Abstract:

Aiming at the gradient vanishing problem caused by the increasing number of network layers, an image recognition method based on the Coupled-Deep Belief Network (C-DBN) is proposed, which introduces the cross-layer linkage to the Deep Belief Network (DBN). The structure and the parameter updating method for the C-DBN are given in detail, while the performance of the DBN and that of the C-DBN are compared with respect to their respective best parameters and the same net-depth on two image datasets. Moreover, the impact of the weights used in the coupling between the primary line and the secondary line is analyzed at the cross-layer linkage. Besides, several classic deep learning based methods are compared with the proposed C-DBN. Experimental results show that the C-DBN is superior to the DBN in terms of convergence speed and accuracy. And, a good performance is achieved by the proposed method in comparison with some classical deep learning methods. This means that the usage of cross-layer linkage can alleviate the gradient vanishing problems effectively in the DBN training, which helps to improve the following recognition performance.

Key words: cross-layer linkage, deep belief network, deep learning, image recognition