Robust emotion recognition techniques from facial expressions using images and videos

Abstract

Emotion recognition from facial expressions is the process of recognizing human facial expressions into six basic emotions namely anger, disgust, fear, happy, sad and surprise. Facial expression recognition has attracted significant attention because of its various applications in robotics, psychology, medicine, security, and computing (human-computer interaction, interactive games, computer-based learning, entertainment, etc). Recognizing emotions from facial expressions using images that are varying in pose, illumination and age at real time are challenging tasks. The most expressive features on the face are eye, eyebrow, nose, chin and mouth regions. In a video, the frame in which peak of an emotion is expressed is called as an apex frame. The apex frame and a suitable classifier are the key elements for emotion recognition. Identifying and extracting the most expressive features on the face from the apex frame that could recognize emotion at high accuracy is a very important problem to be addressed. A suitable quotfeature-classifierquot combination improves the accuracy of emotion recognition. The aforesaid challenges have motivated to work on emotion recognition from facial expressions using images and videos. This thesis addresses appearance and geometric feature based approaches for feature extraction for emotion recognition from images. Using appearance based approach for feature extraction, analysis of spatial and transform domain methods is performed on frontal face and it is observed that transform domain methods outperform spatial domain methods. Using transform domain methods for feature extraction, emotion recognition is performed for images with pose and illumination variations separately. The accuracy gradually reduces when pose changes. Illumination affects emotion recognition and suitable pre-processing of images is required prior to feature extraction. To make emotion recognition invariant to pose and illumination, methods using geometric feature based approach for feature extraction are proposed..