Research On Small Target Traffic Sign Recognition Based On Multi-scale Convolutional Neural Network

Traffic sign recognition has a wide application prospect in the field of automatic driving.In practical application scenarios,occlusion,illumination,detection methods and other factors will lead to the omission and misdetection of small traffic signs.This paper proposes a new multi-scale fusion convolutional neural network model(SC-RCNN).To begin with,the Multi-scale Atrous Spatial Pyramid Pooling(MASPP)module is added to the basic feature extraction network.After multi-scale dilated convolution sampling,the amount of information under each feature is not changed.In this way,the loss of resolution can be reduced and the context information of the same image can be captured.Then,two Fast Concat modules(F-Concat)are added in the network to fuse the information of middle and high level and low level of the model.The fusion of high and low level information can not only enrich semantic information,but also realize the reuse of information at different scales.Furthermore,a Batch Normalization layer(BN)is added before each maximum pooling layer,transforming the data for each layer.Although the addition of modules deepens the network depth.BN layer can accelerate the model convergence speed,so that the whole training time does not change greatly.Finally,the clustering method is used to adjust the size and size of the anchor frame to adapt to the size of the target in the dataset.The experimental results shows that the traffic sign recognition accuracy is significantly improved by using the network model in this paper.On the CCTSDB dataset,the average accuracy of traffic sign recognition reaches 87.48%.The recognition accuracy of warning category reaches 89.87%,prohibition category reaches 90.12%,direction category reaches 82.11%,and indication category reaches 89.20%.On the GTSDB dataset,the average identification accuracy of traffic signs reaches 59.15%,the identification accuracy of warning category reaches 79.48%,the identification accuracy of prohibition category reaches 48.08%,the identification accuracy of direction category reaches 39.24% and the identification accuracy of indication category reaches 69.81%.The model can effectively prevent the omission and misdetection of small target signs,improve the identification accuracy of traffic signs,and has certain application value.There are 35 figures,7 tables and 50 references in this paper.