A new method for multi-wave band spectrum recognition of landmark images is suggested using improved BP network. The method follows the minimum principle of the structural design of the neural network, in which the training sample numbers match the hidden node numbers. Increasing randomly the numbers of sample of each classification plays a smooth role as a result of the increase of noise interference of sample sets,and it avoids overfitting and enhances the generalization of network, as well as speeds up the learning rates. Through the designing the two grades of BP network, this method was verified by a specific example of landmark images including 11 categories. Results show that all 11 categories of landmarks can be distinguished well using trained BP network,and distinctly the goal of recognition of various categories of landmarks can be achieved.