Research On Traffic Sign Image Recognition Based On Convolutional Neural Network

As an important part of intelligent transportation system,traffic sign image recognition has an indispensable guiding significance in driver assistance system and a pivotal application prospect in the field of driverless cars.In this thesis,German traffic sign image database(GTSRB)is taken as the research object.The method based on deep convolutional neural network improves and optimizes the AlexNet network model to improve the real-time performance of traffic sign recognition while ensuring the accuracy of image recognition.The main research contents and results of this thesis are as follows:Firstly,according to the characteristics of traffic sign images in natural scenes,grayscale enhancement and size normalization are adopted to improve the quality of training images and reduce the influence of light and background interference on training images.Secondly,the AlexNet network model based on deep convolutional neural network is improved.On the premise of improving the recognition accuracy,the total number of parameters is reduced by compression and memory consumption is greatly reduced.Thirdly,based on the training depth convolutional neural network model of Ubuntu system platform and caffe deep learning framework,the curve of training accuracy and loss function changing with the training times of iteration can be obtained through visualization operation,demonstrating the convergence of the model and the advanced algorithm.Finally,through the test and verification of the traffic sign image test set,the performance of the improved and optimized Alexnet network model is obtained,and the recognition effect is compared and analyzed.Eventually,the traffic sign image recognition method based on the improved deep convolutional neural network achieved good results,with the recognition accuracy of 96.875% and the recognition speed of 40 milliseconds per image.