融合高分辨率网络的雾天目标检测算法  

A fog target detection algorithm fusing high-resolution network

作　　者：张骞 陈紫强[1] 孙宗威 赖镜安 ZHANG Qian;CHEN Zi-qiang;SUN Zong-wei;LAI Jing-an(School of Information and Communication,Guilin University of Electronic Science and Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《计算机工程与科学》2023年第11期1970-1981,共12页Computer Engineering & Science

基　　金：国家自然科学基金(61861011);广西研究生教育创新计划(2022YCXS060)。

摘　　要：针对雾天场景中因图像模糊不清、目标难以分辨等原因导致错检、漏检的问题,提出了一种融合高分辨率网络的目标检测算法HR-Cascade RCNN。采用高分辨率网络HRNet作为Cascade RCNN的特征提取网络,通过不同分辨率的子网络并行连接,提取多尺度的特征信息,减少下采样过程中的信息损失,增强目标的语义信息表示;使用CIoU损失函数替换原有的Smooth L1损失函数,引入惩罚项度量真实框与检测框之间宽高比的相关性,优化网络的收敛效果,有助于提高检测框的定位精度;最后,采用SoftNMS改进候选框选择机制,降低车辆遮挡等情况下的漏检率,提高网络检测能力。在真实雾天数据集RTTS和合成雾天数据集Foggy Cityscapes上的实验结果表明,HR-Cascade RCNN与原Cascade RCNN相比,mAP分别提高了5.9%和3%。To address the issues of false detection and missed detection in foggy weather scenarios where images are blurred and targets are difficult to distinguish,a target detection algorithm that fuses a high-resolution network,named High Resolution Cascade RCNN(HR-Cascade RCNN),is proposed.This algorithm adopts HRNet as the feature extraction network for Cascade RCNN,connects parallel sub-networks with different resolutions to extract multi-scale feature information,thus reducing information loss during downsampling and enhancing the semantic representation of targets.Secondly,the CIoU loss function is used to replace the original Smooth L1 loss function,and a penalty term is introduced to measure the correlation between the aspect ratio of real bounding boxes and detected bounding boxes,thus optimizing the convergence performance of the network,and helping to improve the positioning accuracy of detected bounding boxes.Finally,SoftNMS is adopted to improve the candidate box selection mechanism,reducing the false negative rate in situations such as vehicle occlusion,and enhancing the detection ability of the network.Experimental results on real foggy weather datasets RTTS and synthetic foggy weather datasets Foggy Cityscapes show that compared with the original Cascade RCNN,HR-Cascade RCNN improves mAP by 5.9%and 3%respectively.

关 键 词：雾天场景 目标检测 深度学习 Cascade RCNN 高分辨率 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]