Wood texture detection based on optimized deep belief network model with multiple feature fusion

Abstract

Texture, as an important component of wood quality classification, is difficult to extract and distinguish due to its complex features. Based on the the traditional gray level co-occurrence matrix (GLCM), this paper introduces the local binary pattern (LBP) operator to extract the uniform rotation invariance characteristics of features for multi-feature fusion, resulting in more expressive texture feature expression. For the deep belief network (DBN) training algorithm, which may have problems such as low computational efficiency, slow convergence rate, and "dead zone", Leaky ReLU is introduced as an activation function and adaptive learning rate to optimize the DBN network model. The experimental results show that the proposed method has better recognition speed and accuracy compared to BP, ELM, SVM, CNN, etc.