Research On Key Object Detection Of Transmission Lines Based On Deep Supervised Neural Network

In the transmission line scene,key objects such as dampers,insulators,and wires that are damaged indirectly affect the normal and stable operation of the power supply system.Regular inspection of transmission lines can effectively avoid power supply system failure.Currently,UAVs are usually used to inspect transmission lines,take pictures and save images of all inspected transmission lines.Finally,manually inspect all these pictures for defective devices.Therefore,to improve the efficiency of key objects defect inspection of power transmission lines,it is necessary to study the detection of key objects of power transmission lines to achieve the goal of rapid positioning of key objects of power transmission lines,and ultimately improve the efficiency of inspection of power transmission lines.In the research of key object detection in transmission lines,the traditional object detection methods,which are mainly based on manual features,can only target specific types of objects with easily distinguishable features,and are not so robust.The mainstream object detection algorithms based on deep neural networks make up for the shortcomings of traditional object detection methods and improve the robustness of object detection algorithms.However,in the research of key object detection of transmission lines,the following problems still exist when using deep-supervised neural networks: First,there is a lack of sufficient training data;second,the shallow features of the network are not effectively used.Third,object space positioning is not accurate.Therefore,this thesis mainly solves the above three problems as the starting point of the research,the main work is as follows:(1)After referring to the VOC official Annotation guidelines,a dataset containing23024 key objects of the transmission line was established.A scheme for augmenting image data using PCA Jittering and perspective transformation algorithms is proposed,and a corresponding annotated object position conversion algorithm is proposed to solve the problem of lack of training data.(2)Inspired by FPN and Multi-layer Convolutional Feature Fusion Networks,we combined the residual block to innovatively propose a shallow residual projection structure,thereby improving the CNN’s high-resolution feature representation of key objects of the transmission line,especially improved detection performance for small objects.(3)To optimize the RPN of Faster R-CNN,we propose a parameter fine-tuning optimization strategy,which is suitable for our dataset.We fine-tuned the anchor box parameters and the positive and negative sample ratio parameters in the RPN to improve the performance of the transmission line key object detection algorithm.(4)We studied the IoU loss function used for positioning correction,and found that the IoU loss function has the disadvantage that the network parameters cannot be optimized.To improve the shortcomings of the IoU loss function,an IoUC loss function is proposed,which solves the problem that parameters cannot be learned during network training.In this thesis,the augmented dataset,CNN network with shallow residual projection structure,fine-tuned RPN parameters,and IoUC positioning loss function are incorporated into Faster R-CNN to complete comparative experiments.We found that the m AP of wires,insulators,and dampers has been improved.The m AP of key object detection of all power transmission lines is increased by 69.7%.