基于区域预推荐和特征富集的SOD R-CNN交通标志检测网络  

SOD R-CNN Traffic Sign Detection Network Based on Region Proposal and Feature Enrichment

作　　者：周楝淞[1] 邵发明 杨洁[3] 彭泓力 李赛野 孙夏声[1] ZHOU Liansong;SHAO Faming;YANG Jie;PENG Hongli;LI Saiye;SUN Xiasheng(No.30 Institute of CETC,Chengdu Sichuan 610041,China;Army Engineering University of PLA,Nanjing Jiangsu 210007,China;Sichuan Electric Power Corporation,Chengdu Sichuan 610016,China)

机构地区：[1]中国电子科技集团公司第三十研究所,四川成都610041 [2]陆军工程大学野战工程学院,江苏南京210007 [3]四川省电力公司,四川成都610016

出　　处：《信息安全与通信保密》2024年第10期115-126,共12页Information Security and Communications Privacy

基　　金：四川省科技厅重大专项(MWA22Y307)。

摘　　要：基于区域的快速卷积神经网络存在资源的浪费和无法有效应对小目标检测的问题,提出基于高可能性区域推荐网络及特征富集的区域的小目标检测卷积神经网络架构。首先,采用区域推荐网络对锚点区域进行筛选,节约分类阶段的处理时间,提高了系统的处理速度。其次,为了解决无法有效检测小目标的问题,提出了融合视觉几何组16层网络的第三、第四、第五层特征信息的方法来强化特征表达的策略。最后,提出次要感兴趣区域的概念,将交通标志的上下文信息融合到目标特征表达中。这些策略提高了目标检测的准确率和速度。Region-based fast convolutional neural networks suffer from resource wastage and an inability to effectively tackle the challenge of small object detection.To address these issues,this paper proposes convolutional neural network architecture for small object detection based on high possibility region proposal network and feature-enriched region.First,the region proposal network is employed to filter anchor regions,thereby saving processing time during the classification stage and enhancing the system’s processing speed.Then,to address the problem of ineffective detection of small objects,a strategy is proposed to strengthen feature representation by fusing the feature information from the third,fourth,and fifth layers of the VGG 16-layer network.Finally,the concept of secondary regions of interest is introduced to integrate the contextual information of traffic signs into the object feature representation.These strategies improve both the accuracy and speed of object detection.

关 键 词：目标检测 深度特征 感兴趣区域 特征融合 锚点 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]