摘要
针对现有小样本目标检测模型存在的尺度变化问题,支持集与查询集之间的外观变化、遮挡导致的误检与漏检问题,本文提出一种融合多尺度特征与注意力的小样本目标检测模型.首先,采用ResNet-101网络进行特征提取,同时引入ASPP(Atrous Spatial Pyramid Pooling)模块获取不同的感受野,以捕获目标细节信息的多尺度特征.其次,采用Bi-FPN网络进行多尺度特征融合,获得更具代表性的查询特征与支持特征,有效缓解尺度变化问题.然后,利用提出的注意力引导特征增强模块对查询特征与支持特征进行自身关注,使得它们具有更好的判别能力,由此促进查询特征与支持特征的融合,以更好地应对外观变化和遮挡带来的挑战,从而缓解误检、漏检问题.最后,将分类头与边界框回归头进行解耦,分别对RPN网络基于细粒度查询特征产生的候选区域进行目标分类与目标定位.在PASCAL VOC与MS COCO数据集上的实验结果表明,所提模型的检测性能优于主流的小样本目标检测模型,相较于基线模型DCNet,mAP平均分别提升了3.5%与2.1%.
Aiming at the problems of scale variation,the false detection and missed detection caused by the appearance change between the support set and the query set and occlusion in the existing few-shot object detection model,a few-shot object detection model that integrates multi-scale features and attention is proposed in this paper.Firstly,the ResNet-101 network is used for feature extraction,and the ASPP(Atrous Spatial Pyramid Pooling)module is introduced to obtain different receptive fields,so as to capture the multi-scale features of the object detail information.Secondly,the Bi-FPN network is used for multi-scale feature fusion to obtain more representative query features and support features,which can effectively alleviate the problem of scale change.Thirdly,the proposed attention-enhanced feature enhancement module is used to pay attention to the query features and support features,so that they have better discriminative ability,thereby promoting the fusion of query features and support features.This enhances the model’s ability to cope with the challenges brought about by appearance changes and occlusions,so as to alleviate the problems of false detection and missed detection.Finally,the classification head and the bounding box regression head are decoupled,and the object classification and object location are respectively performed on the candidate regions generated by the RPN network based on fine-grained query features.The experimental results on the PASCAL VOC and MS COCO datasets show that the detection performance of the proposed model is better than the mainstream few-shot object detection models.Compared to the baseline model DCNet,the mAP increases by an average of 3.5%and 2.1%,respectively.
作者
张英俊
甘望阳
谢斌红
张睿
ZHANG Yingjun;GAN Wangyang;XIE Binhong;ZHANG Rui(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)
出处
《小型微型计算机系统》
北大核心
2025年第3期689-696,共8页
Journal of Chinese Computer Systems
基金
山西省基础研究计划面上项目(20210302123216)资助
吕梁市引进高层次科技人才重点研发项目(2022RC08)资助.
关键词
小样本学习
元学习
目标检测
多尺度特征融合
注意力机制
few-shot learning
meta-learning
object detection
multi-scale feature fusion
attention mechanism
