection accuracy of the position of hanging objects in the high-voltage power grid image, effectively help maintenance personnel to efficiently inspect the high-voltage power grid and improve the safety of the power grid system. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of a method for detecting hanging objects in a high-voltage power grid based on deep learning provided by an embodiment of the present disclosure;

[0038] Figure 2 It is a structural schematic diagram of a high-voltage power grid hanging object image recognition model provided by another embodiment of the present disclosure. DETAILED DESCRIPTION

[0039] The following will refer to the attached Figure 1 and Figure 2 Specific embodiments of the present disclosure are described in detail. Although specific embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0040] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present disclosure, but the description is based on the general principles of the specification and is not used to limit the scope of the present disclosure. The scope of protection of the present disclosure shall be determined by the attached claims.

[0041] To facilitate understanding of the embodiments of the present disclosure, further explanation will be given below using specific embodiments as examples in conjunction with the accompanying drawings, and each of the accompanying drawings does not constitute a limitation on the embodiments of the present disclosure.

[0042] In one embodiment, if Figure 1 As shown, the present disclosure provides a method for detecting hanging objects in a high-voltage power grid based on deep learning, comprising the following steps:

[0043] S100: Collecting label images of hanging objects in the high-voltage power grid to be tested;

[0044] S200: Inputting the label image of the high-voltage power grid hanging object to be tested into a high-voltage power grid hanging object image recognition model for detection to determine the position of the hanging object in the high-voltage power grid.

[0045] In another embodiment, the high-voltage power grid hanging object image recognition model includes:

[0046] A backbone network, used for extracting and fusing features of the label image of the hanging object in the high-voltage power grid to be tested, so as to generate a feature map;

[0047] A detection component is used to detect the feature map generated by the backbone network.

[0048] In another embodiment, the backbone network includes:

[0049] A feature extraction network, the feature extraction network comprising first to fifth feature extraction layers, for extracting features from the label image of the high-voltage power grid hanging object to be tested;

[0050] A feature fusion network, the feature fusion network includes first to fourth feature fusion layers, which are used to fuse the features of the extracted label images of the hanging objects in the high-voltage power grid to be tested to form a feature map.

[0051] In this embodiment, Figure 2 As shown, the feature extraction network includes a first feature extraction layer C1, a second feature extraction layer C2, a third feature extraction layer C3, a fourth feature extraction layer C4 and a fifth feature extraction layer C5 connected in sequence. The label image of the high-voltage power grid hanging object to be tested is sequentially subjected to feature extraction by the first feature extraction layer, the second feature extraction layer, the third feature extraction layer, the fourth feature extraction layer and the fifth feature extraction layer, and each layer obtains the image features of the high-voltage power grid hanging object.

[0052] The feature fusion network includes the first layer P2, the second layer P3, the third layer P4 and the fourth layer P5 of FPN. The output feature maps of the fifth feature extraction layer C5 and the fourth feature extraction layer C4 are fused using the attention layer and then sent to the fourth layer of FPN for processing. The output feature maps of the third feature extraction layer C3 and the fourth layer of FPN are fused using the attention layer and then sent to the third layer of FPN for processing. The output feature maps of the second feature extraction layer C2 and the third layer of FPN are fused using the attention layer and then sent to the second layer of FPN for processing. The output feature maps of the first feature extraction layer C1 and the second FPN layer are fused using the attention layer and then sent to the first layer of FPN for processing.

[0053] In the specific image processing process, the feature extraction network selects ResNet, and uses the SKNet attention layer to replace the original 3×3 convolution for feature extraction. Specifically, in the FPN structure, each layer of features in FPN is upsampled and scaled before fusion. Among them, the feature map output by FPN is {P2, P3, P4, P5}, and the upsampling process is: P2 is enlarged by 2 times, P3 is enlarged by 4 times, P4 is enlarged by 8 times, and P5 is enlarged by 16 times, and then spliced ​​and fused after unifying the scale.

