基于改进Faster RCNN的少样本目标检测算法  

Few-Shot Object Detection Algorithm Based on Improved Faster

作　　者：杜芸彦 杨锦辉 李鸿 毛耀 江彧[1,2,3] RCNN DU Yunyan;YANG Jinhui;LI Hong;MAO Yao;JIANG Yu(Chinese Academy of Sciences,Key Laboratory of Optical Engineering,Chengdu 610000,China;Chinese Academy of Sciences,Institute of Optics and Electronics,Chengdu 610000,China;University of Chinese Academy of Sciences,Beijing 100000,China)

机构地区：[1]中国科学院光束控制重点实验室,成都610000 [2]中国科学院光电技术研究所,成都610000 [3]中国科学院大学,北京100000

出　　处：《电光与控制》2023年第5期44-51,共8页Electronics Optics & Control

基　　金：国家自然科学基金(61733012,61905253)。

摘　　要：当前大部分目标检测都依赖于大规模的标注数据集来保证其检测的正确率,而在实际场景中,大量数据的获取是十分困难的,且对数据的标注也需要花费大量人力物力。针对这一问题提出了一种基于Faster RCNN的少样本目标检测算法(CA-FSOD),在目标类别仅有少量标注样本的情况下,对目标样本进行检测。为了提高检测性能,首先提出了CBAM-Attention-RPN模块,减少无关候选框的数量;其次提出了全局-局部关系检测器模块,通过关联少量标注样本和待检测样本的特征,获取与目标类别更相关的候选区域;最后提出了基于余弦Softmax损失的分类器作为目标检测的分类分支,能有效地聚合同类别特征、降低类内方差、提高检测精度。为了验证所提算法,在MS COCO数据集上进行了训练和测试,实验结果表明,该方法的AP50为21.9%,优于目前一些少样本目标检测算法。At present,most object detection rely on large-scale annotation datasets to ensure the accuracy of detection.In the actual scene,it is very difficult to obtain a large amount of data,and it also takes a lot of manpower and material resources to annotate the data.To solve the problem,a Few-Shot Object Detection algorithm based on Faster RCNN,CA-FSOD,is proposed,which detects the object samples when there are only a few annotated samples in the object category.In order to improve the detection performance,a CBAM-Attention-RPN module is proposed to reduce the number of irrelevant candidate regions.Secondly,a global-local relation detector module is proposed to obtain candidate regions that are more related to the object category by associating the features of a small number of annotated samples and samples to be detected;Finally,a classifier based on cosine Softmax loss is proposed as the classification branch of object detection,which can effectively aggregate the same category features,reduce the intra-class variance and improve the detection accuracy.In order to verify the proposed algorithm,it is trained and tested on MS COCO dataset.The experimental results show that the AP 50 is 21.9%for this method,which is better than that of some existing few-shot object detection algorithms.

关 键 词：目标检测 少样本学习 少样本目标检测 Faster RCNN 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]