Emotion Recognition for Affective Computing: Computer Vision and Machine Learning Approach

The purpose of affective computing is to develop reliable and intelligent models that computers can use to interact more naturally with humans. The critical requirements for such models are that they enable computers to recognise, understand and interpret the emotional states expressed by humans. The emotion recognition has been a research topic of interest for decades, not only in relation to developments in the affective computing field but also due to its other potential applications. A particularly challenging problem that has emerged from this body of work, however, is the task of recognising facial expressions and emotions from still images or videos in real-time. This thesis aimed to solve this challenging problem by developing new techniques involving computer vision, machine learning and different levels of information fusion. Firstly, an efficient and effective algorithm was developed to improve the performance of the Viola-Jones algorithm. The proposed method achieved significantly higher detection accuracy (95%) than the standard Viola-Jones method (90%) in face detection from thermal images, while also doubling the detection speed. Secondly, an automatic subsystem for detecting eyeglasses, Shallow-GlassNet, was proposed to address the facial occlusion problem by designing a shallow convolutional neural network capable of detecting eyeglasses rapidly and accurately. Thirdly, a novel neural network model for decision fusion was proposed in order to make use of multiple classifier systems, which can increase the classification accuracy by up to 10%. Finally, a high-speed approach to emotion recognition from videos, called One-Shot Only (OSO), was developed based on a novel spatio-temporal data fusion method for representing video frames. The OSO method tackled video classification as a single image classification problem, which not only made it extremely fast but also reduced the overfitting problem.