Feature enhancement modules applied to a feature pyramid network for object detection

A feature pyramid network (FPN) improves the ability of an object detection model to detect multiscale targets. However, the simple upsampling used in an FPN is not conducive to the propagation of deep semantic information, and redundant background information is not conducive to object detection. In this paper, we propose two plug-and-play modules for preexisting FPN-based architectures: a channel filtering module (CFM) and a spatial filtering module (SFM). The CFM learns the correlations between channels to improve the feature maps obtained via upsampling. The SFM introduces global information to improve the detection performance of the network. With the CFM and SFM, we improve the average precision (AP) of Faster R-CNN with an FPN by 0.9% to 1.3% on COCO, and we boost the AP of YOLOv5s with PANet by 2.8%.