Facial expression recognition using 2-D DCT of binarized edge images and constructive feedforward neural networks

Computer-based automatic human facial expression recognition (FER) is fundamental and indispensable in realizing truly intelligent human-machine interfaces. In this paper, a new FER technique is proposed, which uses lower-frequency 2D DCT coefficients of binarized edge images and constructive one-hidden-layer (OHL) feedforward neural networks (NNs). The 2D DCT is thereby used to compress the binarized edge images to capture the important features for recognition. Constructive OHL NNs are then used to realize the mapping from the feature space to facial expression space. Facial expression ldquoneutralrdquo is regarded as a subject of recognition in addition to two other expressions, ldquosmilerdquo and ldquosurpriserdquo. The proposed recognition technique is applied to two databases which contain 2-D front face images of 60 men (database (a)) and 60 women (database (b)), respectively. Experimental results reveal that our proposed technique provides in general improved performance when compared to two other recognition methods that use vector matching and fixed-size BP-based NNs. Our method yields testing recognition rates as high as 100% and 95% for databases (a) and (b), respectively, which clearly demonstrates its promising capabilities.