摘要
针对红外图像特征提取困难、检测准确率低等问题。以YOLOv5为基线网络,在特征提取和特征融合阶段构建多尺度注意力模块,解决特征提取困难问题;对特征融合网络构建加权特征金字塔以保留浅层网络行人特征,解决检测精度低的问题;对红外与可见光图像特征进行融合再分割得到语义信息,利用语义损失引导高级语义信息流回图像融合模块,丰富融合图像特征。为验证所提算法的有效性,在KAIST数据集上与主流算法YOLOv5s和YOLOv7进行对比,本算法mAP分别提高了1.9%和0.8%。实验结果表明,在KAIST数据集上,YOLO-EB检测网络得到的平均精度有明显提高,夜间行人检测效果较好。
It is difficult to extract infrared image features and low detection accuracy.Taking YOLOv5 as the baseline network,the multi-scale attention module is constructed in the feature extraction stage and the feature fusion stage to solve the difficult problem of feature extraction.To solve the problem of low detection accuracy,a weighted feature pyramid is constructed for the feature fusion network to retain the pedestrian features in the shallow network.The infrared and visible image features are fused and segmented to obtain semantic information,and the semantic loss is used to guide the high-level semantic information to flow back to the image fusion module to enrich the fused image features.To verify the effectiveness of the proposed algorithm,compared with the mainstream algorithms YOLOv5s and YOLOv7 on KAIST dataset,the mAP of the proposed algorithm is improved by 1.9%and 0.8%,respectively.Experimental results show that the average accuracy of the YOLO-EB detector is significantly improved on the KAIST dataset,and the pedestrian detection effect at night is better.
作者
王浩
吕晓琪
谷宇
WANG Hao;LV Xiaoqi;GU Yu(Inner Mongolia Key Laboratory of Pattern Recognition and Intelligent Image Processing,School of Information Engineering,Inner Mongolia University of Science and Technology,Baotou InnerMongolia 014010,China;Institute of Information Engineering,Inner Mongolia University of Technology,Hohhot 010051,China)
出处
《激光杂志》
CAS
北大核心
2023年第11期48-53,共6页
Laser Journal
基金
国家自然科学基金项目(No.62001255,No.61771266,No.61841204)
内蒙古自治区科技计划项目(No.2019GG138)
中央引导地方科技发展资金项目(No.2021ZY0004)
内蒙古自治区高等学校青年科技英才支持项目(No.NJYT23057)
内蒙古自治区直属高校基本科研业务费项目优秀青年基金
内蒙古自治区自然科学基金项目(No.2019MS06003,No.2015MS0604)
内蒙古自治区高等学校科学研究项目(No.NJZY145)
教育部“春晖计划”合作科研项目(No.教外司留[2019]1383号)。
关键词
红外图像
行人检测
图像融合
加权特征金字塔
多尺度注意力机制
infrared image
pedestrian detection
image fusion
weighted feature pyramid
multiscale attention mechanism
