摘要
由于存储资源和功耗等限制,在嵌入式设备上部署基于深度学习的目标检测算法存在困难,且效果不佳.基于YOLOv4算法,提出了一种改进的YOLOv4-Light轻量化网络模型,采用GhostNet网络结构替换CSPDarknet53作为主干提取网络,引入CBAM注意力机制关注通道和空间两个维度的特征信息,并利用感知量化方法对权重进行INT8量化处理,在保证精度的情况下降低网络模型规模和参数量.在PC端和NVIDIA Jetson Xavier NX上选用VisDrone无人机数据集分别对网络模型进行测试,结果表明YOLOv4-GhostNet-CBAM模型的尺寸是160M,比YOLOv4降低了34.43%;检测速率最高可达到34.6FPS,比YOLOv4提高了56.6%.YOLO-Light模型的尺寸是40.2M,比YOLOv4降低了83.5%;检测速率最高可达到78.6FPS,为YOLOv4的3.6倍,且交并比为0.5时的平均精度均值(mAP50)与YOLOv4相比仅下降了3%.YOLO-Light模型相较于原模型优势明显,能够在低功耗的嵌入式设备上完成实时目标检测.
It is difficult and ineffective to deploy deep learning-based target detection algorithms on embedded devices,due to storage resources and power consumption limitations.This paper proposes an improved YOLOv4-Light lightweight network model based on YOLOv4 algorithm.In order to reduce the model scale and number of parameters with guaranteed accuracy,GhostNet network is adopted to replace CSPDarknet53 as the backbone extraction network,CBAM attention mechanism is introduced to focus on the feature information of the channel and space dimensions,and the weight of the model is quantified into INT8 type with the perceptual quantization method.The network models are tested on PC and NVIDIA Jetson Xavier NX respectively with VisDrone dataset.The results show that the model scale of YOLOv4-GhostNet-CBAM is 160M,which is 34.43%lower than YOLOv4;the detection rate can reach up to 34.6FPS,which is 56.6%higher than YOLOv4.The model scale of YOLO-Light is 40.2M,which decreases by 83.5%compared with YOLOv4;the detection rate can reach up to 78.6FPS,which is 3.6 times of YOLOv4,and the average mean accuracy at a cross-merge ratio of 0.5(mAP50)is only 3%lower compared to YOLOv4.This YOLO-Light model has obvious advantages over the original model and can accomplish real-time target detection on low-power embedded devices.
作者
石博雅
董学峰
SHI Boya;DONG Xuefeng(School of Electronic and Information Engineering,Tiangong University,Tianjin 300387,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2024年第3期651-656,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61903273)资助
天津市自然科学基金项目(19JCYBJC16200,19JCQNJC03300)资助
天津市科技计划项目(20YDTPJC01530)资助.
