To balance: balanced micro-expression recognition

Micro-expressions are subtle facial movements that expose a person’s hidden emotions. Recognizing the micro-expression has importance for example in criminal investigations and psychotherapy. Compared with the shallower-architecture model, image magnification of these movements, which is also crucial for accurate recognition, has received relatively less attention in the field of micro-expression recognition. In this work, we find that there are some limitations during the training process, in particular, an imbalance in the distribution of motion amplitudes of samples, optical flow features, and semantic features. To mitigate their adverse effects, we propose adaptive balanced magnification, the balance of optical flow features and the balance of enhanced semantic features, to reduce these imbalances. Experimental results from three benchmarks (CASMEII, SAMM, and SMIC) show that our proposed method has higher accuracy and better recognition success than other micro-expression recognition methods.