Design Of Traffic Sign Recognition Algorithm Based On Depth Learning

At present,China's social economy is continuously developing,and the number of cars is also steadily increasing.Although it has made it more convenient for citizens to travel,the increase in vehicles has also caused traffic jams and frequent traffic accidents.Among them,the driver misjudged the traffic indication information,resulting in misoperation during driving,which is the main cause of traffic problems.However,the driver's missing judgment on traffic information is caused by the failure to accurately detect the visual image information.Therefore,how to use computer visual processing technology to efficiently identify traffic sign information in road environment and reduce the occurrence of such traffic accidents is a research topic with positive significance.This paper mainly analyzes the identification technology of traffic signs in road scenes,and uses the identification algorithm model to judge the category information of designated signs to reduce the occurrence of accidents in daily traffic.The main research contents of this paper include the following aspects.Firstly,the theory based on convolution neural network structure is analyzed,and the training process is expanded from the network structure composed of convolution neural network.The paper also studies the existing image recognition technology in transportation field,compares different design schemes,and proposes that this paper adopts the end-to-end network design based on removing candidate regions to extract features.Then,the traffic sign target detection based on the end-to-end YOLO v3 algorithm is designed,and the basic network structure darknet-53,multi-scale region identification characteristics and feature classification are analyzed.The strategy of optimizing the loss function is adopted to solve the problem of category imbalance caused by the large number of prediction frames and the instability of positive and negative sample data.In the original algorithm test,when too many prediction frames lead to multiple repeated frames in the final recognition result,the non-maximum suppression module NMS is improved,and a reduced score target is adopted to replace the set threshold score,and finally the detection frame with the highest score of the target category is retained to ensure the accuracy of detection.Subsequently,the algorithm model parameters are set,including input size 416*416,learningrate0.001,iteration times 500,000,etc.,and the parameters are adjusted.The cross-union ratio threshold of the model is adjusted and analyzed,and the iou threshold of 0.3 is selected to complete the model training.Finally,through experimental comparison,compared with the original YOLO v3' s 88.6% accuracy rate and 88.9% recall rate,the optimized algorithm achieves 89.1% accuracy rate and89.8% recall rate.Then,the optimized YOLO network model is compressed and improved.The convolution in the basic network darknet-53 is replaced by the depth separable convolution of Depthwise+Pointwise.The spatial information and channel information of the extracted feature map are processed separately to reduce the number of parameters,thus improving the training rate of the model.Using the pyramid feature fusion method of FPN,the output result of the original YOLO network after 16 times of downsampling fusion is upsampled,fused with the 8 times of downsampling result output by the network,and a fused feature detection layer of 8times of downsampling is established to realize accurate identification of small area targets in the image detection process.Finally,the functional test of the system is implemented on pycharm compiler using python language and tensorflow's deep learning framework.The improved network model is analyzed and compared with the original algorithm model under the indexes of accuracy rate,false detection,missing detection,etc.Finally,after completing the training,the experiment shows that the average detection accuracy m AP of the compressed improved model is 92.4%,which is nearly 3% higher than the average detection accuracy of 89.5% of the optimized network,and the loss value of the model is 0.2010,which is 0.1 lower than the loss value of 0.3011 of the original YOLO v3 algorithm,ensuring the stable convergence of the training of the model.