Few-shot object detection algorithm based on improved faster R-CNN

The task of object detection requires extensive, high-quality image datasets to acquire the ability to recognize objects in images and determine their positions. By utilizing rich and diverse datasets, model can achieve better generalization capability and mitigate overfitting issue. However, in real-world industrial settings, annotating datasets requires a significant investment of human and material resources. Addressing these challenges, this paper proposes a few-shot object detection algorithm (SFA-FSOD) based on Faster R-CNN. Firstly, the SimAM-FPN module is proposed, which improves the quality of the Region Proposal Network in locating Regions of Interest by integrating multi-scale features and attention mechanisms into the feature pyramid network, thereby reducing the number of irrelevant candidate boxes. Secondly, the Feature Aggregation (FA) module based on prototypical networks is proposed, aggregating crucial regions around instances to generate basic feature for each category. Finally, Euclidean distance is employed to select and fuse the most similar basic feature from the feature prototype library, thereby enhancing weak features in the few-shot object detection task. Extensive experiments on the Pascal VOC and MSCOCO datasets validate the effectiveness of SFA-FSOD.