摘要
针对现存交通标志识别模型参数量过大、检测速度慢和检测精度较低的问题,本文提出一种改进YOLOv4-tiny的交通标志识别算法.该算法将深度可分离卷积应用到YOLOv4-tiny的特征提取网络中,显著降低了主干网络的参数量和计算量.在特征融合阶段,将特征提取网络得到的不同层次特征图输入双向特征金字塔网络结构(BiFPN)中进行多尺度特征融合.最后,在损失函数设计过程中,使用Focal损失函数代替二分交叉熵损失函数,使检测过程中的正负样本数量不均衡问题得以解决.在TT100K数据集上的测试结果表明,该算法的平均精度均值达到87.5%,相比于YOLOv4-tiny提升了3.9%,模型大小为14MB,仅为YOLOv4-tiny的58%.该算法一定程度上减少了计算量和模型大小,并带来了检测速度和精度的提升.
An improved Yolov4-Tiny traffic sign recognition algorithm is proposed to solve the problems of excessive number of parameters,slow detection speed and low detection accuracy in the existing traffic sign recognition models.This algorithm applies depthwise separable convolution to Yolov4-Tiny feature extraction network,which significantly reduces the number of parameters and computation of the backbone network.In the stage of feature fusion,feature maps of different levels obtained from feature extraction network are input into bidirectional feature pyramid network structure(BiFPN)for multi-scale feature fusion.Finally,in the design process of loss function,Focal loss function was used to replace the dichotomous cross entropy loss function,so as to solve the problem of unbalanced number of positive and negative samples in the detection process.The test results on TT100K data set show that the algorithm has an average mean accuracy(mAP)of 87.5%,which is 3.9%higher than Yolov4-Tiny,and the model size is 14MB,which is only 58%of Yolov4-Tiny.The algorithm reduces the computation and model size to some extent,and brings about the improvement of detection speed and accuracy.
作者
郭继峰
孙文博
庞志奇
费禹潇
白淼源
GUO Ji-feng;SUN Wen-bo;PANG Zhi-qi;FEI Yu-xiao;BAI Miao-yuan(School of Information and Computer Engineering,Northeast Forestry University,Harbin 150000,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2022年第7期1471-1476,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61300098)资助
黑龙江省自然科学基金项目(LH2019C003)资助.
