摘要
为了解决复杂背景条件下,红外目标检测存在的准确率低、召回率低、以及网络模型在嵌入式计算平台上推理速度慢的问题,以轻量化网络YOLOv4-Tiny作为算法的基本架构,结合视觉注意力机制和空间金字塔池化思想,提出两种面向嵌入式系统的红外目标检测网络,利用迁移学习策略进行训练,在以昇腾310 AI芯片为核心的Atlas 200 DK嵌入式计算平台进行部署。实验结果表明,在该嵌入式计算平台上推理分辨率为640 pixel×512 pixel的红外图像,相较于原始网络YOLOv4-Tiny,所提网络YOLOv4-Tiny+SE+SPP的平均准确率和召回率分别提升12.36%和18.6%,推理速度达到78 fps;所提网络YOLOv4-Tiny+CBAM+SPP的平均准确率和召回率分别提升15.94%和22.89%,推理速度达到71 fps,可兼顾准确率和实时性,能够满足军事和安防领域对红外目标进行实时检测和跟踪的需要。
In order to solve the problems of low accuracy and recall rate of infrared target detection under complex background conditions,as well as slow inference speed of network model on embedded computing platform,lightweight network YOLOv4-Tiny was taken as the basic architecture of the algorithm,combined with visual attention mechanism and spatial pyramid pooling idea.Two real-time infrared target detection networks for embedded systems are proposed.Among them,there are a lot of background interference information in target detection in infrared complex scenes.Therefore,the visual attention mechanism is used to effectively learn the weight distribution of the feature map,recalibrate the feature map,strengthen the focus on the target,reduce the influence of irrelevant background information and improve the detection and recognition ability of the model.Spatial pyramid pooling can fuse multi-scale features,enrich the information of feature maps and improve the ability of infrared target recognition and location at different scales.Grad-CAM was used to visualize the feature map strengthened by the attention mechanism,showing the attention of the network model to the target region.The training is carried out on a 2080Ti GPU computer platform using the transfer learning strategy,and deployed on the Atlas 200 DK embedded computing platform with Ascend 310 AI chip as the core.The experimental results show that compared with the original network YOLOv4-Tiny,the infrared images with a resolution of640 pixels×512 pixels are detected on the computer platform.The average accuracy and recall rate of the proposed YOLOv4-Tiny+SE+SPP network were improved by 13.96%and 20.14%,respectively,and the inference speed reached 212 FPS.The average accuracy and recall rate of the proposed YOLOv4-Tiny+CBAM+SPP network were improved by 15.75%and 22.41%,respectively,and the inference speed reached202 FPS.On Atlas 200 DK embedded computing platform,infrared images with a resolution of 640 pixel×512 pixel are detected,compared with the original network YOLOv4-Tiny.The average accuracy and recall rate of the proposed YOLOv4-Tiny+SE+SPP network were improved by 12.36%and 18.6%,respectively,and the inference speed reached 78 FPS.The average accuracy and recall rate of the proposed network YOLOv4-Tiny+CBAM+SPP are improved by 15.94%and 22.89%,respectively,and the inference speed reaches71 FPS,which can meet the needs of real-time detection and tracking of infrared targets in military and security fields.
作者
张鹏辉
刘志
郑建勇
何博侠
裴雨浩
ZHANG Penghui;LIU Zhi;ZHENG Jianyong;HE Boxia;PEI Yuhao(School of Mechanical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;Nanjing Planetech Intelligent Technology Co.,LTD,Nanjing 210014,China;Institute of Artificial Intelligence,Shanghai University,Shanghai 200444,China)
出处
《光子学报》
EI
CAS
CSCD
北大核心
2022年第2期193-202,共10页
Acta Photonica Sinica
基金
国家自然科学基金(No.51575281)
中央高校基本科研业务费专项资金(No.30916011304)。
关键词
红外图像
注意力机制
迁移学习
目标检测
嵌入式平台
Infared image
Visual attention
Transfer learning
Target detection
Embedded platform
