Research On Small Target Detection Technology Based On Improved Faster RCNN

In recent years,target detection algorithms are developing rapidly.At present,it is widely used in face detection,video surveillance,medical diagnosis and intelligent transportation.These applications have completely changed people's lives.At this stage,the object detection technology based on deep learning has excellent detection effects in general detection tasks.However,the effect of small target detection is not ideal.In this paper,we focus on the research of small object detection technology based on the deep separable grouped convolutional Res Net based on multiple convolution kernels and improved RPN.An improved Faster RCNN algorithm is proposed.The main research contents include:A deep separable packet convolution Res Net based on multiple convolution kernels.Based on the analysis of the small target feature extraction network,Res Net based on deep separable convolution is designed as the feature extraction network.On this basis,the topology of multiple convolution kernels is used to enhance the adaptability of the network;The convolution form is improved by using the method of block convolution to compress the amount of parameters and computation in the network;After packet convolution,an improved channel shuffling is proposed to enhance the feature information exchange between different packets.In addition,the use of mish activation function can avoid the collapse of low dimensional data.Finally,combined with the residual connection form,a deep separable packet convolution Res Net(MDWSR)with multiple convolution cores is constructed.Experiments are carried out on the DOTA data set of aerial images.Compared with traditional Res Net,the Top 1 error rate and Top 5 error rate of MDWSR are reduced by 3.34% and1.56%.The complexity of the model is reduced by 23%.Compared with other network models,it also has significant advantages.Small object detection algorithm based on improved RPN.Firstly,DCP clustering algorithm is used to redesign the anchors in RPN network.The design of anchors based on kl?DCP is proposed.The difference measurement of kl dispersion is introduced into DCP clustering algorithm.The new design of anchors is more suitable for small target size and proportion candidate box.Then,in the post-processing of RPN network candidate frames,a Softer-NMS algorithm with adaptive threshold is proposed.The loss function is designed as kl loss.Regression adds a branch prediction standard deviation and introduces the standard deviation into the loss function.By predicting the density around the candidate box,different threshold suppression strategies are adopted.Finally,the weighted average processing and coordinate updating of candidate frames are performed.The improved RPN is compared and tested on the DOTA?6 data set.Compared with the traditional RPN network,the m AP of the improved RPN network is increased by 6.76%.FPS also achieves about 42.87 frames per second.The improved strategy for small object detection proposed in this paper uses MDWSR as the feature extraction network and the improved RPN as the region suggestion network to obtain the final improved Faster RCNN algorithm.Compared with the traditional Faster RCNN,the improved algorithm m AP has increased by 13.98%.FPS reached 43.25 frames per second.Compared with other networks such as R-FCN and YOLO V3,the algorithm in this paper also has better detection results.It proves the effectiveness and feasibility of the improved strategy in this paper on the task of small object detection.