10th International Congress on Information and Communication Technology

Authors - Robert, Tubagus Maulana Kusuma, Hustinawati, Sariffudin Madenda
Abstract - The process of forming a good dataset is a very decisive step in the success of a facial expression recognition/classification system. This paper proposes 24 scenarios for the formation of facial expression datasets involving the Viola-Jones face detection algorithm, YCbCr and HSV color space conversion, Local Binary Pattern (LBP), and Local Monotonic pattern (LMP) feature extraction algorithms. The results of the 24 dataset scenarios were then formed into five dataset categories to be used as training datasets and testing of two Machine Learning calcification models, namely Support Vector Machine (SVM) and Convolutional Neural Network (CNN). The SVM classification model is designed using four different kernels: radial, linear, sigmoid, and polynomial basis functions. Meanwhile, the CNN classification model uses the MobileNetV2 architecture. From testing the five categories, the best accuracy result is 83.04% provided by the SVM classifier that uses the sigmoid kernel and a combined dataset of LBP and LMP features extracted to focus only on the facial area from the results of the Viola- Jones face detection algorithm. In addition, for the CNN classifier, the best accuracy was obtained at 82.14% by using the Y-grayscale dataset which also focuses only on the facial area but without the feature extraction process. The results of the best accuracy for the two classifiers show that the face detection stage plays an important role in the facial expression recognition/classification system. The LBP and LMP algorithms are good enough to use for feature extraction in forming datasets in the SVM classification model.