EFR-ACENet: Small object detection for remote sensing images based on explicit feature reconstruction and adaptive context enhancement

Small object detection in remote sensing images has been a major challenge in the research field of remote sensing technology. Traditional deep learning models perform well in detecting multiple objects, but lack in detecting small object detection, resulting in decreased detection accuracy. In order to reduce the information loss of small objects in object detection as well as to effectively manage and utilize different features in multiple receiver domains, this study proposes an architecture for a small object detection network for remote sensing images based on explicit feature reconstruction and adaptive context enhancement (EFR-ACENet). An explicit feature reconstruction module is designed to preserve feature details in the image and reduce the loss of small object information. Also, an adaptive context enhancement module is introduced, which is capable of detecting objects more comprehensively by integrating features such as contextual environment. And the features of perception fields of different sizes are extracted by the adaptive perception capability. We evaluate EFR-ACENet against ten state-of-the-art algorithms on multiple datasets such as the dataset for Tiny Object Detection in Aerial Images (AI-TOD) and the dataset for object Detection in Optical Remote sensing images (DIOR). The experimental results show that EFR-ACENet improves the average precision by3.3 % and 13.6 % on the AI-TOD and DIOR datasets respectively, which fully demonstrates its effectiveness in small object detection tasks. This result signals the great potential of EFR-ACENet for future airborne real-time applications.