改进ResNet50和FPN的多尺度目标检测算法研究  

Research on multi-scale object detection algorithm based on improved ResNet50 and FPN

作　　者：郭宝鑫 谢晓尧[1] 刘嵩[1] GUO Baoxin;XIE Xiaoyao;LIU Song(Key Laboratory of Information and Computing Science of Guizhou Province,Guizhou Normal University,Guiyang,Guizhou 550001,China)

机构地区：[1]贵州师范大学贵州省信息与计算科学重点实验室,贵州贵阳550001

出　　处：《贵州师范大学学报（自然科学版）》2024年第1期94-101,126,共9页Journal of Guizhou Normal University：Natural Sciences

基　　金：四川省重点研发计划项目(2021YFS0401;2022YFS0558)。

摘　　要：针对ResNet50和FPN结构无法将浅层的细节信息和深层的语义信息充分融合利用等问题,提出了一种改进ResNet50和FPN结构的算法,在ResNet50网络结构不同层次中引入了改进的通道和空间注意力模块,充分利用不同特征层的细节信息和语义信息。此外,在FPN结构中,为了能让浅层特征层更好的利用深层特征层的语义信息,在FPN自上而下的路径中,不同特征层之间增加了旁路来加强特征的重用。实验结果表明,在MS COCO数据集训练以后在PASCAL VOC 2012测试的均值平均精度(mAP)达到了83.2%,提升了2.7%,在MS COCO数据集上的mAP提升了1.5%,具有不错的检测性能。For the ResNet50 and FPN structures,the shallow detail information and the deep semantic information cannot be fully integrated and utilized.An algorithm to improve the structure of ResNet50 and FPN is proposed,introducing improved channel and spatial attention modules in ResNet50 network structure to take advantage of detailed and semantic information from different feature layers.In addition,in the FPN structure,in order to allow the shallow feature layer to better utilize the semantic information of the deep feature layer,in the top-down path of FPN,a bypass is added between different feature layers to enhance feature reuse.The experimental results show that the method performs well in PASCAL after training on the MS COCO dataset.The mean average precision(mAP)of the VOC 2012 test reached 83.2%,with an increase of 2.7%,and the mAP on the MS COCO dataset increased by 1.5%.It has good detection performance.

关 键 词：注意力机制 特征金字塔 特征重用 特征融合 特征层信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]