Research On Target Detection Algorithm Based On Deep Learning

In the field of target detection,there is a problem of lack of background samples(regions not belonging to any object class in the picture).In the case of a more complex background or high similarity of targets,the detection capability of the detector will be limited.In addition,the quality of the data directly affects the final result of the test model.This paper is based on the Faster RCNN detection framework,which is mainly improved from difficult sample mining and data equalization sampling to improve the performance of the detector.The main tasks include the following:1.Difficult sample mining.This paper introduces online hard examples mining(OHEM)in the framework of Faster RCNN detection,and improves the mining method.In the feature extraction stage,the Residual Nets(ResNets)network with strong feature extraction capability was used to mine the difficult cases through an up-convolutional network to form a new network architecture FOHEM,which combined the context information and expanded the receptive field at the same time.Experimental results show that the network framework can improve the detector's detection capability.2.Data equalization sampling method.This paper first studies the effects of different sampling methods on target detection.Based on the existing problems,we increase the richness of the whole sample set by balancing the intra-class sample size,and eliminate the influence of the intra-class sample size.Finally,the effect of the method was verified by experiments.