Novelty Detection Applied to Recognizing Facial Expressions in Video Stream
Novelty Detection; Artificial Neural Networks; Viola-Jones; Kanade-Lucas-Tomasi; Principal Component Analysis;
This work investigates the capacity of the Multilayer Perceptron (MLP) and Radial Basis Function (RBF) networks in the task of Novelty Detection (ND) in the recognition of facial expressions in video stream. The video data set used is produced by professional actors in the studio with basic affective states of the human face. The Viola-Jones, Kanade-Lucas-Tomasi (KLT) and Principal Component Analysis (PCA) algorithms are used in the pre-processing phase to extract features from the face. The results evaluate the performance of the MLP and RBF networks in the ND task, using new facial expressions compatible with those used in the training phase and also examines the capacity of the networks in ND using the faces of actors never before seen by the networks. In this process, the MLP and RBF networks have an accuracy of 98% for classification task, 69,1% and 91,5% for ND with data similar to the data from the training phase and 100% for ND with totally new data. Thus, this research brings together methods and techniques applied in ND using Artificial Neural Networks (ANN) aiming at the production of interactive cognition systems in the field of affective computing, based on techniques of Artificial Intelligence (AI) and Computer Vision.