[0054] In another embodiment, the detection component includes:

[0055] A dimensionality scaling feature fusion (Conv) module, used to scale down the feature map output by the feature fusion network;

[0056] A detection-side attention (SKNet) module, used to perform feature enhancement processing on the feature map scaled down by the dimension scaling feature fusion module;

[0057] A region of interest pooling (ROIPooling) module, used to amplify the feature map processed by the detection end attention module through interpolation calculation to obtain a feature amplification map;

[0058] A non-maximum suppression (NMS) module is used to merge the overlapping feature magnification images output by the region of interest pooling module to output a prediction box.

[0059] In this embodiment, the feature map output by the feature fusion network is sequentially sent to the dimension scaling feature fusion unit and the detection end attention module for processing, wherein the dimension scaling feature fusion module expands the output features of the feature fusion network to the same scale vertical axis splicing, and then performs convolution, batch normalization, and maximum pooling processing to achieve scale reduction to 1 / 4 of the original, and the dimension is reduced to 1 / 4 of the original. The scale-reduced feature map is input into the detection end attention module, and then the feature map processed by attention is used to perform pooling interpolation of the region of interest to generate a detection map, and finally the detection map is sent to the non-maximum suppression module to obtain the prediction result.

[0060] In another embodiment, the detection end attention module includes a convolutional layer, a BN layer, a RULE layer, a transposed convolutional layer, a BN layer, a transposed convolutional layer and a parallel attention layer connected in sequence.

[0061] In another embodiment, the training process of the high-voltage power grid hanging object image recognition model includes the following steps:

[0062] S1000: Use the COCO2017 dataset to pre-train the high-voltage power grid hanging object image recognition model;

[0063] In this step, the high-voltage power grid hanging object image recognition model is pre-trained using the COCO2017 dataset, and the model learns network parameters in multiple iterations, which enables the model to have good detection performance.

[0064] S2000: Divide the constructed high-voltage power grid hanging object label image dataset into a training set and a test set, input the training set into the pre-trained high-voltage power grid hanging object image recognition model to train the model, and after the training is completed, input the test set into the trained high-voltage power grid hanging object image recognition model to test the model. If the model accuracy is greater than 90% and the recall rate is greater than 90% during the test, the test passes, and a trained high-voltage power grid hanging object image recognition model is obtained.

[0065] In this step, this embodiment uses the general COCO2017 dataset format to create a high-voltage power grid hanging object image dataset. First, the Labelme open source dataset annotation tool uses the four-point annotation method to annotate the high-voltage power grid hanging object image, and then converts the annotation file into the COCO2017 dataset format. Each high-voltage power grid hanging object image in the constructed dataset D corresponds to a label file. Finally, the high-voltage power grid hanging object images in the dataset D are divided into training set D according to a ratio of 1:3. train and the test set D test .

[0066] The detection results of the high-voltage power grid hanging object image recognition model trained in the above embodiment and the existing image recognition model are shown in Table 1:

[0067] Table 1

[0068]

[0069] In Table 1, the detection performance of the image recognition model described in the present disclosure and the existing ones is evaluated by using the detection accuracy P (Presision, which is a detection index), the detection recall R (Recall, which is an index for judging missed detection) and the number of images processed per second (Frames Per Second, the number of images processed per second is recorded as FPS, which is the number of images processed in 1 second, used to indicate the computational complexity of the model. The larger the FPS, the lower the computational complexity of the model). As can be seen from Table 1, the various indicators of the model described in the present disclosure are better than those of the existing image recognition models including Faster RCNN, YOLOV3 and EfficientDet3, indicating that the detection accuracy of the model described in the present disclosure for hanging objects on the high-voltage power grid is better than that of the existing models.

[0070] In another embodiment, the present disclosure further provides a high-voltage power grid hanging object detection device based on deep learning, comprising:

[0071] A collection unit, used to collect label images of hanging objects on the high-voltage power grid to be tested;

[0072] The detection unit is used to input the label image of the high-voltage power grid hanging object to be detected into the high-voltage power grid hanging object image recognition model for detection to determine the position of the hanging object in the high-voltage power grid.

[0073] In another embodiment, the detection unit comprises:

[0074] A feature extraction subunit is used to extract features from the label image of the hanging object on the high-voltage power grid to be tested;

[0075] The feature fusion subunit is used to fuse the image features extracted by the feature extraction subunit to generate a feature map.

