论文部分内容阅读
为了提高交通标志识别的鲁棒性,提出了一种基于图模型与卷积神经网络(CNN)的交通标志识别方法,建立了一个面向应用的基于区域的卷积神经网络(R-CNN)交通标志识别系统。构造了基于超轮廓图(UCM)超像素区域的图模型,有效利用自底向上的多级信息,提出了一种基于图模型的层次显著性检测方法,以提取交通标志感兴趣区域,并利用卷积神经网络对感兴趣候选区进行特征提取与分类。检测结果表明:针对限速标志,基于UCM超像素区域的图模型比基于简单线性迭代聚类(SLIC)超像素的图模型更有利于获取上层显著度图的大尺度结构信息;基于先验位置约束与局部特征(颜色与边界)的层次显著性模型有效地融合了局部区域的细节信息与结构信息,检测结果更加准确,检测目标更加完整、均匀,查准率为0.65,查全率为0.8,F指数为0.73,均高于其他同类基于超像素的显著性检测算法;基于具体检测任务的CNN预训练策略扩展了德国交通标志识别库(GTSRB)的样本集,充分利用了CNN的海量学习能力,更好地学习目标内部的局部精细特征,提高了学习与识别能力,总识别率为98.85%,高于SVM分类器的95.73%。
In order to improve the robustness of traffic sign recognition, a traffic sign recognition method based on graph model and convolutional neural network (CNN) is proposed. An application-oriented region-based convolutional neural network (R-CNN) Logo recognition system. A graph model based on super-pixel map of super-contour map (UCM) is constructed and a multi-level bottom-up information is effectively used. A hierarchical saliency detection method based on graph model is proposed to extract the interest areas of traffic signs. Convolutional neural networks feature extraction and classification of regions of interest. The test results show that the graph model based on the UCM superpixels is more conducive to obtaining the large-scale structure information of the upper saliency map than the graph model based on the simple linear iterative clustering (SLIC) The hierarchical saliency model of constraint and local features (color and boundary) effectively fused the detail information and structure information of local area, the detection results are more accurate, the detection target is more complete and uniform, the precision is 0.65 and the recall rate is 0.8 , And the F-index was 0.73, all of which were higher than other similar superpixel-based saliency detection algorithms. The CNN pre-training strategy based on specific testing tasks extended the GTSRB sample set and fully utilized CNN’s massive learning Ability to better learn the local fine features inside the target and improve the learning and recognition ability, the total recognition rate is 98.85%, which is 95.73% higher than the SVM classifier.