Small-Object Detection in Remote Sensing Images With Super-Resolution Perception

Abstract

Small objects are widely distributed on remote sensing images (RSIs), and most of them are achieved by super-resolution (SR) reconstruction followed by detection. However, due to the independent training of the SR network and the detection network, the lack of interaction between them leads to the limited performance of small object detection (SOD). Furthermore, time accountability is increased since the SR task is performed before the detection task. To address these problems, we develop a new SOD network to improve the SOD performance in RSIs, which embeds the SR task into the SOD task. First, a channel attention weighting module is proposed before the backbone to assign weights to different channels of the input image, allowing the network to selectively focus on different channels. Second, a self-attention encoding module is designed between the backbone and neck to add self-attention weight to the features extracted from the backbone and enhance the feature representation ability of small objects. Finally, the SR perceptual branch and perceptual loss are designed so that the SR task and the detection task can be associated through the SR perceptual loss, and the SR perceptual branch can guide the backbone network to learn high-resolution features through joint training, thus improving the detection performance of small objects. In the inference phase, the SR perceptual branch has been removed to improve the speed. Extensive experimental results on VEDAI and DOTA datasets show that the proposed method achieves an accuracy of 82.63% and 78.63%.

Keywords

• Attention mechanism
• remote sensing images (RSIs)
• small object detection (SOD)
• super-resolution (SR)
• YOLOV5