Small Target Detection Algorithm Based on Two-Stage Feature Extraction

In the field of target detection, the traditional general algorithms based on open data sets such as Imagenet and COCO can not detect small targets well because of their small size and insignificant features. Considering that the feature information of small targets is mostly retained in the shallow layer of the network, this paper adds a detection head on the shallow layer of the backbone network on the basis of the general target detection algorithm, and introduces the shuffle attention mechanism to deeply integrate the feature information from different layers, so as to enhance the expression ability of the network for small targets. Compared with the original network, mAP has improved by 6.25% on the VisDrone dataset. At the same time, considering the difficulty of extracting the feature information of small targets in single-stage detection, we propose a two-stage detection framework that can be used on any detection algorithm. We process the images into blocks according to the size characteristics of the targets, carry out interpolation and redetection on the region of interest with low confidence, which makes the recall ability of small targets significantly improved. When the inference confidence threshold is 0.5, compared with the original network, the performance is improved by 4.24
^%
, which can significantly improve the ability of small targets detection.