Research On Object Detection Based On Improved Classification And Regression

Object detection is one of the most important and challenging branches of computer vision,which is widely used in security monitoring,unmanned driving,vision robot and other people's lives.In recent years,with the rapid development of deep learning networks for detection tasks,the performance of object detector has been greatly improved.But there are still too many challenges.For example,in the case of blurred images,multi-target overlapping,and too small scale,classification and regression are too difficult.This paper focuses on the existing problems of single-stage detection algorithm,Combined with domestic and foreign research progress,the classical algorithm framework and the leading research results.The main work is as follows:(1)An object detection method based on selective two-stage classification regression is proposed.For the speed and accuracy imbalance between the single-level detection algorithm and the two-level detection algorithm,this paper readjusts the twostage classification and regression tasks in the Refine Det network.First,the network becomes more efficient by selectively performing only classification tasks at the lower detection level and regression tasks at the higher detection level.Secondly,in order to solve the problem of misalignment between features and anchor frames existing in the original algorithm and most single-stage detection algorithms,a deformable convolution Ro IConv is introduced,which achieves the purpose of feature alignment by directly calculating the offset between features and anchor frames.Finally,this paper conducted related experiments on the MSCOCO dataset and Pascal VOC dataset.The results show that,compared with the original algorithm and most single-level detection algorithms,the algorithm in this paper has higher accuracy.It is important that these improvements will hardly cause network computing burden,so the speed of network inference is still real-time.(2)A vehicle detection algorithm based on the improved residual classification network is proposed to improve the classification network which is not suitable for object detection.Traditional classification networks are specially designed for image classification tasks.In order to solve this problem,an improved residual classification network structure called D-Res Net-50 was proposed,which was optimized for regression tasks and improved classification tasks significantly.On the basis of improving the structure of residual classification network,on the one hand,selective classification and selective regression are used.On the other hand,random occlusion enhancement strategy is used in the preprocessing stage.The evaluation results on KITTI data set show that the algorithm in this chapter is superior to most classical detection algorithms in vehicle detection accuracy.The network processing time of single image detection is only 0.3 seconds,which can meet the real-time requirements and has certain practical application value.