F-SRAF: a small target detection method based on attention mechanism and feature enhancement

Abstract

In target detection, small target detection is always a thorny problem, because there is little pixel information of small target in a picture. However, in the feature extraction phase of most target detection network architectures, the deep feature maps lack information on the edge or geometric details of small targets. Although the shallow feature map is rich in geometric details, it lacks the semantic features of small objects. To solve the above problems, this paper proposes a small target detection model F-SRAF, which combines feature enhancement and attention mechanism. Specifically, firstly, the feature enhancement module is used to extract credible geometric details by combining the shallow high-resolution features with the deep super-resolution semantic features. Then, the channel attention mechanism is added to the deep super-resolution semantic features to enhance the features useful to the target and suppress the background features. Finally, the related detail feature map learned by the network is fused with the semantic feature map added with the channel attention mechanism to output a high-resolution feature map with semantic information. In the experiment of this paper, our method has achieved good results on MS COCO dataset and Tsinghua-Tencent 100K dataset. At the same time, it is proved that this detection algorithm can enhance the semantic information of shallow feature map and has a good detection effect for small targets.