Research On Super-resolution Algorithm For Face Recognition

Super-resolution refers to getting a clear high resolution image from one or multiplelow resolution images through feature extraction and reconstruction. Throughsuper-resolution technology, better images can be obtained with the original hardwarefacilities. Therefore, super-resolution technology has been widely used in videosurveillance, medical imaging and other fields. Face recognition is one of the biometricresearches because the uniqueness of human face. Hardware facilities may bring lowresolution face images, which has a significant impact on the recognition rate. Therefore,super-resolution for face recognition has became an important branch of thesuper-resolution research.In this paper, we do some research on the algorithm of face recognition with the lowresolution face image. Through super-resolution technology, we improve resolution of theface image, and then use it for face recognition. The main content of this paper is asfollows:1. The background and development of the face recognition and super-resolutionresearch both at home and abroad are introduced in this paper. Several classic methods forfeature extraction of face images are summarized and applied to the research ofsuper-resolution.2. Canonical Correlation Analysis(CCA) is a classic method dealing with therelationship of multiple data. Supervised Canonical Correlation Analysis(SCCA)comprehensively considers the within-class information and CCA method, which canreduce sensitivity to light, posture, facial expressions, and other external factors. Thence,SCCA is better for feature extraction. There are the same intrinsic geometries between highresolution and low resolution face images, and relationship learning is used to build abridge for their transformation. High resolution face image feature can be achieved through low resolution facial feature and mapping relationship, and be used for face recognition.3. To enhance the applicability of Locality Preserving Projections(LPP) insuper-resolution of face images, we propose an improved method, Correlation EnhancedLocality Preserving Projection(CELPP), which introduces the method of CCA into LPP. Inthis paper, CELPP is used for feature extraction, and relationship learning or RBF is tobuild a bridge for the transformation of high resolution images and low resolution images.Then, high resolution images can be achieved and used for face recognition. CELPP takesthe advantage of manifold learning, and also guarantees the largest correlation of the samepair of high resolution image and low resolution image.