Face recognition classifier based on dimension reduction in deep learning properties

Abstract

Nowadays, with the increasing use of biometric data, it is expected that systems can give successful results against difficult situations and work robustly. Especially, in face recognition systems, variables such as direction of light, facial expression and reflection are making difficult to identify. Thus, in recent years, Convolutional Neural Network (CNN) models, which are deep learning models as an alternative to traditional feature extraction and artificial neural network methods, have begun to be developed. In this work, for face recognition, VGG Face deep learning model is compared with our proposed model which uses Multi Layer Perceptron (MLP) classifier and reduced deep features by principal component analysis. The Kinect RGB image dataset belonging to 40 people with different facial expressions and lighting conditions has been tested with 4-fold cross validation method. While 97.18% classification ratio was achieved with the first model, 100% recognition accuracy has been obtained by the second model. The results show that deep learning achieves a high performance in face recognition under different light and expression conditions, however, the proposed classification method based on dimension reduction in deep learning properties achieves better performance.