Balanced DBN Algorithm with Adaptive Integration Parameter Adjustment

Abstract

ABSTRACT At present, deep belief networks (DBN) is widely used in the field of image processing. However, its random processing and Gibbs sampling make the same topology network extract the same image with different characteristics. In order to reduce this kind of difference, this paper proposes a dual DBN model (Dual-DBN), which consists of left-right symmetric sub-networks with the same dimensional feature extraction function. Each sub-network uses the matrix coefficient to adaptively adjust feature dimension of hidden layer, to obtain the optimal number of nodes. In order to optimize the network, the difference metric function is used as model loss function, and then the model parameters are fine-tuned by BPalgorithm. The experimental data is the traffic history monitoring data of six different types of roads in Jinan City, Shandong Province. Experimental tests show that the model has a higher accuracy compared with that of the traditional deep network.