A New Facial Expression Recognition System Using Decision Tree and Deep Neural Networks

A new efficient facial expression recognition system is proposed in this paper that uses a decision tree framework and deep neural networks. As an unsupervised feature extractor, the 2-D DCT is applied to difference images between neutral and expression images, and the low-frequency DCT coefficients are organized as expression features to be learnt. Geometric features are not used since their extraction is basically impractical in real-life applications. Three expressions (namely, surprise, smile, sadness) are subjects of the first two decision tree nodes that are implemented by three efficient one-hidden-layer MLPs (OHL-MLP) that are trained by utilizing the well-known back-propagation (BP) algorithm. Three other expressions (namely, disgust, anger, fear) are processed and analyzed by the third node in the decision tree which is realized by a deep neural network, where an autoencoder is first applied to the directly concatenated DCT coefficients of different facial parts such as eyes, mouth and nose to fuse and smooth the features and an OHL-MLP is then trained to classify the target expressions. The proposed recognition system is applied to the extended Cohn-Kanade (CK+) database, with 123 subjects and 70 images for each expression being collected and used (50 for the training phase and 20 for the testing phase). The recognition rate of the first node is 99.17% for both the surprise and smile facial expressions. Sadness is perfectly (100%) recognized by the second node. The last node presents 100.00%, 70.00%, and 55.00% accuracy corresponding to the fear, anger and disgust facial expressions, respectively.