A recursive attention-enhanced bidirectional feature pyramid network for small object detection

Single Shot MultiBox Detector (SSD) method shows outstanding performance by using multiscale feature maps in object detection task. However, the SSD method exhibits low accuracy in small object detection. In this paper, A Recursive Attention-Enhanced Bidirectional Feature Pyramid Network (RA-BiFPN) is proposed. Firstly, we designed the attention-enhanced bidirectional feature pyramid network (A-BiFPN) to improve the detection accuracy of the small object. The A-BiFPN is composed of bidirectional feature pyramid network (BiFPN) and the coordinate attention. Among them, the BiFPN employs top-down and bottom-up paths to aggregate features at different scales so that features at all scales contain rich semantic and detailed information. These features help coordinate attention that embeds positional information into channel attention so that the network can easily focus on the channels and locations related to the object in the feature map. Secondly, in order to enhance the ability of the A-BiFPN to characterize small targets, we adopted the recursive structure to feed back the output feature of the A-BiFPN into the backbone network. In this way, the recursive structure goes through the bottom-up backbone repeatedly to enrich the representation power of the A-BiFPN. The experimental results show that the detection accuracy of our method in PASCAL VOC, NWPU VHR-10 , KITTI and RSOD dataset is improved by 2.65%, 7.98% ,7.02% and 5.63% respectively compared to the original SSD.