Power transmission line anomaly target detection method based on improved YOLOX algorithm

Abstract

The application relates to a power transmission line abnormal target detection method based on an improved YOLOX algorithm, which comprises the following steps: 1) inputting a power transmission line abnormal image to be identified, adjusting the image into a standard input image after pretreatment, and training based on the improved YOLOX algorithm; 2) using a Focus model to split a high-resolution image into multiple low-resolution images; 3) extracting the features of the input image as abnormal features through a feature selection module; 4) continuously extracting the abnormal features and dividing them into three effective feature layers P3, P4 and P5; 5) inputting the feature maps of the effective feature layers into a PAFPN network for feature extraction, efficiently fusing the feature maps of different feature layers, and outputting the feature maps of different feature layers; and 6) detecting and positioning the abnormal targets in the effective feature maps of different feature layers. Compared with the prior art, the application has the advantages of high abnormal target detection precision and reduced algorithm complexity.

Description

TECHNICAL FIELD

[0001] The application relates to a power transmission line abnormal target detection method, in particular to a power transmission line abnormal target detection method based on an improved YOLOX algorithm. BACKGROUND

[0002] With the development of science and technology, the laying range of power transmission lines gradually covers all parts of the country, and the structural design form of power equipment and related power components is also becoming more and more complex. However, as the power demand of people's daily life and production gradually increases, the probability of power supply and distribution line operation failure is also increased. If the safe and stable operation of the power system cannot be guaranteed, it will affect people's daily life and also cause significant social losses. Therefore, how to ensure the safe and reliable operation of the power supply and distribution line has become a major challenge in the power transmission link.

[0003] By detecting the targets in the power transmission line image, such as loose cables and damaged equipment, the fault conditions on the power transmission line can be found in time to avoid power interruption or safety accidents. At the same time, the detection of abnormal targets can also improve the maintenance efficiency of the power transmission line, reduce the maintenance cost, and improve the reliability and safety of the power system.

[0004] However, the traditional detection of power transmission lines is realized by manual operation, which is low in efficiency and consumes manpower and resources. Therefore, it is of great practical significance and application value to study the power transmission line abnormal target detection algorithm based on deep learning. However, the use of unmanned aerial vehicles for power transmission line detection has the following problems:

[0005] (1) In the process of collecting power transmission line images, the unmanned aerial vehicle will pass through areas with different topographic features, such as villages, farmlands, rivers, mountains, glaciers, forests, and roads, which makes the collected images have constantly changing background information.

[0006] (2) The light and weather conditions (such as sunny, rainy days, etc.) cannot be determined in advance, so the contact network images collected by the equipment will be seriously disturbed by fog, haze and different light, increasing the detection difficulty.

[0007] (3) The unmanned aerial vehicle needs to be manually operated by power grid staff, and in order to avoid accidents such as collision, the precise fault position cannot be collected. Generally, the field of view of the unmanned aerial vehicle is wide, so the resolution of the images taken is high. However, the abnormal targets on the power transmission line are mostly hidden and easily ignored, and the detection is prone to deviation.

[0008] (4) In the field of transmission line detection, the physical size of the abnormal target photographed by the unmanned aerial vehicle is small. Moreover, due to the complex background, occlusion, edge blur and lack of samples, it is difficult for the neural network to extract enough target feature information. At the same time, small target detection is also a complex problem and a hot research topic in computer vision, which is a challenging task.

[0009] (5) At present, the detection algorithm based on deep learning mainly aims at the detection of general targets, and when applied to the transmission line abnormal detection scene, the performance of the algorithm is obviously reduced, and the detection performance cannot meet the requirements of industrial application.

[0010] Therefore, how to realize a small target detection method capable of detecting transmission line abnormalities in real time has become a technical problem to be solved. SUMMARY

[0011] The purpose of the present application is to overcome the defects of the prior art and provide a transmission line abnormal target detection method based on an improved YOLOX algorithm.

