Improving face recognition using deep autoencoders and feature fusion

Uncontrolled environments are the main challenges of real face recognition systems, recent success of deep learning and features fusion has led to various performance improvements. This paper proposes a novel scheme called feature autoencoder (FAE), where an autoencoder model is not trained directly from the raw facial images, rather it uses a fusion of features constructed by Gabor filter, local binary pattern and local phase quantisation. For each feature, a linear discriminant analysis is applied to reduce its high dimensionality and a limited adaptive histogram equalisation process is employed for contrast enhancement. The proposed scheme has been evaluated using known datasets such as AR, ORL and YALE, and the experimental results carried out on these databases have been compared using three classifiers: k-nearest neighbour, multiclass support vector machine and softmax classifier, demonstrating the effectiveness of proposed approach and parameters. The experimental results obtained and compared with recent and similar approaches on six databases: ORL, YALE, AR, extended YALE B, CMU PIE, and LFWcrop, suggest that the proposed technique outperforms similar techniques. The recognition rates got from them are 100%, 100%, 99.66%, 99.40%, 97.31%, and 90.68% respectively.