Robust Facial Expression Recognition Based on Local Tri-directional Coding Pattern

Automatic facial expression recognition system has broad applications prospect in the field of computer vision, and of which the precision of facial expression feature extraction is the crucial factor for improving recognition accuracy. Whereas the traditional Local Binary Pattern has the shortcomings of inaccurate feature description and large feature size, in this paper we propose a novel feature descriptor, namely local tri-directional coding pattern (LtriDCP), to overcome the above two flaws. First, we adopt Kirsch masks to compute convolution values of each pixel in 8 directions, which can accurately represent the texture information of facial expression compared with the coding scheme of LBP. Second, considering that the difference of facial expression is mainly characterized in the texture change of horizontal, vertical and diagonal directions in eye, mouth, forehead and other facial areas, we only encode three convolution values in the corresponding directions to reduce feature size but still retain superior performance. Experimental results on JAFFE database show that LtriDCP outperforms several state-of-the-art feature descriptors and demonstrate its effectiveness.