Cross-Layer Attention Network for Small Object Detection in Remote Sensing Imagery

In recent years, despite the tremendous progresses of object detection, small object detection has always been a challenge in the field of remote sensing. The main reason is that small objects cover few features that are easily lost during down-sampling. In this article, we propose a cross-layer attention network aiming to obtain stronger features of small objects for better detection. Specifically, we designed an up-sampling and down-sampling feature pyramid to obtain richer context information by bidirectionally fusing deep and shallow features, as well as skipping connections. Moreover, a cross-layer attention module is designed to obtain the nonlocal association of small objects in each layer, and further strengthen its representation ability through cross-layer integration and balance. Extensive experiments on the publicly available datasets (DIOR dataset and NWPUVHR-10 dataset) and the self-assembled datasets (SDOTA dataset and SDD dataset) show the excellent performance of our method compared with other detectors. Moreover, our method achieved 74.3% mAP on the public DIOR dataset without any tricks.