[0012] The purpose of the present application can be achieved by the following technical solutions:

[0013] A transmission line abnormal target detection method based on an improved YOLOX algorithm, the method comprising the following steps:

[0014] 1) input the transmission line abnormal image to be identified, adjust the preprocessed standard input image, and train the preprocessed standard input image based on the improved YOLOX algorithm;

[0015] 2) use the Focus model to split the high-resolution image into multiple low-resolution images;

[0016] 3) extract the features of the input image as abnormal features through the feature selection module;

[0017] 4) continuously extract the abnormal features and divide them into three effective feature layers P3, P4 and P5;

[0018] 5) input the feature maps of the effective feature layers based on the new double-channel attention mechanism into the PAFPN network for feature extraction, efficiently fuse the feature maps of different feature layers, and output the feature maps of different feature layers;

[0019] 6) detect and locate the abnormal target for the effective feature maps of different feature layers entering the detector.

[0020] Further, the transmission line image to be identified is obtained by a unmanned aerial vehicle carrying a camera, which is for transmission line images in different positions, different weather and different illumination.

[0021] Further, the preprocessing is specifically: normalizing the image, all images are cropped to a resolution of 1368*912, using labeme software to label the abnormal target, obtaining the power transmission line abnormal data set, and dividing the data set into a training set and a test set.

[0022] Further, the power transmission line abnormal data set includes a high-voltage tower bird nest data set and a self-broken insulator data set.

[0023] Further, the 3) is specifically:

[0024] 301) The input feature mapping is divided into C feature sub-mappings, and an N-dimensional vector of each feature sub-mapping is obtained through a full connection layer;

[0025] 302) The N-dimensional vector is constrained into [0, 1] using a Softmax activation function, and then combined with the corresponding feature sub-mapping as a weighted weight;

[0026] 303) The C feature sub-mappings are combined into a new feature mapping containing all basic features of the power transmission line abnormal target, as an abnormal feature.

[0027] Further, P3, P4 and P5 are used for detecting small targets, medium targets and large targets respectively; the FReLU activation function is used for optimization in the continuous extraction of the abnormal feature, and is expressed as:

[0028]

[0029] Wherein, T(x) is a simple and effective spatial context feature extractor, and x represents an abnormal feature.

[0030] Further, the 5) is specifically: the input is divided into two branches along the channel direction, respectively generating attention sub-graphs, and then merging after processing to obtain new feature maps of different feature layers.

[0031] Further, the new feature maps of different feature layers include cross-channel information, direction information and position information of the abnormal target.

[0032] An electronic device comprising a memory and a processor, the memory having a computer program stored thereon, and the processor implementing the method as described above when executing the program.

[0033] A computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the method as described above.

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

[0035] I. The present application improves the YOLOX algorithm, adds a feature selection module in the backbone network, extracts important feature information in the effective feature map by assigning weights, and reserves practical features for subsequent feature extraction, further improving the detection accuracy of abnormal targets in the power transmission line. By optimizing the residual structure in the feature extraction network, SimAM is introduced in the residual edge for processing, and the neurons on the residual edge are optimized to improve the target feature representation and the ability of the network to detect small abnormal targets in the power transmission line.

[0036] II. The present application improves the activation function of the YOLOX network, uses FReLU to enhance the nonlinear representation of small abnormal objects in the power transmission line. FReLU function is used in the residual module of feature layer P3, so that the ability to detect small targets can be improved while preventing excessive network overhead and reducing algorithm complexity.

[0037] III. The present application proposes a new attention mechanism feature enhancement network (AM-FPNet) based on a dual-channel attention idea. This attention mechanism is embedded in the shallow feature enhancement map of the FPN module, i.e. P3, to improve the feature information of small abnormal targets in the power transmission line. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is the structure diagram of the YOLOX algorithm before the improvement of the present application;

[0039] Figure 2 is a schematic diagram of the feature selection module of the present application;

[0040] Figure 3 is a schematic diagram of the residual network optimized by SimAM of the present application;

[0041] Figure 4 is a schematic diagram of the improved feature extraction network AM-FPNet of the present application;

[0042] Figure 5 is a schematic diagram of the feature enhancement network fused with the attention mechanism of the present application. DETAILED DESCRIPTION

[0043] The present application will be described in detail below in conjunction with the drawings and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present application, and gives detailed implementation and specific operation process, but the protection scope of the present application is not limited to the following examples.

