基于空间通道注意力机制与多尺度融合的交通标志识别研究
Research on traffic sign recognition based on spatial channel attention mechanism and multi-scale fusion
  
DOI:
中文关键词:  交通标志;轻量化网络;YOLOV3 3ctiny;多尺度融合;特征金字塔;空间通道注意力机制
英文关键词:traffic signs;lightweight network;YOLOV3-3ctiny;multi-scale fusion;feature pyramid;spatial channel attention mechanism
基金项目:国家自然科学基金(51305472)和重庆市研究生联合培养基地项目(JDLHPYJD2018003)资助项目
作者单位
黄志强 重庆交通大学 机电与车辆工程学院,重庆 400074 
李 军 重庆交通大学 机电与车辆工程学院,重庆 400074 
摘要点击次数: 3102
全文下载次数: 312
中文摘要:
      通过YOLOV3深度神经网络算法可以实现道路交通标志的自动检测与识别,由于YOLOV3运算量较大,很难在小型嵌入式平台上使用,针对这一问题,文中提出了改进型的轻量化YOLOV3 3ctiny神经网络模型。为了融合浅层特征图的空间信息与深层特征图的语义信息,将第19层卷积层通过上采样后与第7层卷积层相连接,多尺度融合后输入YOLO层形成新的特征金字塔,以此提高小目标的识别率。同时,为使网络更加关注交通标志的细节信息,在特征金字塔网络中增添能够增强前景信息降低背景信息的空间通道注意力机制。使用Kmeans聚类算法对数据集作聚类处理,获得一组先验框。在长沙理工大学交通标志数据集上进行测试,实验结果表明,改进后算法的识别率达到91.8%,与YOLOV3 tiny算法相比提高了24.9个百分点,而与YOLOV3算法相比,每张图片的检测时间降低至0133 s,降低了49.6%,该算法具有较强的实时性和准确性。
英文摘要:
      The automatic detection and recognition of road traffic signs can be realized through the YOLOV3 deep neural network algorithm. But YOLOV3 costs considerable computation resources and can hardly be used on small embedded platforms. To solve this problem, this paper proposes an improved lightweight YOLOV3 3ctiny neural network. In order to fuse the spatial information of the shallow feature map and the semantic information of the deep feature map, the 19th convolutional layer is connected to the 7th convolutional layer after upsampling, and the multi scale fusion is input to the YOLO layer to form a new feature pyramid to improve the recognition rate of small targets. Meanwhile, in order to make the network pay more attention to the details of traffic signs, a spatial channel attention mechanism that can enhance the foreground information and reduce the background information is added to the feature pyramid network. The K means clustering algorithm is used to cluster the data set to obtain a set of a priori boxes. Finally, a series of tests are conducted on the traffic sign dataset of Changsha University of Science and Technology, and the results show that the recognition rate of the improved algorithm reaches 91.8%, 249 percent higher than that of the YOLOV3 tiny algorithm; and the time is reduced to 0.133 s per image, lower by 49.6%. The algorithm has strong real time performance and accuracy.
查看全文  查看/发表评论  下载PDF阅读器

你是第3756232访问者
版权所有《南京邮电大学学报(自然科学版)》编辑部
Tel:86-25-85866913 E-mail:xb@njupt.edu.cn
技术支持:本系统由北京勤云科技发展有限公司设计

欢迎访问《南京邮电大学学报(自然科学版)》编辑部!