Development of an Adaptive Trait-Aging Invariant Face Recognition System Using Convolutional Neural Networks

Abstract

Trait-aging causes large intra-subject variations and negatively impacts on the accuracy of face recognition systems. With lapse in time, it returns a false match when identifying and authenticating the same individual. In this paper, an augmented database using pre-processing measures, a feature extracting algorithm, a modified classifier and an adaptive trait-aging invariant face recognition system were developed. The database was populated using data augmentation processes on the subjects got from the Face and Gesture Recognition Network Aging Database (FG-NET AD). The data augmentation process increased the amount of available data, thus making it suitable for deep learning operations. Cross-validation operations were also performed on the augmented dataset and each face image resized to fit into the adopted Convolutional Neutral Network Model (CNN). The CNN model was re-trained with FG-NET AD after it augmented, cross-validated and divided into mini-batches. The re-training was achieved using transfer learning technique. The resulting adaptive face recognition algorithm was validated using the data reserved for testing to ascertain its performance. The performance of the proposed adaptive face recognition system was evaluated using cumulative score curve and Mean Square Error (MSE). It had a 99.90% testing accuracy, a testing loss of 0.003%, Mean Absolute Error (MAE) of 0.0637, and Mean Squared Error (MSE) of 0.0158. Thus, out-performing the best result recorded in the available literature using similar database.