Abstract: Parallel flatness pattern recognition based on a binary tree hierarchical back propagation (BP) model and Legendre orthodoxy polynomial decomposition was presented aiming at the illegibility in physical meaning and poorness in robust stability of traditional flatness defect pattern recognition by the least squares method (LSM) proximity algorithm and the low accuracy of a common BP neuron network. It reduces the prediction range of each network and uses more networks for degree elevation. Experimental results show that the system performances are improved not only in robust ability but also in precision.