Emotion recognition from facial image analysis using composite similarity measure aided bidimensional empirical mode decomposition

The aim of this work is to automatically detect and analyse the emotions from the digital videos and images. Initially the images are extracted from pre-recorded videos, from which the faces are cropped automatically. The training dataset is formed with minimal number of images per subject for each emotion. Bi-dimensional Empirical Mode Decomposition (BEMD) is used to decompose the images in its Intrinsic Mode Functions (IMF). Composite Similarity Measure (CSM) based classification has been employed to detect the correct emotion from the images. "ENTERFACE'05 Audio-Visual Emotion Database", "JAFFE Database" and a database developed in laboratory called "DCAB database" are used to test the performance of the proposed method. The advantage of this method is to be able to classify or rank the emotions found in an image or a video even when the image or video is subjected to feature occlusion such as the subject putting on spectacles or sunglasses. Moreover, it is robust to illumination, different view point and background colour of the image or video. The performance is also invariant to the dress, hair style, facial hair or moustache of the subject. This method is also able to overcome the problem related to ageing to some extent.