Detection Algorithm Of Untwisted Or Broken Strands Based On Four-channel Faster R-CNN

Electricity power has been an essential part in our modern society for a long time.The process of delivering power from power plant to end-users spans a really long distance which makes routine inspection and maintenance time-consuming,labor-intensive,and inefficient which we need to resolve.It is critical to get rid of untwisted strands to ensure the security of transmission system.Detection algorithm is a core part of maintenance process.However,the detection algorithm was considered and designed in a traditional digital image processing way in the past which has limited capability in a reality image content such as sophisticated background,different weather and illumination.This thesis introduces deep learning into detection algorithm,which provides a new idea for the detection.The beginning of the thesis demonstrates the background and significance of this research.Then it introduces the basic algorithms that commonly used in traditional detection methods and some theoretical knowledge of Faster R-CNN.Finally,this thesis proposes two algorithms for detecting untwisted or broken strands based on deep learning,which overcomes the lack of generalization of traditional detection algorithms and improves the performance of general deep learning methods.The main work of this thesis is as follows:Inspired by attention mechanism,this thesis proposes a four-channel Faster R-CNN which has an additional attention channel on input channel of the residual network.The data in the attention channel can guide the network to improve the recognition confidence of attention area and finally improve the model performance.Experiments show that the fourchannel method works.In order to solve the problem that the target is too small in the picture,a Faster R-CNN detection algorithm based on sliding cropping is proposed.The steps of sliding cropping are: place a crop window with a fixed size on the picture.After each crop,the window slides down a fixed distance.After sliding to the bottom of a column,it will switch to the beginning of the next column and repeat the above process until all the picture is cropped.Then,the cropped subimages will be sent to Faster R-CNN for training or detection.Finally the result is obtained by using result fusion and the improved NMS algorithm in this thesis.Sliding cropping can prevent the network from down-sampling input pictures,and retain the target information to the greatest extent.Experiments show that,based on the correct selection of sliding cropping algorithm parameters,the algorithm greatly reduces the number of false detections and improves network performance.The attention mechanism can improve the confidence of detected targets,and sliding cropping can reduce the number of false detections.The thesis proposed to integrate this two into a new detection algorithm,so that the two algorithms have complementary advantages.The experiment proves that the combination works.This thesis proposes two algorithms for the detection.From the P-R curve of the experimental results,their detection capabilities have been improved.During the detection process,there are still many points that need to be improved,such as insufficient data and inaccurate line segmentation of transmission line.In the future,we can start in these points to further improve the detection ability of the model.