Micro-Expression Recognition Based on Spatiotemporal Gabor Filters

For its wide range of applications in public security and psychotherapy, recognizing micro-expressions from facial image sequences has gain increasing attentions recently. Subtlety and short duration are major challenges for micro-expression recognition. In this paper, we propose a method for micro-expression recognition based on spatiotemporal Gabor filters. In preprocessing, for each video clip, the intensities of facial movements are first magnified by Eulerian video magnification (EVM), and then a sequence of frame difference is generated by subtracting a non-expression frame from all the frames in original video clip. Following that, we convolve a bank of spatiotemporal Gabor filters with the difference sequences, and the magnitudes of Gabor filter responses are used as features. The final features are fed up into a linear SVM for classification after spatiotemporal max pooling. The proposed method is evaluated on two micro-expression datasets, CASME2 and SMIC. Experimental results on CASME2 demonstrate the importance of preprocessing for micro-expression recognition. Furthermore, the proposed method achieves better recognition performance than some popular methods on both CASME2 and SMIC datasets, such as LBP-TOP and HOOF, in micro-expression recognition.