The classification of traffic sign base on fuzzy characteristics training set

The BP networks have the ability of nonlinear mapping to realize the function of N categoriser. However, BP networks need to be trained again when the training set is changed. Meanwhile, the larger the network is, the slower convergence rate is, and the poorer result of classification and recognition is. So, two-hierarchy neural network classifier for recognition of traffic signs is presented, the first hierarchy classification which consists of a single BP network is used to coarsely classify indicative signs, warning signs and prohibitive signs. Classify the sample of similar characteristic and superpose the image pixel of each type, then a new sample called fuzzy characteristic training set is obtained. The correctness of classification is up to 100% for testing set with white Gaussian noise, while the first hierarchy neural network is trained. Taking fuzzy characteristic training as the first classification in the two-hierarchy neural network not only reduces the scale of training set but also makes the neural network fault tolerant. The convergence rate and classification also improve. Moreover it indicates that this kind of training method is similar to thinking models of biological intelligence.