We propose a novel and robust deep learning method for face recognition, which uses effective image representations learned automatically to handle big data. There are two stages of the deep learning architecture in real-time application. First, in the offline training procedure, we train a stacked denoising autoencoder to learn generic image features from 80 million images from the Tiny Images Dataset used as auxiliary offline training data. Second, in the supervised object recognition procedure, we construct five layers as a feature extractor to produce an image representation and an additional classification layer, which we can use to further tune generic image features to adapt to specific object recognition by online training of the corresponding objects. Comparison with the state-of-the-art face recognition methods shows that our deep learning algorithm in face recognition is more accurate and it is a perfect processing tool for the big data problem.
© 2015 Optical Society of AmericaPDF Article