One Sample Face Identification Under Complex Environment

As one of the most classical biometric methods, face identification technology hasraised much attention in the field of computer vision and pattern recognition. With themerits of both high accuracy and low intrusiveness, face identification has a variety of ap-plications such as financial security, smart cards recognition and airport surveillance. Aconsiderable amount of research has been devoted to human face identification. There aremany successful methods for the problem of automatic face recognition. Unfortunately,most of these approaches severely rely on the size and representative of the training dataset.While these methods focus on improving the recognition accuracy, they ignore the problemoriginated from the dataset without sufficient training samples. In real-world face recogni-tion system, because of the difficulty in grasping face photographs or limited ability of thesystem for data storage, there is usually only one image for each individual. On the otherhand, storing only one sample for each person in the database has many advantages: theyare easy to collect, save storage space and reduce computational cost. Therefore, developingan efficient method for the task of face identification with one-sample per person is of greatsignificance.In this thesis, we study the problem of one-sample face recognition under complex en-vironment. For a face identification system, the most critical obstacles towards real-worldsapplication are often caused by the disguised, corrupted and varying illuminated images in alimited sample set. To solve the above problems, we propose a two-steps scheme by posingthe one-sample face identification problem as a representation and matching problem. Forthe representation step, we present a novel manifold embedding method, namely sparsediscriminative multi-manifold embedding (SDMME), to learn the intrinsic representationbeneath the raw data. We construct two sparse graphs based on two structured dictionariesfor measuring the similarity between samples. Multiple feature spaces are learned tominimize the bias of data points from the subspace of the same class and maximize thedistance to the subspaces of other classes simultaneously. For the matching step, we presenta distance metric based on the learned manifold structure to identify the person. Extensiveexperiments demonstrate that the proposed method outperforms other state-of-the-artmethods for the problem of one-sample face identification, while the robustness withocclusion and illumination variances highlights the innovational contribution of our work.