An Efficient Method For Tiny Object Detection

Small object detection is a hot research topic.The main challenge of small object detection is that it is difficult to extract effective feature information due to the lack of pixels for small objects.For instance,small and medium-sized objects in dataset of COCO exist in about a half of the images,which leads to the failure of the network to learn the features of small objects in about a half of the time during training.This problem leads to low detection accuracy of small objects.With the rapid development of the deep learning technology,small object detection based on deep learning has made great progress.A series of methods have been proposed to improve the performance of small object detection in images from the aspects of network structure,training strategy,data processing,et al.In this thesis,a review is made for small object detection methods.Existing works are categorized and their characteristics and performance are analyzed.In addition,small object detection dataset is introduced.Because the anchors in the shallow layers are sparse,an anchor densification scheme is adopted in this thesis.For each center of the anchors,the anchor density is calculated.Then,the anchors are densified to make anchor density of the small object be consistent with the large object to improve the detection performance.Experimental results on dataset of MS-COCO 2017 with different object detection methods(Faster R-CNN,Retina Net)with different backbone networks(Res-50-FPN,Res-101-FPN)demonstrate the effectiveness of the proposed anchor densification scheme.Additionally,a novel anchor-free two-stage object detection algorithm is proposed in the thesis.In the first stage,region proposals are produced via corner points extracted based on Cornet Net.In order to improve the perception ability for the object internal information,center pooling is introduced to enhance the representation ability for the features of the internal regions,and the internal key points are detected.A large number of false-positive proposals can be filtered out by checking whether the internal key points exist in the internal area.In the second stage,the kept proposals are fed into a multivariate classifier to obtain the final result.The proposed algorithm has been tested on the data set of MS-COCO with an accuracy of 47.6%,which is a competitive result compared to that of the state-of-the-art object detection methods.The proposed algorithm outperforms Corner Net by 7.1% in accuracy.For the objects with special(huge,tiny,or large aspect-ratio)shapes,higher accuracy increments can be obtained which demonstrates the effectiveness of the proposed algorithm.The main contributions of this work include:(1)In order to solve the problem of insufficient matching of anchors in the detection model,an anchor densification strategy is proposed.Combining this strategy can effectively improve the detection performance of various current anchor-based detection algorithms for small objects;(2)The concept of centrality is introduced into the detection network based on keypoint detection.Compared with the current method based on corner points,it can effectively improve the perception ability of the network to the internal information of the object,so as to obtain the region proposals of the object more efficiently and accurately.(3)The anchor-free two-stage object detection algorithm combining the region proposal strategy based on the key points and the multivariate classifier is proposed,which can effectively improve the detection performance.Especially for the objects with special shape(huge,tiny or extremely high aspect ratio),the accuracy improvement is more obvious.