Few-Shot Object Detection with Contrastive Enhancement Loss Guided by Object Association Graph in Remote Sensing Images

Object detection in remote sensing images is widely used in military reconnaissance and environmental monitoring, but the high cost of obtaining large-scale labeled data limits the popularity of deep learning methods. Although few-shot learning methods that utilize a large number of annotated remote sensing images for auxiliary training are regarded as effective solutions, in practical applications, data scarcity leads to difficulties in learning discriminative features for novel classes and insufficient object feature representation, which seriously affects the detection performance of the model. To address these issues, we propose a few-shot object detection with contrastive enhancement loss guided by object association graph (OAG-FSOD). Aiming at the class-agnostic characteristics of proposals generated by the region proposal network, this method models the relationships between proposals through object association graph to guide contrastive learning in generating positive and negative pair labels. Subsequently, by maximizing the consistency of features from the same object proposal while minimizing the consistency of features from different object proposals, discriminative contrastive representations are learned. Additionally, the semantic discriminability of novel class features is further enhanced via contrastive enhancement loss. The OAG-FSOD method effectively strengthens the feature representation of objects, thereby improving the performance of few-shot object detection in remote sensing images. Extensive experiments on two commonly used remote sensing object detection datasets, DIOR and MAR20, validate the effectiveness of the proposed method.