The majority of studies on texture analysis focus on classification and generation, and few works concern perceptual similarity between textures, which is one of the fundamental problems in the field of texture analysis. Previous methods for perceptual similarity learning were mainly assisted by psychophysical experiments and computational feature extraction. However, the calculated similarity matrix is always seriously biased from human observation. In this paper, we propose a novel method for similarity prediction, which is based on convolutional neural networks (CNNs) and stacked sparse auto-encoder (SSAE). The experimental results show that the predicted similarity matrixes are more perceptually consistent with psychophysical experiments compared to other predicting methods.