摘要
针对交通标志图像中目标物较小,SSD(单次多框检测)模型对其检测精度不佳的问题,提出一种基于SSD模型改进的卷积网络算法。在原SSD特征层基础上加入低层特征图,并将低层邻近特征图进行融合,实现不同特征层的多元信息分类预测与位置回归。对SSD默认框的大小选取进行k-means聚类分析,调整原有默认框比例,加快模型收敛。通过不同数据集进行验证,实验结果表明,该算法表现出较好的检测效果,同时满足实时性的要求。
Aiming at the problem that the traffic sign objects is small,meanwhile the SSD(single multi-box detection)model has poor detection accuracy,an improved SSD model is proposed.On the basis of the original SSD feature layer,the feature map of the lower layer was added,and the adjacent feature map of the lower layer was fused to realize the classification and prediction of multiple information and location regression of different feature layers.The size default box of SSD was carried out by k-means clustering analysis,and it adjusted the proportion of the default box size and accelerated the convergence of the model.Finally,the experiments were carried out through different data sets and the results show that the proposed model improves the detection effect and meets the requirement of real-time performance.
作者
田智慧
孙盐盐
魏海涛
Tian Zhihui;Sun Yanyan;Wei Haitao(College of Earth Science and Technology,Zhengzhou University,Zhengzhou 450052,Henan,China;College of Information Engineering,Zhengzhou University,Zhengzhou 450001,Henan,China)
出处
《计算机应用与软件》
北大核心
2021年第12期201-206,共6页
Computer Applications and Software
基金
河南省重点研发与推广专项(科技攻关项目)(192102210124)。
关键词
神经网络
目标检测
特征融合
交通标志
Neural network
Target detection
Features fusion
Traffic signs
