This study aims to develop a deep learning model to improve the accuracy of identifying tiny targets on high resolution remote sensing (HRS) images. We propose a novel multi-level weighted depth perception network, which we refer to as MwdpNet, to better capture feature information of tiny targets in HRS images. In our method, we introduce a new group residual structure, S-Darknet53, as the backbone network of our proposed MwdpNet, and propose a multi-level feature weighted fusion strategy that fully utilizes shallow feature information to improve detection performance, particularly for tiny targets. To fully describe the high-level semantic information of the image, achieving better classification performance, we design a depth perception module (DPModule). Following this step, the channel attention guidance module (CAGM) is proposed to obtain attention feature maps for each scale, enhancing the recall rate of tiny targets and generating candidate regions more efficiently. Finally, we create four datasets of tiny targets and conduct comparative experiments on them. The results demonstrate that the mean Average Precision (mAP) of our proposed MwdpNet on the four datasets achieve 87.0%, 89.2%, 78.3%, and 76.0%, respectively, outperforming nine mainstream object detection algorithms. Our proposed approach provides an effective means and strategy for detecting tiny targets on HRS images.
© 2023. Springer Nature Limited.