Research On Traffic Sign Image Recognition Based On Deep Learning

The recognition of traffic road signs is not only a key technology in autonomous vehicle,but also an auxiliary tool for advanced vehicle-mounted systems.Although the traditional traffic sign recognition method can be used to detecte,its accuracy rate is difficult to be further enhanced.Due to the deep learning method has a strong ability of feature representation,which applied to the detection and recognition of traffic sign can effectively help the driver(or unmanned vehicle)to complete the correct operation,so as to reduce or avoid the occurrence of traffic accidents.In view of the difficult problems in the detection of traffic road signs,this paper improves the SSD model to achieve the detection and recognition and accurate positioning of small traffic road signs.Related work and innovations are as follows:(1)Aiming at the problems of missing detection and low detection accuracy of the SSD model in the detection of small targets,a Dalation_DenseNet_SSD model was proposed.The feature extraction capability of the network was improved,and the loss of information caused by pooling operations can be reduced and also the detection performance for small targets was improved effectively by the application of Dalation_DenseNet_SSD model.Experimental verification was carried out on the CTSD dataset.The experimental results show that the mAP of the Dalation_DenseNet_SSD model has increased by 3.9% and 6.4% respectively compared to the Dalation_VGG_SSD model and the SSD model,the losses have decreased by 0.5 and1.8,and the recall rate has reached 92.6%,and the accuracy rate has reached 94.4%.(2)Due to the weak correlation between the prediction frame classification confidence and position confidence,which results in the inaccurate positioning of the prediction frame,a kind of model as KL_Softer-NMS_DD_SSD is proposed.The classification confidence and the correlation between the prediction frame and the real frame IoU was enhanced.And also weighted the obtained multiple detection frames to update the final detection frame,effectively improving the precise positioning and classification confidence of the prediction frame.Experiments show that the mAP of the KL_Softer-NMS_DD_SSD model further raises1.2% compared to Dalation_DenseNet_SSD,and the loss decreases by about 0.4,and the recall rate and accuracy reaches 95.3% and 96.6% respectively.(3)In order to improve the recognition effect by increasing the complexity of the network model in the recognition of traffic road signs,it has caused an increase in the number of parameters.An improved SqueezeNet-IR-GRU model is proposed.The lightweight SqueezeNet was used to reduce the number of covariates;the ELU activation function was used to solve the gradient descent problem;the number of covariates was further reduced by improving the depth residual network and GRU neural network,which reduced the training time of the network and ensured the stability of the model.The experimental verification was carried out on the GTSRB dataset.The experimental results show that the parameter amount and comparison model of the proposed SqueezeNet-IR-GRU model are reduced by a maximum of 1/5 times,and at the same time,the stability and convergence of the model are improved,and the accuracy rate It also reached 99.13%.In addition,the improved model was verified on the CAFIR10 data set,and an accuracy of 88.25% was also achieved.