RST R-CNN: a triplet matching few-shot remote sensing object detection framework

Detecting remote sensing objects on few-shot datasets remains a challenging in the domain of object detection and recognition. As a feasible solution to this challenge, RST R-CNN, a novel approach for detection of remote sensing objects based on meta-learning strategy and has high detection accuracy on few-shot remote sensing objects is put forth in the present work. Relying on the meta-learning strategy, the proposed method turns out to have a high accuracy on few-shot remote sensing objects. In the proposed method, the feature enhancement module will precisely capture the shape features of the objects, and the respective field enhance module enlarges the respective field in feature extraction. Meanwhile, a feature interaction module is introduced to compute similarities between the support and query sets, thereby enhancing the quality of candidate box generation in few-shot object detection tasks. Besides, a triplet matching detector is designed, which can improve the model’s performance on new classes of data by making the data in the same class move closer and those in different classes to move farther from each other. Comparative experiments have been performed on two datasets—NPWU VHR-10 and the DIOR between our method and some state-of-the-art (SOTA) models to verify the effectiveness and superiority of our model Our model is found to have outstanding performance on both datasets and have outperformed the other methods. Ablation experiments are conducted, which prove the self-developed modules have contributed to improved performance of our proposed model. The research here is expected to provide new solutions to object detection tasks in remote sensing scenarios.