Research On Video Face Image Super-Resolution Reconstruction Technique

Super-resolution(SR)reconstruction technology is a key research technique in the field of image enhancement.It can produce high-resolution(HR)images with more details based on low-resolution(LR)images without changing the hardware devices.Hence,SR technology is widely used in many industries that require high resolution of digital images or video,such as medical,security and criminal investigation,digital high definition and so on.In this paper,face details in surveillance video are research priorities.Through face reconstruction techniques,a HR face image with distinct facial structure and fine detail can be generated from a single or a group of LR images obtained from the same scene,which can provide more information for many real-world applications such as face recognition.This paper first introduces the basic theory of super-resolution reconstruction technology,including the theoretical observation model of SR reconstruction,the classification of generalized SR algorithm,the basic link of SR reconstruction of video image and the evaluation method of reconstructed image quality.Secondly,the existing face SR algorithms are analyzed and introduced,and the problems of these methods are pointed out.Based on this,this paper proposes a novel hybrid face SR reconstruction framework.The framework first selects images with relatively clear,frontal,and bright lighting from the various poses of face images taken by the video to participate in the multi-frame face SR reconstruction,which can effectively reduce registration pressure.Then the weight matrix based on canonical correlation analysis(CCA)theory are structured to weighted combine the reconstructed results of LR video face sequence via single frame and multi-frame SR algorithm.Specifically,the method uses the CCA theory to analyze the image patch with more relevance to the original scene face image in the result of the multi-frame method and then add to the result of single-frame method as the constraint that enhance the consistency between the reconstructed face image and the original image,making the reconstructed HR face image can maintain the overall appearance and individual details at the same time.Finally,a large number of experiments on four public video datasets are performed to certify the performance of the proposed method.The experimental results show that the proposed method presents competitive performance both in terms of visual perception and evaluation of objective indicators in the reconstruction of LR video face image sequence.Compared with the advanced SR method,the PSNR gain value of the image reconstructed by the proposed method for the same input image can be as high as 0.67 dB.In order to demonstrate the role of the proposed SR method in video face recognition scenes,the face recognition experiments based on SR technology are further developed.The recognition experiment results demonstrate that the proposed face image selection mechanism and hybrid SR method can effectively improve the recognition accuracy of the lower resolution input image.Our method is able to increase the recognition accuracy by up to 3.5%,and it can also bring about 1.1% improvement in recognition accuracy when the input image resolution is 24×24.