Research For Single Sample Face Recognition Under The Conditions Of Pose And Illumination Variant

Automatic Face Recognition (AFR) develop to nowadays, posture and illumination variation has been the main bottleneck of automatic face recognition under the conditions of uncertain environment and without user's cooperation. They are also the main reasons that automatic face recognition's further application been restricted. In addition, the single sample is a basic condition in many practical applications. So, the research of the single sample face recognition with variable posture and variable illumination has great value and has been widely concerned by researchers.This paper firstly introduces the research background of face recognition. Then around the two key issues in posture and illumination variation in face recognition: posture and illumination emendation, we make a detailed analysis of feature point detection and location and the 3D modeling technology. And the 3D data acquisition, specific face modeling and various posture and illumination emendation methods are also introduced and analyzed. The application scope and their advantages and disadvantages of respective algorithms are pointed out.Through the research of Candide-3 model and combine the advantages of initiative shape model and initiative appearance model, we propose a specific 3D face reconstructing method based on Candide-3 in this paper. This method can improve the specific 3D face reconstruction's time efficiency, and at the same time guarantee the degree of accuracy of 3D face reconstructing.Under the conditions of single sample, a new face recognition approach based on the multi-profile samples generation and face characteristics enhancement was proposed. This method make a specific face modeling by apply Candide-3 model to generate a number of face images with different postures through postures emendation of the frontal face, and use the enhanced training samples based unity of space-based singular value decomposition method to represent and enhance the features, and combined with the use of the nearest neighbor classification. This new method raises the face recognition rate under the condition of single sample.In the analysis of existing face authentication algorithm, a better method based on a Candide-3 and Support Vector Machine (SVM) is proposed, which is made used of single training sample face authentication. Through rotated the Candide-3 reconstruction model, we got.many digital faces of different postures and trained them with original samples together, and then applied SVM classification match for face verification. Experiments showed that it has almost the same recognition rate with the multi-sample face recognition algorithm by using this method of face authentication.Based on the harmonic images model, this method use the illumination coefficient vector of different face images under real illumination condition as the training sets of illumination coefficient generation algorithm firstly. Then, improved equal-error competitive learning algorithm is used to calculate the various illumination environments'center illumination coefficient vector. At last, random perturbation factor is added to center illumination coefficient vector to generate different illumination vector. The experiment's results show that this method can generate illumination coefficient vector which is a simulation of realistic illumination and provides the basis for generating digital face of various illumination effect.In order to better tolerate illumination and disadvantages introduced by posture adjustment, this paper proposed a method which use single sample reconstruction digital faces of different posture and illumination as training sets, and then build its face HMM (Hidden Markov Model). Under the constraints of the single sample, this method use a particular face of the 3D models, to generate new human face images with different illumination and posture. In order to effectively identify the human face images under different posture and illumination conditions, this method builds the user's unique face HMM and uses different illumination and posture face images as the model's training data, the model contains the observed value distribution under different illumination and posture condition and transfer information of status. HMM is a double random process, it can better tolerate observed value's deviation, which can reduce the dependence on illumination and posture adjustment. We need not to do posture emendation and illumination compensation on images to be recognized after the contraction of the model, and experiments showed that this method can improve the recognition rate and reduce the processing time to identify significantly at the same time, it can adapted to large-scale, real-time face recognition application better.Finally, the paper completed the design and development of a test platform. The software implementation of the test platform was described, and the operation interface was presented. The experimental platform almost implements all the key algorithms mentioned in all chapters, the stabilization and recognition results are also good on several face database. And propose a system solution of candidates identification based on automatic face recognition method to meet the practical need of the verification of candidates who take part in examination.