Research On Small Target Detection Algorithm Based On Deep Learning

Due to the problems of low resolution,disorientation and complex background of small targets in the image data sets such as aviation,medical treatment and autopilot,most of the target detection algorithms have a single feature map scale,which can not effectively integrate the feature semantic information,and the feature extraction network produces errors after multiple sampling,which results in the loss of the edge feature information of small targets in the image,resulting in low recognition rate and false detection Missed inspection and other problems.In order to solve the problem of insufficient target feature information in SSD(single shot multibox detector)model,this paper proposes a method of deconvolution and feature fusion.In order to solve the problem of low resolution of the underlying feature map conv4?3 of SSD,the deconvolution method is used to increase its resolution and map the ROI features.Then,the new feature map is spliced with the convolution layer conv10?2 after the upper sampling.Finally,the feature map of the inherent four scales of SSD is fused into multi-scale detection.By testing the improved method in coco,Pascal VOC and other data,the detection accuracy is significantly improved compared with SSD method.In addition,this paper proposes a parallel high-resolution hrnet combined with LSTM feature extraction network.Firstly,the method of constructing parallel high-resolution feature map in hrnet is used to construct four parallel subnets.Then,the channel number of the feature map obtained by each subnet is spliced to form multi-scale detection.Finally,the sequence feature information before and after the fusion of two-way LSTM is used to improve the feature expression.This network replaces vgg16 network in SSD method,and tests in coco,ucas-aod and Kitti data sets.The average detection accuracy of this network is 41.6%,69.8% and 69.4% respectively,which is significantly higher than SSD,dssd and other methods.The results show that the two different methods proposed in this paper can effectively reduce the missed detection rate of small targets and improve the average accuracy of the overall target.