Abstract: Hand-designed features (such as Gabor, LBP) has been widely employed in facial expression recognition. In the real-world applications of facial expression recognition, it is very difficult to achieve perfect face alignment because of the impact of complex background and the limitations of face alignment approaches. Independent Subspace Analysis (ISA) is an unsupervised feature learning method, which can be used to learn phase-invariant visual features from images. The problem of facial expression recognition based on ISA in the situation of not precise face alignment was investigated. Through analyzing the facial expression recognition performances with different subspace size, it was turned out that choosing an appropriate subspace size is important to improve the robustness of learned features for facial expression recognition in the situation of not precise alignment.

Key words: facial expression recognition, independent subspace analysis, unsupervised feature learning, spatial-temporal feature learning