Multi-view Facial Expression Recognition Based on Fusing Low-level and Mid-level Features

Multi-view facial expression recognition (MFER) is one of the more active research projects in human-computer interaction. Aiming at the problem of low recognition rate of single low-level feature for multi-view facial expression recognition, a recognition method fusing low-level and mid-level features is proposed, which recognizes an expression from the coarse to the fine pattern. First of all, we extract mid-level feature based LLC (locality-constrained linear coding) in traditional SPM on facial active regions. Then we compute PHOG descriptor as low-level feature on the whole face. Next, the mid-level and low-level features are concatenated, which is simple but effective for MFER. We evaluate our approach with extensive experiments on SDUMFE and Multi-PIE datasets, which shows that our approach achieves promising results for multi-view facial expression recognition.