Research On Face Super Resolution Reconstruction In Image And Video Sequences

With the rapid development of video and image processing technology in recent years, thereis a high demand for high-resolution (HR) images, video and image sequences. Super resolution(SR) reconstruction technology can break through the physical limitations of the image, it hasproved to be useful in many practical applications, such as penal reconnaissance, facerecognition, medical diagnosis, etc. Super resolution reconstruction can predict thehigh-resolution image through one image or multy low-resolution (LR) images.The mainresearch methods based on the reconstruction method and learning-based method. Face imagesuper-resolution is super resolution technology in face image, mainly include face recognition,face recovery, facial expression analysis, face video transmission and other fields.This thesis focus on the single and multi-frame face super resolution reconstruction research.For single frame low resolution face image, it uses the learning-based method to get the highresolution reconstruction image; for multi-frame low resolution face images, it fusescomplementary information of several low resolution images to get a high resolution image. Thedissertation investigates the key issues of face image super resolution reconstruction. It integratesinto the Bayesian maximum a posterior（MAP）frame. The main contributions and innovationpoints of the dissertation are as follows:For single frame face image super resolution reconstruction, in order to decrease the facehallucination algorithm’s noise, a new learning-based super-resolution algorithm is presented.Pyramid is used to extract the facial gradient distribution features, establish the standard facetraining database for the study model; these features are combined with pyramid-like parentstructure to predict the best prior. And then through the Bayesian MAP frame capture the high resolution face image. Experimental results show that the proposed approach synthesizeshigh-resolution faces eliminates the noise with better visual effect, and the peak signal-to-noiseratios is improved about1.19dB to2.4dB compared with some existing face super-resolutionmethods.In order to decrease noise and save time, a novel face hallucination method is proposed inthis thesis for the reconstruction of a high-resolution face image from Sparse Representation. Bytraining two dictionaries for the low-and high-resolution image patches, the method efficientlybuilds sparse association between high-frequency components of HR image patches and LRimage feature patches, and defines the association as a prior knowledge to guide super-resolutionreconstruction with respect to their own dictionaries. The learned Dictionary pair is a morecompact representation of the patch pairs, compared to MRF approaches, which simply sample alarge amount of image patch pairs, reducing the computational cost substantially. Experimentsshow that the proposed method generates higher-quality images and costs less computationaltime than some recent face image super-resolution (hallucination) techniques, and achieves muchbetter results than many state-of-the-art algorithms in terms of both PSNR and visual perception.For muti-frame face image super resolution reconstruction, Registration of consecutiveframes and selection of frame is quite essential in multi-frame image super-resolution. In order todecrease the larger inter-frame motion, an adaptive frame selection principle is proposed, whichjoints the Optical flow algorithm for motion estimation registration and the super-resolution isdesigned. First, using the Optical flow algorithm to calculate the inter-frame motion estimation,designing an adaptive frame selection method to discard some of the larger inter-frame motionframes, and then through sub-pixel image registration to calculate the accurate motion estimationparameters, and finally combine the MAP method for image super-resolution which take intoaccount two iterations of the difference between the resulting image vectors of the next iterationalgorithm. Experimental results show that this method not only achieve sub-pixel accurateregistration, but also achieve better results in the visual effects and the peak signal to noise ratio.