Research On Recognition Of Traffic Sign Based On Deep Learning

Traffic signs recognition technology is an important part of intelligent transportation and pilotless system,it is also one of the hot topics in the field of machine vision.The German Traffic Sign Recognition Benchmark(GTSRB)is employed in this thesis to study the characteristics of traffic sign s by a new proposed algorithm to enhance the detailed information about image and one modified Le Net-5 model to improve the extracting performance and the recognition accuracy in image processing.The main contents and achievements of this thesis as follows:(1)Considering of the c haracteristics of the traffic sign s in the GTSRB,a self-adaptive contrast enhancement method based on gradient and intensity histogram is proposed to overcome the over-enhancement shortage of traditional algorithm that caused by the high peak value of his togram.(2)A series of digital image processing methods are adopted to improve the quality of images,including regional cutting,gray,dimension normalization,and contrast enhancement,to reduce the effects of background information and noises.(3)The parallel convolutional layer,preventing over fitting strategy,and batch gradient algorithm are introduced to modify the Le Net-5 model,therefore,the performance of extracting image feature information and the accuracy of recognition are improved.(4)To improve the efficiency of model training,the modified model is trained based on the Ali cloud machine learning platform,the convergence and the superiority of algorithm are verified by studying the variation of training precision and cross entropy loss function with the iterative times.(5)The performance of modified model is tested by investigation of the effects of recognition with several typical traffic sign images.The results suggests that the proposed traffic signs classification algorithm based on deep learning can efficiently classify the target images,the average accuracy of which is up to 97.16%.