Abstract: Aiming at the characters of pattern recognition of fracture surface images, a character pick-up method by wavelet transform is proposed. Then, based on BP neural network theory, a nonlinear classifier is designed specially for such a pattern recognition, and its network structure was determined by test to give the way to choose parameters involved. Experimental investigations were carried out on computer aiming at several typical metal fracture surface images. The results showed that the average rate of correct recognition achieves 93.75%, even 95% if using the energy as character parameters alone. It means that a nonlinear classifier has a higher and more reliable rate of correct recognition than the linear classifier. The results showed that the nonlinear pattern recognition and classification method based on wavelet transform and BP neural network theory can recognize correctly the images in which the texture is complex, with a better adjustability provided.