The Key Technology Of Face Recognition Under Uncontrolled Conditions

Face recognition (FR) technology has been significantly improved in the nearly fifty years. Under controlled situations, face recognition can perform well in commercial applications. However, the recognition efficiency deteriorates dramatically under uncontrolled situations. Negative factors include data collection, feature extraction, as well as the generalization ability of kernel algorithm. Therefore, FR research under uncontrolled situation becomes a challenge in the fields of computer vision and artificial intelligence.This thesis systematically studies many topics in FR, especially the relevant solutions of preprocessing, feature extraction and classification. We have got into deep researches with regard to the variations of pose, expression, occlusion and low-quality image problem, and proposed several innovative ideas and effective solutions. The contributions of this thesis include four innovations and three improvements.The first innovation is the preprocessing method based on key-points and Gaussian Weights. Currently, texture feature and shape feature are two important feature descriptions with their respective characteristics. So a new challenging is how to combination of these two features.we proposed a new facial feature that is a fusion of texture and shape in2D space. In this new method, five key-points are extracted from central points of key facial organs. Facial pixel values are assigned dynamically according to their distance to the relevant key-point. Combining both texture and shape characteristics together can not only make the new feature (compared with raw facial feature) more efficient, but also provide a valuable reference for raw feature preprocessing.The second innovation is the feature extraction method based on factor analysis. We innovatively proposed a framework of factor analysis based on deep study of subspace analysis and factor analysis.The framework consists of two important principles. One is to apply different transform models in different scenarios; the other one is to search for an optimal weighting scheme that fits best for the current classification. Meanwhile, we have proved that some classical algorithms like PCA, LDA and LPP are specific cases within our framework. By utilizing this framework as a tool, we proposed two new feature extraction algorithms called Factor analysis based LDA and Factor analysis based LPP.The third innovation is multi-instance face recognition method. All methods based on multiple parts or blocks are summarized as multi-instance method in this thesis. Two critical problems in multi-instance face recognition are solved:the definition of multi-instance and the fusion of multiple instances. The multi-instance method combined with sparse representation can achieve satisfactory experimental results.The fourth innovation is to propose the RSIF method that is resolution-invariant. RSIF method is based on the linear combination in a high-and low-resolution training dictionary pair. It is a new solution for feature extraction and pattern recognition, which distinguishes from typical feature extraction methods.The first improvement is for sparse representation classification. Sparse representation classification (SRC) is currently a popular method that is famous for its robustness to facial noise, occlusion and illumination variation. However sparse representation has two disadvantages, small sample size and dense correspondence. In order to overcome these two problems, we proposed an improved method by combing SRC and multi-instance learning.The second improvement is about3D face modeling. Morphable Model and SFS are two main methods used in3D face modeling. This thesis presents an improved method that is a fusion of them two to further enhance their merits. Further more, textural compensation is applied to correct the modeling error of2D virtual face caused by pose variation in3D-2D face recognition.The third improvement is focused on face super-resolution (SR) problem. We propose a modified face hallucination method based on non-local similarity and multi-scale linear combination (NLS-MLC). Different from most of the established learning-based algorithms, non-local similarity is considered in the proposed model to enhance the performance under noise and other complicated situations. The aim of applying this SR algorithm to face images is not only for visual quality enhancement, but also for recognition accuracy improvement.