Local feature representations for facial expression recognition based on differences of gray color values of neighboring pixels
Mohammad Shahidul Islam · NIDA Wisdom Repository · 2013
A fruitful way for creating any innovative system in a computing environment is to integrate a sufficiently user-friendly interface with the ordinary end user. Successful design of such a user-friendly interface, however, means more than just the ergonomics of the panels and displays. It also requires that engineers precisely define what information to use and how, where, and when to use it. Facial expression as a natural, non-intrusive and efficient way of communication has been considered as one of the possible inputs of such interfaces. The research interest in facial expression recognition has grown due to its potential applications. Many local feature representations such as Gabor filters, LBP had been proposed for facial expression recognition. However, accuracy rates and running time of facial expression recognition achieved by these representations have yet to be improved. The work of this thesis aims at designing a robust and effective Facial Expression Recognition (FER) system that uses simple but effective gray scale invariant local descriptors for face expression recognition. The local feature pattern at a pixel is computed based on differences of gray color values of its neighboring pixels. The pattern represents the changes of gray color values of pixels in its surrounding area. Four alternative local feature representations are proposed depending on the numbers of considered directions and neighboring pixels. To create the feature vector for an image, the facial image is divided into blocks and the histograms counting the occurrences of all possible local patterns for all blocks are computed then concatenated. A variance-based feature selection method is also proposed to reduce the length of the descriptor, thus, help the running time for feature extraction, training and classification time. It can also be shown that the more the numbers of considered directions and neighboring pixels, the richer the information the local pattern represents and so the higher accuracy of facial expression recognition achieved from using the local pattern. However, longer pattern length and longer processing time will be needed. Experimental results show that the proposed feature representations along with Support Vector Machine are more effective than some other well-known local feature representations for facial expression recognition. The feature representations are also easy to compute and suitable for real time applications.