Publication

Abstract : Emotion Recognition through facial expressions has become an active area of research. Challenges like pose and illumination variations in images increases the scope of research. Several machine learning and deep learning techniques have been used to perform emotion recognition. However, these models need huge volume of data for training and are very time-intensive. Few aforesaid challenges are addressed in this work by using meta-learning techniques which require very few training samples for training and helps the model to generalize to new tasks. In this work, the concept of meta-learning with Siamese Networks to recognize emotions from facial expressions for arbitrary poses is used and the model is coined as MREAP (Meta learning approach to Recognize Emotions with Arbitrary Poses). An in-house dataset AED-1 (Amrita Emotions Dataset-1) has been used which consists of images expressing five basic emotions: anger, disgust, happiness, neutral and surprise in seven arbitrary poses and very less training samples are used for training. MREAP predict emotions with poses that are never seen in training which is a novelty in FER tasks. In this work, the model learns and builds feature vectors for making predictions using Euclidean distance as the metric and it achieved an accuracy of 80%.