Classification for Real Traffic signs Based on Neural Network

摘要

This paper presents a study of the traffic sign classification using the neural network classifier. Firstly, 23 ideal signs are chosen as the training set and by doing so 531 real signs are roughly classified. The recognition rate is 65.9%. For the well-known reasons, the colors of the actual signs have been distorted probably. Secondly, two kinds of different real signs subset are selected for training the neural network, followed by more signs classifications for other samples. Results on two conditions show the detection rate of 99.0% and 89.3%, respectively. Thirdly, four kinds of fuzzy characteristics training set are chosen for the purpose of net training, and thus an approximated classification for 531 real signs is performed. The recognition rates on the four conditions are: 58.4%, 83.6%, 79.8 and 87.6%. By comparing the results, it may be concluded that the key factors affecting recognition rate are the color distortion and the complexity of the problem.