Local binary pattern network: a deep learning approach for face recognition

Deep learning is well known as a method to extract hierarchical representations of data. This method has been widely implemented in many fields, including image classification, speech recognition, natural language processing, etc. Over the past decade, deep learning has made a great progress in solving face recognition problems due to its effectiveness. In this thesis a novel deep learning multilayer hierarchy based methodology, named Local Binary Pattern Network (LBPNet), is proposed. Unlike the shallow LBP method, LBPNet performs multi-scale analysis and gains high-level representations from low-level overlapped features in a systematic manner. The LBPNet deep learning network is generated by retaining the topology of Convolutional Neural Network (CNN) and replacing its trainable kernel with the off-the-shelf computer vision descriptor, the LBP descriptor. This enables LBPNet to achieve a high recognition accuracy without requiring costly model learning approach on massive data. LBPNet progressively extracts features from input images from test and training data through multiple processing layers, pairwisely measures the similarity of extracted features in regional level, and then performs the classification based on the aggregated similarity values. Through extensive numerical experiments using the popular benchmarks (i.e., FERET, LFW and YTF), LBPNet has shown the promising results. Its results out-perform (on FERET) or are comparable (on LFW and FERET) to other methods in the same categories, which are single descriptor based unsupervised learning methods on FERET and LFW, and single descriptor based supervise learning methods with image-restricted no outside data settings on LFW and YTF, respectively. --Leaves i-ii.