Based on the geometrical invariability of Hu invariant moment, and taking the characteristics of image invariant moment as an input, an improved error back propagation (BP) neural network for batch training is established. According to Bayesian normalization Levenberg-Marquardt algorithm, the calculation accuracy of error functions is optimized and the network is improved so as to realize the optimized parameter combination. An improved BP neural network object recognition model based on invariant moment is built in MATLAB environment. Experimental results show that accurate target recognition and correct interferential image estimation are implemented with the method.