Object Detection for Optical Remote Sensing Images with Self-supervised Feature Representation

Automatically detecting objects of interest on remote sensing images is crucial for earth observations. Existing remote sensing object detectors mainly rely on supervised methods, and the quality and quantity of annotated samples determine the detection performance. However, obtaining large-scale labeled images is labor-intensive and requires domain expertise, which hinders the advancement of remote sensing object detection. To solve the problem, a method based on self-supervised feature representation is presented, with the goal of investigating how to utilize a large number of unlabeled remote-sensing images to enhance detection performance. The method contains three steps. Firstly, the presented method collects many unlabeled remote-sensing photos and reconstructs them to suppress the ineffective expression of background information while preserving the key features of the object. Then, object-level contrastive learning is used to acquire the generalized feature representation. Finally, the extracted feature expression is transferred to downstream task to improve the final detection performance. Experiment results show that with only a few training epochs on the NWPU VHR-10.v2 dataset, the proposed method outperforms supervised-only methods.