Histogram distance metric learning for facial expression recognition.

• A new metric learning method is presented for histogram data classification. • A convex metric learning cost function is proposed based on modified chi distance. • Local Metric Learning is proposed for facial expression recognition in the wild. • Dropout based regularizer is employed to avoid over-fitting on training data. Facial expression recognition is an interesting and challenging problem in computer vision. So far, much research has been performed in this area; however, facial expression recognition in uncontrolled conditions has remained an unresolved problem. The widely-used feature descriptors in computer vision are often histogram data. In this paper, a new metric learning method is presented for histogram data classification. In this method, chi-squared distance is appropriately modified for metric learning. Then, a convex cost function is proposed to use in metric learning optimization. Moreover, the proposed algorithm is redefined as Local Metric Learning for facial expression recognition problem. In this definition, the proposed metric learning method is applied locally on facial sub-regions. Experimental results on four histogram datasets (dslr, webcam, amazon, and caltech) as well as controlled and uncontrolled facial expression recognition datasets (CK+, SFEW, and RAF-DB) show that the proposed method has superior performance compared to the state-of-art methods. [ABSTRACT FROM AUTHOR]