Facial recognition has now played a pivotal role in many applications, including biomechanics, sports, image segment, animation, and robotics, etc. Although commercial facial recognition is matured, micro-expression recognition is still in its infancy and has attracted more attention from researchers in recent years. Usually, test and training samples can be recorded by different equipment throughout a variety of conditions, or by heterologous species. As a result, it is necessary to investigate whether the common micro-expression recognition algorithm is still feasible when the test samples are different from the training samples. In the present study, a series of well-developed algorithms for multi-source domain adaptation, the basic principles of multi-source domain adaptation, and the feature representation method has been discussed. A new method called the novel super-wide regression network (SWiRN) model has been introduced. Finally, some loss functions that are commonly used in neural networks for multiple source domain adaptations have been presented.
Xiaorui Zhang, Tong Xu, Wei Sun, Aiguo Song