Research On Small Target Detection Algorithm Based On Deep Learnin

With the development of deep learning,major breakthroughs have been made in object detection,which are used in important occasions such as airports,railway stations,and ports to complete tasks such as video surveillance,face recognition,and intelligent transportation,and have achieved good results.However,in the field of small object detection,the ideal effect cannot be achieved.Small object detection can be said to be a "hard bone" in the field of object detection,and it is still a very challenging task.In this paper,a series of researches are carried out to improve the performance of small object detection,and several improvement methods are proposed which based on some existing excellent object detection methods.The main research contents of this paper are as follows:1.Aiming at the difficulty of small object feature extraction and the weak expression ability of the extracted shallow features,this paper proposes a shallow feature enhancement method.The receptive field is expanded by paralleling multiple dilated convolutions with different dilation rates,to obtain richer contextual information for enhancing shallow feature representation.In addition,to solve the problem that the general feature fusion method is not effective,a quadratic feature fusion method is proposed.The method uses top-down feature fusion to transfer deep rich semantic information to shallow layers in the first round of feature fusion,and uses bottom-up feature fusion to transfer rich and detailed information from shallow layers to deep layers;in the second feature fusion The feature maps of the corresponding scales of the two types of fused feature maps obtained by the first round of feature fusion is fused by convolution,so as to obtain features that contain both rich semantic information and rich detailed information.Finally,it is verified on the PASCAL VOC and COCO datasets.The results show that the shallow feature enhancement method and the quadratic feature fusion method can effectively improve the performance of small object detection.2.The Anchor-based object detection method often needs to carefully design the anchor frame,and then fine-tune the candidate frame during the detection,which makes the Anchorbased method poor generality and unfriendly to small object detection.Aiming at the problems of the Anchor-based method,and referring to the Corner Net method,an Anchor-free two-stage object detection method based on corner point and centripetal offset is proposed.In the first stage,the upper left and lower right corner key points are obtained through improved corner pooling and corner prediction.Since corner pooling can only perceive the information near the boundary of the object bounding box,the improved corner pooling method is used to perceive the internal information of the object can improve the positioning accuracy of corner points,and exhaust possible candidate boxes according to the position of the corner points.At the same time,predict the corresponding centripetal offset for each corner point,calculate the corresponding center point position of the corner point through the corner point and the centripetal offset,and determine whether the center point corresponding to the corner point of the candidate frame falls in the center area.to filter out false positives.The second stage uses a multivariate classifier to classify and regress the candidate boxes that survived the first stage.With the Anchor-free two-stage method,setting the hyperparameters of the anchor box can be avoided.Finally,the performance is evaluated on the COCO dataset.The experiments show that our algorithm greatly improves the performance of object detection,and achieves excellent performance in small object detection.