[0044] EMBODIMENT

[0045] As Figure 1As shown, the structure diagram of the YOLOX algorithm before improvement, which is mainly aimed at the detection of general targets, when applied to the power line anomaly detection scene, the performance of the algorithm is obviously decreased, and the detection performance cannot meet the requirements of industrial application.

[0046] As shown in Figures 2-5 A power line anomaly target detection method based on an improved YOLOX algorithm, characterized in that the method comprises the following steps:

[0047] 1) Input the power line anomaly image to be identified, adjust it to a standard input image after preprocessing, and train the preprocessed standard input image based on the improved YOLOX algorithm.

[0048] The power line anomaly image to be identified is obtained by a camera carried by a drone. The operator controls the drone to fly on the ground and collects power line images at different positions, in different weather and under different light. The preprocessing is as follows: normalize the image, crop all images to a resolution of 1368x912, label the abnormal target using labeme software (data labeling software), obtain the power line anomaly data set, including the high-voltage tower bird nest data set and the self-blowing insulator data set, and divide the data set into a training set and a test set. In the training process, n images are randomly selected as input images each time.

[0049] 2) Use the Focus model to split the high-resolution image into multiple low-resolution images.

[0050] 3) Extract the features of the input image through the feature selection module as abnormal features.

[0051] 301) Divide the input feature mapping into C feature sub-mappings, and obtain an N-dimensional vector of each feature sub-mapping through a fully connected layer;

[0052] 302) Use the Softmax activation function to constrain the N-dimensional vector to [0, 1], and then combine it with the corresponding feature sub-mapping as a weighted weight;

[0053] 303) Combine the C feature sub-mappings into a new feature mapping containing all basic features of the power line anomaly target, as abnormal features.

[0054] 4) Continuously extract abnormal features and divide them into three effective feature layers P3, P4 and P5.

[0055] With the increase of network layers, the size of the extracted abnormal feature map is also different, and P3, P4 and P5 are used for detecting small targets, medium targets and large targets respectively; since most of the targets on the transmission line are small targets, the position of the SPPBottleneck structure is adjusted from the P5 layer to the P3 layer. Compared with the original backbone network, the adjusted network can reduce part of the network parameters, while improving the receptive field of small targets. In addition, in order to improve the detection performance of the network on small targets, the FReLU activation function is used for optimization in the process of continuously extracting abnormal features, which is expressed as:

[0056]

[0057] Wherein, T(x) is a simple and effective spatial context feature extractor, and x represents an abnormal feature.

[0058] 5) The feature map of the effective feature layer is input into the PAFPN network based on the new dual-channel attention mechanism for feature extraction, and the feature maps of different feature layers are efficiently fused, and the feature maps of different feature layers are output.

[0059] The input is divided into two branches along the channel direction, and attention subgraphs are generated respectively by using the relationship between channels, and the direction perception and position perception information of the target are captured by pooling. In the channel attention subgraph, the convolution kernel is used to learn the corresponding relationship between channels, and after calculation by the Sigmoid function, the input is multiplied by the value. In addition, in order to better learn the detailed target information, different pooling kernels are used to capture the long-range dependencies along the two spatial directions. Then, different size subgraphs are combined and converted into subgraphs with the same number of channels. Finally, a 1x1 convolution operation is performed, and then multiplied by the input as the attention weight. After obtaining two attention subgraphs, they are combined to obtain new feature maps of different feature layers; the new feature maps of different feature layers include cross-channel information, direction information and position information of abnormal targets. The new dual-channel attention mechanism is embedded into the P3 shallow feature enhancement map to improve the feature information of small abnormal targets in the transmission line.

[0060] 6) For the effective feature maps of different feature layers entering the detector, the detection and positioning of abnormal targets are carried out.

[0061] Particularly, all the parameters of the present application are deep network parameters except for the hyperparameters indicating the values, which are autonomously learned and optimized by the Adam optimizer method.

[0062] An electronic device includes a memory and a processor, the memory having a computer program stored thereon, and the processor implements the method as described above when executing the program.

