Özet:

Abstract Facial Expressions are quite personalized and may look different for different individuals. Whereas there are certain facial muscles which shows some common features for certain human expressions across the cultures and facial shapes. Convolution Neural Network have shown tremendous success in Facial Expression Recognition task. In recent past many researchers have proposed multiple models with manageable size solution to Facial Expression Recognition task. In the current work, we have considered shape, complexion and other identity related information separate from certain specified muscle movements which are specific for emotion recognition. This is done by a novel Emotion-Generative Adversarial Network. This saves a lot of effort and simplifies the Facial Expression Recognition process. We then apply Scale Invariant Feature Transformation and vola john’s face extraction method for pre-processing and face image extraction from background.  This enables us to train our model accurately irrespective of scale, orientation, illumination etc and with very less training samples accurately. We feed the feature extracted facial image to an attention-based Convolutional Neural Network. This will ensure more emphasis on critical areas for expression recognition of facial image. Finally, we have used Local Binary Pattern for classification of the input image to a particular emotion class. We have tested our model on CK+, OULU- Casia and FER-2013 datasets and it is at par with performance of all major state-of-art models. Proposed model may be utilized by various automated interactive systems, such as robot to human communication, automated customer care systems etc. The proposed work may also be quite useful for observing reaction of viewers to a particular advertisement or article automatically and use this information for various purposes like user’s interest, product feedback etc.

Anahtar Kelimeler: