Abstract:

In traditional BP neural network weight is adjusted by steepest gradient descent method and training process could attain error-free result by following the ramp of the error function from the start,but it has some disadvantages such as slow convergence rate of training process and is prone to fall into local minimum.An improved algorithm is proposed in this paper,which makes a combination of momentum gene and adaptive learning rate,thus allowing it to keep network from falling into local minimum and accelerate training speed.Experiment shows that the improved algorithm has higher recognition accuracy rate compared with simple BP algorithm.

Key words: BP neural network;image recognition;feature extraction;improved algorithm