The Research On A Single Training Sample For Face Recognition Per Person

Single sample for face recognition technology has become one of the important topics in the field of computer vision. It involves many fields, such as omputer vision, pattern recognition, machine learning, artificial intelligence, etc. In terms of the ID card identification,electronic passports and law enhancement, due to the large amount of face data and limited storage equipment, only one image can be stored for each person's face. Because each person only has one sample, it presents a great challenge for face recognition. Therefore, the study of a single training sample for face recognition technology has great significance.Firstly, this paper studied a single training sample method for face recognition based on subspace and traditional manifold learning. In the method based on the subspace, it introduces the principal component analysis (PCA), projection-combined principal component analysis((PC) A), fisher linear discriminant analysis (FLDA) and locality preserving projection (LPP)algorithms. Although these methods can be used for face recognition from a single training sample per person, it is difficult to achieve the desired results of recognition rate. Most existing discriminative manifold learning algorithms assume that face samples from different classes are considered as a manifold, and then extract features. Although these methods consume less time in the feature extraction and recognition, the recognition rate is not very high. Secondly, we briefly introduce DMMA (Discriminative Multi-Manifold Analysis)algorithm. Unlike traditional manifold learning methods, the algorithm models the local patches of each face image as a manifold and each class can get a facial feature matrix by manifold learning. The algorithm uses the local geometric features of the image. Despite the fact that it obtains a higher recognition rate, but consumes more time in the feature extraction stage and can't meet real-time face recognition system requirements.In order to effectively extract facial expression feature in the single face recognition system, this paper proposes a novel method using the fusion Uniform LBP features and DMMA features. First, each face image is partitioned into several nonoverlapping patches to form an image set for each sample per person. Second, the Uniform Local Binary Pattern(Uniform LBP) operator is used to extract image histogram of each image set. So the histogram of each image subsets forms a statistics manifold. Third, this paper applies DMMA algorithm to obtain the low-dimensional face image feature. Finally, it uses the reconstruction-based manifold-manifold distance to identify the unlabeled face images. This method combines the face image texture information and geometrical information.Experimental results show that the algorithm is superior to the general recognition DMMA algorithms on the AR database and ORL database.Finally, In order to extract effective features of face image in the complex environments,a novel method by using the fusion HOG features and discriminative multi-manifold analysis features was presented. A new adaptive method was proposed to calculate similarity between patches of the face image. First, we partition each face image into several nonoverlapping patches to form an image set for each sample per person. Second we use Histogram of the Oriented Gradient (HOG) operator to extract image histogram of each image set. The histogram of each an image set forms a statistics manifold. Then we apply DMMA algorithm to obtain the low-dimensional face image feature. Last we apply the reconstruction-based manifold-manifold distance to identify the unlabeled subjects. Experimental on the AR database and FERET database results show that the proposed algorithm has better recognition rate on the face images with variations in rotation illumination.