2019-11-11

This thesis provides a technique for Facial expressions from still images of human faces based on Dual-Feature-Selection (DFS) concept. This work includes substantive problems such as; variety in facial datasets, the instances consisted in each dataset, the number of facial expressions experienced, utilizing an accurate technique for face detection, utilizing multiple feature selection methods for selecting the minimum number of most characterized facial features, utilizing multiple classification algorithms, and diversity of structure models. A newly collected dataset named (KURD) dataset along with two existed datasets CK+ and JAFFE are used as a data source for validating this approach. Viola-Jones algorithm is used for face detection. Correlation Feature Selection (CFS), Information Gain (IG), Gain Ratio (GR) and Relief Feature Selection (ReliefF) are used as a feature selector. K-Nearest Neighbor (KNN), Multi-Layer Perceptron (MLP), Decision Tree (DTJ48) and Naïve Bayes (NB) algorithms are used as a classifier of expressions. For discovering a suitable model of each classifier, hundreds of tests are performed dynamically. The results of various models are compared, and it is perceived that these techniques provide high rates of recognition. The accuracy of classifiers was quite high for six classes of KURD and CK+ datasets. Experiments are performed using Matlab programming and Weka Data Mining Tool.