Research On Detection Algorithm For Multi-scale Small Targets Based On Improved Faster R-CNN

In recent years,with the rapid increase of the total mileage of national high-speed railways and transmission lines,the way of checking bird's nests and key parts on transmission lines by manual has caused great challenges to the investment of human and material resources,and it is difficult to ensure the timeliness and accuracy of detection results.Therefore,with the proposal of intelligent railway and smart grid,it has become an inevitable trend for the future development of power patrol inspection to use UAV-based aerial image target detection technology to detect bird's nests and key parts along transmission lines.Based on convolution neural network,this paper studies a multi-scale small targets detection algorithm for UAV aerial images based on improved Faster R-CNN.The main work and innovations are as follows:Firstly,by comparing VGG,Inception and ResNet three basic networks briefly,ResNet50,which has the highest cost performance,is selected as the basic network of this paper.On this basis,the realization principle of target detection network based on Faster R-CNN is emphatically studied.And after analyzing the scale of small targets on UAV aerial images,the main defects of Faster R-CNN in small targets detection are found.Secondly,aiming at the main defects of Faster R-CNN in small targets detection,the target detection algorithm based on Faster R-CNN is improved in many aspects.Through the improvement of RPN network,the problem that small targets can't be recalled due to the unreasonable design of initial proposal box is solved.By using Rol Align algorithm,the two successive quantization errors caused by using Rol Pooling algorithm is avoided,and the pooling accuracy of the proposed target area features is improved.By using deconvolution operation in ResNet₄f feature mapping layer of target proposal box,not only the resolution of the layer feature map is increased,but also the semantic information of the high-level convolution feature map is fused,thus the detection accuracy of small targets on UAV aerial images is improved significantly.In addition,by using global average pooling instead of full connection,the detection speed of the network is improved significantly.At last,the rationality and effectiveness of the improved detection network design methods proposed in this paper are verified on UAV aerial image nest test set and VOC2007 test set respectively.Then,in order to further improve the detection accuracy for multi-scale small targets of the UAV aerial images,this paper improves ResNet50 network structure by increasing network width and embedding SE module.At last,experiments are carried out on UAV aerial image nest test set and VOC2007 test set respectively,and the experimental results show that the two design methods can significantly improve the detection accuracy of the previous network model.Finally,this paper builds a deep learning experimental platform with Tensorflow,and designs and develops the interactive interface for multi-scale small targets detection of the UAV aerial images based on this platform,and realizes the visual operation of UAV aerial image targets detection,and provides the text calling interface.