Multilayer Position-patch Based Face Image Super Resolution Reconstruction

Based on image processing technology,face image super-resolution reconstruction uses one image or several low resolution images to reconstruct the high frequency information of the image,so as to obtain a high resolution image,it has been widely used in many areas which have a high demand in image resolution.In the case of small samples,how to make full use of the information provided by the training set to make the reconstructed images more similar to the original images has always been the target of face image super-resolution reconstruction.To solve this problem,a new multilayer position-patch based face image super-resolution reconstruction is proposed in this thesis,which improves the similarity between the reconstructed image and the original image.The work of this thesis are as follows:First of all,to find the main problems,this thesis studied the development of face image super-resolution reconstruction,then proposed this thesis's research direction.To implement the algorithm,this thesis also studied some typical face image super-resolution reconstruction algorithms.Secondly,owing to some training set have small samples,this thesis proposed a method to optimize the training set.A training set includes many high resolution images and low resolution images,to reconstruct a high resolution image,we flip these training images from left and right respectively.Then we crop and align these images,extract one step degree and two step degree features of these images,divide these face images and their feature images into many patches.Using the non-local similarity of the low resolution feature images,the algorithm searches the similar patches of the reconstructed patch.The training set are formed by the patches at same positions.Finally,in order to make full use of the information of the training set,this thesis uses high and low resolution images simultaneously to build weight matrix.The constraints of weight matrix can reduce local ghosting of the reconstructed image effectively.Multilayer model is built by the images which are divided into different scale of patches.The model can reflects the degradation process of an image better.The experiments are based on the databases of FERET,CAS-PEAL-R1 and FEI Compared to the interpolation-based methods,learning-based methods and location patch based methods,the peak signal to noise ratio and the structural similarity index of these images are better.The algorithm proposed in this thesis can make full use of the information of training set and obtain better reconstructed face images.