Research On Learning-based Face Superresolution Reconstruction

The resolution of an image may reflect its clarity.The number of pixels per unit area on the digital image determines its resolution and it is a key indicator to evaluate the quality of the image.Super-resolution reconstruction of face refers to a kind of software technologies of obtaining a high-resolution face from one low-resolution face and it’s also called face hallucination.Based on the human face playing a vital role in computer vision,machine learning and other fields, face hallucination has been a hot research topic and has broad prospect of application in multi-media communication,face recognition and video surveillance,et al.In this thesis learning-based face hallucination is studied.This thesis proposes two face hallucination algorithms through the study of face hallucination method based on similarity constraints and face hallucination method based on neighborhood embedding.One is an improved face hallucination algorithm based on similarity constraints and iterative learning and the other is an improved neighbor-embedding face hallucination algorithm based on the position.The two contributions described in this thesis can be summarized as follows.1.The traditional learning-based image super-resolution reconstruction algorithm has high space complexity.Thus, this thesis proposes a face hallucination algorithm based on similarity constraints and iterative calculation.Firstly,for the face to be reconstructed,make use of principal component analysis(PCA) and choose a certain of high-resolution(HR) and low-resolution(LR)face pairs with the highest similarity with it to form the new training set in the initial face database for the subsequent iterative learning and reconstruction.Then only one HR-LR face pair in training set is involved in the training process in each iteration,reducing the space complexity.Otherwise,due to the defect of the one-to-many mapping from the low-dimensional space to the high-dimensional space in manifold learning,the similarity constraints method is adopted to calculate the reconstruction weight in each iteration in order to reduce the reconstruction error brought by this inconsistency.Experimental results show that the improved face hallucination algorithm not only has smaller space complexity, but also has better subjective and objective effects.2.The search is to study face hallucination method based on neighbor embedding.Firstly like above,to search the new HR-LR face set from the original face training set by PCA projection method with the highest similarity for each face to be reconstructed. In order to represent theinformation of human face better, this thesis proposes a new method to extract facial feature and a new method to search neighborhood facial blocks with a joint learning of the HR-LR facial block’s features. Considering the structural features of face images,to adopt a kind of classifying human face image blocks based on position.Taking the inconsistency between high-resolution representation manifold(HRM) and low-resolution representation manifold(LRM) in manifold learning into account,LR facial block neighboring the reconstructed facial block in a certain location and the corresponding HR facial neighborhood blocks are projected to a common manifold space to regulate the manifold.And then the following adopting the local linear embedding(LLE)algorithm for solving reconstruction weight coefficients is based on the above common manifold.Experimental results show that the improved face hallucination algorithm based on neighbor-embedding method and the position comparing to the traditional neighbor-embedding algorithm,the former could generate a final HR result with better quality objective and subjective visual effects.