[0063] A computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the method as described above.

[0064] The preferred embodiments of the present application have been described in detail above. It should be understood that modifications and variations can be made by persons of ordinary skill in the art without creating spurious equivalents while remaining within the scope of the present application. Therefore, the technical solutions obtained by logical analysis, reasoning or limited experiments based on the conception of the present application and the prior art in the technical field should be within the protection scope defined by the claims.

Claims

1. A power transmission line anomaly target detection method based on an improved YOLOX algorithm, characterized in that, The method comprises the following steps: 1) inputting an image to be identified of a power transmission line anomaly, adjusting the image to a standard input image after preprocessing, and training the standard input image after preprocessing based on an improved YOLOX algorithm; 2) using a Focus model to split a high-resolution image into multiple low-resolution images; 3) extracting features of the input image as abnormal features through a feature selection module; 4) continuously extracting abnormal features and dividing them into three effective feature layers P3, P4 and P5; 5) inputting feature maps of the effective feature layers based on a new dual-channel attention mechanism into a PAFPN network for feature extraction, efficiently fusing feature maps of different feature layers, and outputting feature maps of different feature layers; 6) detecting and locating abnormal targets for effective feature maps of different feature layers entering the detector. The improved YOLOX algorithm adds a feature selection module in the backbone network, extracts important feature information in the effective feature map by assigning weights, optimizes the residual structure in the feature extraction network, introduces SimAM in the residual edge for processing, optimizes the neurons on the residual edge, improves the target feature representation and the ability of the network to detect small abnormal targets in the power transmission line, and improves the activation function of the YOLOX network, uses FReLU to enhance the nonlinear representation of small abnormal objects in the power transmission line, and uses the FReLU function in the residual module of the feature layer P3.

2. The power line anomaly target detection method based on the improved YOLOX algorithm according to claim 1, characterized in that, The image to be identified of the power transmission line anomaly is obtained by a camera carried by a drone, and is an image of the power transmission line in different positions, different weathers and different illuminations.

3. The power line anomaly target detection method based on the improved YOLOX algorithm according to claim 1, characterized in that, The preprocessing specifically comprises: normalizing the image, cropping all images to a resolution of 1368x912, labeling the abnormal target using labeme software, obtaining a power transmission line anomaly dataset, and dividing the dataset into a training set and a test set.

4. The power line anomaly target detection method based on the improved YOLOX algorithm according to claim 3, characterized in that, The power transmission line anomaly dataset comprises a high-voltage tower bird nest dataset and a self-explosion insulator dataset.

5. The power line anomaly target detection method based on the improved YOLOX algorithm according to claim 1, characterized in that, 3) specifically comprises: 301) dividing the input feature map into C feature sub-maps, and obtaining an N-dimensional vector of each feature sub-map through a fully connected layer; 302) using a Softmax activation function to constrain the N-dimensional vector to [0, 1], and then combining it with the corresponding feature sub-map as a weighted weight; 303) combining the C feature sub-maps into a new feature map containing all basic features of the power transmission line abnormal target, as an abnormal feature.

6. The power line anomaly target detection method based on the improved YOLOX algorithm according to claim 1, characterized in that, P3, P4 and P5 are respectively used to detect small targets, medium targets and large targets; the FReLU activation function is used for optimization in continuously extracting abnormal features, and is expressed as: wherein T(x) is a simple and effective spatial context feature extractor, and x represents an abnormal feature.

7. The power line anomaly target detection method based on the improved YOLOX algorithm according to claim 1, characterized in that, 5) specifically comprises: dividing the input into two branches along the channel direction to generate attention sub-maps, merging them after processing, and obtaining new feature maps of different feature layers.

8. The power line anomaly target detection method based on the improved YOLOX algorithm according to claim 7, characterized in that, The new feature maps of different feature layers include cross-channel information, direction information and position information of the abnormal target.

9. An electronic device comprising a memory and a processor, said memory having stored thereon a computer program, characterized in that, The program, when executed by the processor, implements the method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by the processor, implements the method of any one of claims 1-8.

Images