Few-Shot Object Detection on Remote Sensing Images Based on Decoupled Training, Contrastive Learning, and Self-Training

Few-shot object detection (FSOD) in remote sensing imagery faces two critical challenges compared to general methods trained on large datasets: (1) Only a few labeled instances leveraged as the training set significantly limit the feature representation learning of deep neural networks; (2) Remote sensing image data contains complicated background and multiple objects with greatly different sizes in the same image, which leads the detector to large numbers of false alarms and miss detections. This paper proposes a FSOD framework (called DeCL-Det) that applies self-training to generate high-quality pseudo-annotations from unlabeled target domain data. These refined pseudo-labels are iteratively integrated into the training set to expand supervision for novel classes. An auxiliary network is introduced to mitigate label noise by rectifying misclassifications in pseudo-labeled regions, ensuring robust learning. For multi-scale feature learning, we propose a gradient-decoupled framework, GCFPN, combining Feature Pyramid Networks (FPN) with a Gradient Decoupled Layer (GDL). FPN is to extract multi-scale feature representations, and GDL is to decouple the modules between the Region Proposal Network (RPN) and RCNN head into two stages or tasks through gradients. The two modules, FPN and GDL, train Faster R-CNN in a decoupled way to facilitate the multi-scale feature learning of novel objects. To further enhance the classification ability, we introduce a supervised contrastive learning head to enhance feature discrimination, reinforcing robustness in few-shot object detection. Experiments on the DIOR dataset indicate that our method performs better than several existing approaches and achieves competitive results.