Face Verification Based On Feature Learning And Similarity Metric Learning

Face Recognition includes two tasks: identification and verification. Face verification is to predict whether pairs of images are from the same person or not. The large intra-personal variations in complex background, expression, pose and lighting have been the main challenge for unconstrained face verification. This thesis deals with the face verification. The main work of this thesis is as follows:(1) On the problem of feature extraction, we achieve extracting automatically local face feature based on unsupervised feature learning. Firstly, we learn a set of filter operators based on sparse auto-encoder model, especially with the local image patches randomly sampled pixel-wisely from the training images, which enhances the discriminative of extracted feature. Secondly, we convolve each image using these learned filter operators by applying nonlinear mapping, which allows us to obtain flexible abundant information. Thirdly, we perform pooling on these convolved images to remove most of redundant information, then whitened principal component(WPCA) is followed for reducing the pooled features and suppressing noise. Lastly, metric learning is combined with concatenated descriptors from learning, and we validate our method on the Labeled Faces in the Wild(LFW) dataset.(2)On the problems of metric learning, we develop a framework to learn weighted subspace and similarity metrics for unconstrained face verification. Firstly, we learn the weighted intra-personal covariance matrix and obtain the robust intra-subspace, by projecting the intra-subspace, the large intra-personal variations will be reduced. Secondly, we add the prior information of faces pair to the similarity metric model, and improve the regularization item. Which guarantees the further robustness to the intra-personal variations and the discriminative power to inter-personal variations for the learned metrics. Experiments show that our method achieves very competitive face verification performance on the widely used Labeled Faces in the Wild(LFW) dataset.