Interpretable Tongue Constitution Recognition via Reshaped Wavelet Attention

Abstract

Currently deep neural networks have been used to perform the tongue constitution recognition, but they are still challenged, failing to extract nice multi-scale and multi-level features. This paper proposes a novel interpretable tongue constitution recognition method based on the reshaped wavelet attention. It separates multi-scale features through discrete wavelet transform and then uses the attention mechanism to weight them. Subsequently, these features are reshaped to the high-dimensional space where the association knowledge of multi-level features are mined and hierarchized so as to fuse them efficiently. Finally, both are integrated into the framework of convolution neural network to generate the more accurate tongue image attributes, by which the tongue constitution recognition is performed. The proposed method not only obtains the higher performance with small cost, but also nicely interprets them. Experimental results show that the proposed method is effective, efficient, and interpretable.