摘要
针对交通标志独特的形状特点,提出一种改进的anchor-free卷积神经网络检测算法FCOS(fully convolutional one-stage object detection)。由于原算法在训练过程中直接将真实框内的位置点标记为正,会产生大量标签噪声,在FCOS网络结构的基础上融合交通标志的形状特征,减少没有辩证能力的噪声标签,设计新的正样本选择策略。实验结果表明,改进后FCOS算法在处理后的TT100K数据集上的检测mAP(mean average precision)在不增加计算量的情况下提升到83.2%,检测性能高于FCOS。
According to the unique shape characteristics of traffic signs,an improved anchor-free convolutional neural network detection algorithm FCOS(fully convolutional one-stage object detection)was proposed.It was analyzed that the original algorithm directly marked the position points in the real frame as positive in the training process,which would produce a lot of label noise.Therefore,based on the FCOS network structure,the shape features of traffic signs were integrated to reduce the noise labels without dialectical ability,and a new positive sample selection strategy was designed.Experimental results show that the detection mAP(mean average precision)of the improved FCOS algorithm on the TT100K data set after processing is improved to 83.2%without increasing the amount of computation,and the detection performance is higher than that of FCOS.
作者
崔港涛
马社祥
CUI Gang-tao;MA She-xiang(School of Electrical and Electronic Engineering,Tianjin University of Technology,Tianjin 300382,China)
出处
《计算机工程与设计》
北大核心
2023年第10期3153-3159,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(61371108)。
关键词
交通标志检测
FCOS
深度学习
正标签
回归位置
卷积神经网络
噪声
traffic sign detection
FCOS
deep learning
positive label
regression position
convolutional neural network
noise