In recent years, the importance of catching the human's emotions grows larger as the Artificial Intelligence (AI) fields developed. The Facial Expression Recognition (FER) is part of understanding the emotion of the human through facial expression. Th...
In recent years, the importance of catching the human's emotions grows larger as the Artificial Intelligence (AI) fields developed. The Facial Expression Recognition (FER) is part of understanding the emotion of the human through facial expression. This technique is highly utilized in the various fields because it can identify person’s expression from the face and provide the customized services.
With growth of deep learning, several approaches have been proposed to recognize the human's expression. In early stage, most of the studies use classical features such as holistic features or the local features of the face. As time goes on, the various neural networks appears such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM). The classical features become information which are fed to the networks to automatically discover the representations needed for object detection or classification.
In this thesis, we propose an algorithm which can efficiently classify the expression through feeding various and reinforced features to the joint fusion classifier. We design the inputs which are goes into the multi-depth networks be minimum overlapped so these can give a various information to the networks. To utilize 3D CNN, we propose a multi-rate-based 3D CNN based on multi-rate signal processing scheme. Also, we make the input images to be normalized based on the intensity of the given image and reinforce the output features by the self-attention. Then we concatenate the reinforced features and classified the expression by joint fusion classifier.
Through the proposed algorithm, for the CK+ database, the result of the proposed joint fusion classifier shows comparable accuracy of 96.23%. For the MMI and the GEMEP-FERA database, it outperforms other state-of-the-art models with accuracy of 96.69% and 99.79%. For the AFEW database, the proposed algorithm shows the accuracy of 31.02%.