To address the problem of inaccurate detection caused by scale variation and poor feature correlation in the few-shot scenario,a few-shot object detection algorithm based on multi-scale fusion and contrastive mechanism was proposed.Unlike previous methods that are limited to surface feature transfer,the potential correlation between the feature space of base classes and novel classes was deeply investigated.Multi-scale recursive projection fully invoked feature correlation.By recursively pro-jecting features across multiple scales to increase feature correlation,and leveraging contrast mechanisms to fully exploit base class space and channel information,the extraction,screening,and matching of novel class features were maximized,resulting in a significant performance improvement.Superior performance is demonstrated in Pascal VOC datasets and MS COCO datasets,which provides new theoretical support and research solutions for the study of few-shot object detection.
维普
万方数据