Feature-enhanced small-target detection

Abstract

Detecting small targets from images is still a challenging problem in computer vision due to the limited size, few appearance and geometric cues, and the lack of large-scale small target datasets. To address this problem, an adaptive feature-enhanced target detection network (YOLO-FENet) is proposed to improve the detection accuracy of small targets. Firstly, an improved adaptive two-way feature fusion module is designed by introducing a feature fusion factor to make full use of the feature maps of various scales to improve the feature expression ability of the network; secondly, a spatial attention generation module is proposed by combining the characteristics of the network, which improves the feature localization ability of the network by learning the positional information of the region of interest in the image. The experimental results on the UAVDT dataset show that the average precision (AP) of the proposed YOLO-FENet is 6.3 percentage points higher than that of the pre-improvement YOLOv5, and it is also better than other target detection networks.