Face Super Resolution Algorithms Research

Human face has been a hot research point in computer vision, pattern recognitionand image/video processing field and it is a credible source for biometric. In videomonitoring systems, face images acquired by camera are usually in low-resoltion,blurring and difficult to recognize, which sets a barrier for police department to crackdown criminal cases. Face hallucination is used to improve the resolution of faceimages and make recognition easy and fast.Our research is to recover a high-resolution face image from a low-resolution one.For this purpose, it is required to draw some important characteristics like contours anddetails. Our method is toughly divived into two parts: global reconstruction and localdetail compensation.Our innovation points are:1) In the global reconstruction phase, we build two eigenface dictionaries in twodifferent scales using sample face images. In low-resolution scale, coeffieentsvector is obtained with the help of Matching Pursuit algorithm.2) In the local detail compensation phase, we divide all face images includedglobal reconstruction result, high-resolution sample face images and lapacianpryriam image generated by them into patches.We untilize Markov RandomField to form a detail face image to compensate the high-frequencycomponents in the global face. The simplified pairwise compatitiblityfunctions accelerate the energy convergence and reduce computation load.A series of experiments from both subjective and objective aspects demonstratesthat our method is able to achieve satisfactory results. And we develop a PC softwarewhich implemented our method and make it possible for practical use.