摘要
针对倾斜视角下的红外行人小目标难以快速准确检测的问题,提出了一种红外行人小目标轻量化实时检测网络模型DRA-YOLO。首先,使用K-means++锚框聚类自适应不同大小尺度目标,从而加快网络收敛并提高检测精度。其次,融入不同注意力机制来重新设计特征提取网络,提高特征定位与计算效率,并搭配改进特征金字塔结构提取关键特征和提升模型稳定性。最后,颈部去掉下采样重新搭配SimAM形成新的特征融合结构,并重新设计检测头来适应本文数据集。对比实验显示,相对原始YOLOv5s模型,在自制和公共数据集上表现突出。m AP50达到94.5%,检测速度提高20.8%,模型大小压缩至10.1 MB,降低了30.3%,且GFLOPs下降了29.1%。这些改进实现了对目标的准确快速检测,有效地平衡了模型大小、检测精度和推理速度。
To address the challenge of fast and accurate detection of small infrared pedestrian targets at inclined viewing angles,a lightweight real-time detection network model for small infrared pedestrian targets(DRA-YOLO)was proposed.First,K-means++anchor box clustering was utilized to adapt to targets of different size scales,thereby accelerating network convergence and improving detection accuracy.Second,different attention mechanisms were integrated into the redesigned feature extraction network to enhance feature location and computational efficiency.This was coupled with an improved feature pyramid structure to extract key features and enhance model stability.Finally,the neck was redesigned by eliminating down-sampling and reorganizing it with SimAM to form a new feature fusion structure.Moreover,the detection head was redesigned to suit the dataset used in this study.Comparative experiments showed that,relative to the original YOLOv5s model,the proposed method performed excellently on both self-made and public datasets.The mAP50 reached 94.5%,detection speed improved by 20.8%,model size was compressed to 10.1 MB(30.3%reduction),and GFLOPs decreased by 29.1%.These improvements facilitated the accurate and rapid detection of targets,effectively balancing model size,detection accuracy,and inference speed.
作者
张飞
王剑
张岳松
ZHANG Fei;WANG Jian;ZHANG Yuesong(School of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650504,China;Yunnan Provincial Key Lab.of Artificial Intelligence,Kunming University of Science and Technology,Kunming 650504,China)
出处
《红外技术》
北大核心
2025年第2期217-225,共9页
Infrared Technology
基金
国防科技特区创新项目(2016300TS00600113)。
关键词
图像处理
行人检测
红外场景
模型优化
YOLOv5算法
image processing
pedestrian detection
infrared scenery
model optimization
YOLOv5 algorithm
