【摘要】 交通指示标志的自动识别是确保行车安全最关键的技术之一。在SSD网络的基础上，论文提出一种基于分组残差卷积模块的神经网络模型（Grouped Residual Convolutional Neural Network,GR-CNN）。首先，在SSD网络上，利用ResNet50网络进行标志的特征信息提取，将原有的瓶颈残差块替换成分组可分离卷积残差块，强化特征提取能力并减小模型复杂度。然后，在每个分组残差块末端加入SA(Shuffle Attention)模块，以消除因为图像中的无效信息和池运算共同造成的特征损失。最后，引入双线性插值，聚合不同尺度间特征信息。在CCTSDB上的实验结果表明，GR-CNN网络模型的交通标志的平均识别准确率mAP达到96.15%，识别速度FPS达到27.7，模型的参数量仅为18 M。模型实现了高识别精度与低复杂度的均衡统一。更多还原

【Abstract】 Automatic recognition of traffic signs is one of the most critical technologies to ensure driving safety. On the basis of the SSD network,this paper proposes a neural network model(grouped residual convolutional neural network,GR-CNN)based on grouped residual convolution module. First,on the SSD network,the ResNet50 network is used to extract the feature information of the sign,and the original bottleneck residual block is replaced by a grouped separable convolution residual block,which strengthens the feature extraction ability and reduces the complexity of the model. Then,an SA(Shuffle Attention)module is added at the end of each grouped residual block to eliminate the feature loss caused by the invalid information in the image and the pooling operation. Finally,bilinear interpolation is introduced to aggregate feature information between different scales. The experimental results on CCTSDB show that the average recognition accuracy of traffic signs of the GR-CNN network model reaches 96.15% mAP,the recognition speed FPS reaches 27.7,and the number of parameters of the model is only 18 M. The model achieves a balance between high recognition accuracy and low complexity.更多还原