Face Recognition using Deep CNN Models

Abstract

This paper presents the use of a deep learning convolutional neural network to recognize different human faces images. This study used three pre-trained deep CNN models to solve a three-class classification problem, which are three different person faces images. This work also aims to provide a quantitative assessment of the classifier's dependability based on performance measures. In this preliminary study, we used 120 human faces images to train deep CNN models. Using the results from three different deep CNN models (AlexNet, DenseNet201 and GoogLeNet), this study assesses the performance of the trained classifier and finds that it performs quite well in the prediction of validation data. The best performance result is AlexNet. The parameter metrics of the batch size, the number of epochs, validation frequency and learning rate are using to train CNN models of AlexNet are 10, 25, 35 and 0.001, respectively. Therefore, the mean accuracy, training accuracy, validation accuracy and error rate using AlexNet are 90.48 percent, 87.00 percent, 90.48 percent and 0.52 percent, respectively. It is concluded that the trained model is capable of classifying three different person faces well. In the future, more data could be utilized in the training network to improve the accuracy of the models. Another possible approach would be to integrate human faces images from different imaging modalities in the training to increase the variety of datasets on which a deep learning model can develop.