BP algorithm has been widely applied to vehicle license plate recognition. However, this approach is challenged by the basic problem in the recognition system itself, that is, the accuracy of recognition. Aiming at the limitation of standard BP algorithm, this papers puts forward a method to improve the BP algorithm by modifying the action function and regulating the learning rate. A tested comparison shows that the improved algorithm is much more efficient in license plate recognition thanks to its fast convergence speed, rapid recognition ability and high recognition rate.