Analytic Approach To Face Emotion Recognition With SVM Kernels

摘要

Face emotion recognition is one of the challenges known with emotion recognition and it has received much attention during the recent years due to its application in different fields. SVM kernels were adopted to increase the robustness of face emotion recognition systems and to identify the most suitable kernel for emotion recognition. This paper uses radial basis function, linear function, sigmoid and polynomial function to identify the six basic emotions and neutral inclusive. In an attempt to achieve this aim the following steps were taken; collection of face emotion images, image pre- processing, features extraction and classification. Face emotion database was created by taken emotional photographs of persons who willing volunteer to help in this paper. The database contains 714 images from 51 persons. However, the photographs were converted from colored images to grayscale images for uniform distribution of colors. Relevant features for classification were extracted from the processed images such as the eyelids, cheeks, nose, eyebrows and lips. Our face emotion database was splitted into two dataset: training set and testing set. SVM classifier used images in the training set to train while images in the testing set were used to test SVM models. The evaluation of the system was performed on MATLAB using classification accuracy and classification time to identify the most suitable kernel for the system. The results obtained shows that sigmoid outperformed other kernels in terms of classification accuracy with overall performance accuracy of 99.33% while polynomial achieved the shortest classification time. In the future, we intend to investigate other classifiers for face emotion recognition and to classify more emotions.