[0076] In another embodiment, the detection unit further includes:

[0077] A feature map scaling subunit, used to reduce the scale of the feature map output by the feature fusion subunit;

[0078] A feature map enhancement subunit, used for performing feature enhancement processing on the feature map scaled down by the feature map scaling subunit;

[0079] A region of interest pooling subunit, used for amplifying the feature map processed by the feature map enhancement subunit through interpolation calculation to obtain a feature amplified map;

[0080] The non-maximum suppression subunit is used to merge the overlapping feature magnification images output by the region of interest pooling subunit to output a prediction box.

[0081] The above specific embodiments are used to illustrate the present invention, which is only used to help understand the present disclosure and is not used to limit the present invention. Any local modification or replacement within the technical scope disclosed by the present invention by any technician familiar with the technology should be included in the scope of the present invention.

Claims

1. A method for detecting hanging objects in a high-voltage power grid based on deep learning, comprising the following steps: S100: Collecting label images of hanging objects in the high-voltage power grid to be tested; S200: inputting the label image of the high-voltage power grid hanging object to be tested into a high-voltage power grid hanging object image recognition model for detection to determine the position of the hanging object in the high-voltage power grid; The high-voltage power grid hanging object image recognition model includes: A backbone network, used for extracting and fusing features of the label image of the hanging object in the high-voltage power grid to be tested, so as to generate a feature map; The backbone network includes: A feature extraction network, the feature extraction network comprising first to fifth feature extraction layers, for extracting features from the label image of the high-voltage power grid hanging object to be tested; A feature fusion network, the feature fusion network comprising first to fourth feature fusion layers, for fusing the features of the extracted label image of the hanging object of the high-voltage power grid to be tested to form a feature map; A detection component, used to detect the feature map generated by the backbone network; The detection component comprises: A dimensionality scaling feature fusion module, used to scale down the feature map output by the feature fusion network; A detection-side attention module, used for performing feature enhancement processing on the feature map scaled down by the dimension scaling feature fusion module; A region of interest pooling module, used for amplifying the feature map processed by the detection end attention module through interpolation calculation to obtain a feature amplification map; The non-maximum suppression module is used to merge the overlapping feature magnification images output by the region of interest pooling module to output a prediction box.

2. The method according to claim 1, wherein: The detection end attention module includes a convolutional layer, a BN layer, a RULE layer, a transposed convolutional layer, a BN layer, a transposed convolutional layer and a parallel attention layer connected in sequence.

3. The method according to claim 1, wherein: The training process of the high-voltage power grid hanging object image recognition model includes the following steps: S1000: Use the COCO2017 dataset to pre-train the high-voltage power grid hanging object image recognition model; S2000: Divide the constructed high-voltage power grid hanging object label image dataset into a training set and a test set, input the training set into the pre-trained high-voltage power grid hanging object image recognition model to train the model, and after the training is completed, input the test set into the trained high-voltage power grid hanging object image recognition model to test the model. If the model accuracy is greater than 90% and the recall rate is greater than 90% during the test, the test passes, and a trained high-voltage power grid hanging object image recognition model is obtained.

4. A high-voltage power grid hanging object detection device for implementing the method as claimed in claim 1, comprising: A collection unit, used to collect label images of hanging objects on the high-voltage power grid to be tested; The detection unit is used to input the label image of the high-voltage power grid hanging object to be detected into the high-voltage power grid hanging object image recognition model for detection to determine the position of the hanging object in the high-voltage power grid.

5. The device according to claim 4, wherein: The detection unit comprises: A feature extraction subunit is used to extract features from the label image of the hanging object on the high-voltage power grid to be tested; The feature fusion subunit is used to fuse the image features extracted by the feature extraction subunit to generate a feature map.

6. The device according to claim 5, wherein: The detection unit also includes: A feature map scaling subunit, used to reduce the scale of the feature map output by the feature fusion subunit; A feature map enhancement subunit, used for performing feature enhancement processing on the feature map scaled down by the feature map scaling subunit; A region of interest pooling subunit, used for amplifying the feature map processed by the feature map enhancement subunit through interpolation calculation to obtain a feature amplified map; The non-maximum suppression subunit is used to merge the overlapping feature magnification images output by the region of interest pooling subunit to output a prediction box.

Images