FE R-CNN: Feature Enhance R-CNN for Few-Shot Ship Object Detection

With ongoing advancements in navigation technology, ship object detection has emerged as a focal point of current research. However, in real-world scenarios, annotated ship samples are often insufficient. To address these challenges, this paper proposes the Feature Enhance R-CNN (FE R-CNN) algorithm, aimed at enhancing the performance of ship object detection in few-shot scenarios. Firstly, the algorithm utilizes the Feature Refine Feature Pyramid Network (FR-FPN) module to improve upsampling and eliminate the aliasing effect. It leverages bidirectional cross-scale channels to comprehensively extract global multi-scale and semantic information. In the model fine-tuning phase, the Enhance Region Proposal Network (ERPN) module is proposed to effectively reduce the imbalance among the candidate regions for base and novel classes. Simultaneously, the Comparative Learning Branch (CLB) is proposed to increase the difference between features of different classes and the similarity between features of the same class. The synergistic effect of these modules significantly enhances the model's accuracy in locating few-shot objects and its capability in multi-scale object detection. Experiments on our custom maritime ship dataset MSD and VOC2012 show that our method outperforms various advanced algorithms, enhancing the model's applicability in real-world few-shot ship object detection scenarios.