As a cross-disciplinary topic of many research areas (e.g. computer vision, pattern recognition, image processing, machine learning, at.al.), face recognition has attracted more and more researchers’ attention in recent years. In the past few decades, much progress has been made in this area. Many performance evaluation results show that the best face recognition algorithm has achieved high performance under the ideal constrained environment. However, under the unconstrained environment, the variations caused by the change in factors such as illumination, pose, expression and so on, could be larger than those caused by identity change. This makes face recognition systems unable to obtain a desirable recognition rate. In order to develop a robust and practical face recognition system, many key problems need to be solved, such as effective feature representation and robust recognition algorithm. Feature representation is generally regarded as fundamental of face recognition algorithm and has received much attention. This thesis reviews the widely used feature representation methods in face recognition from global, local and fused aspects. Local feature is robust to the variations due to lighting, pose, expression or other factors. Thus it attracts much attention in face recognition. This dissertation proposes a novel local feature representation by fusing local phase quantization and ordinal measures features. Related literatures state that global and local features are both essential to face recognition. Basing on this conclusion, this thesis proposes to combine global and local features by a parallel manner, and obtains the boosted face performance. On the whole, the contributions of this dissertation are summarized as follows: (1) Propose a novel local feature representation named Histograms of Local Phase Quantization Ordinal Measures (HOLPQOM). The LPQ is robust to image blurring and the OM is robust to the illumination. We fuse these two features in order to inherit the advantages of them and get a novel representation. The process of extracting this local feature is as follows: we extract the LPQ feature of the original image and then extract the OM feature of the obtained LPQ feature. The value of OM feature is binary. In order to obtain a compact representation, we encode the values of each LPQOM for four orientations at a given lobe number and a given inter-lobe distance into a single decimal number ranging from 0 to 15. Finally, we com...
修改评论