递归投影融合对比机制的少样本目标检测方法

Few-shot object detection based on recursive projection fused contrast mechanism

针对少样本场景中尺度混乱、特征关联性差导致检测不精准的问题,提出一种基于多尺度融合对比机制的检测算法。相比先前方法仅关注表层特征迁移,该方法深刻探讨基类与新类特征空间的潜在联系。通过多尺度递归投影增加特征关联性,利用对比机制充分挖掘基类空间和通道信息,最大化引导新类特征的提取、筛选以及匹配,取得显著性能提升。在Pascal VOC和MS COCO数据集实验中验证了该方法的优越性,为少样本目标检测研究提供了新的理论支撑和研究方向。

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 projecting 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.