Position-patch Based Face Hallucination

In image processing, image resolution represents the quality of a given image, and provides important information. As a branch of generic image, face image plays an important role in face recognition, face detection, security monitoring, etc. However, face images captured by surveillance cameras and other equipment are usually of low resolution due to the limits of these devices and some other environmental factors. Further applications are thus obstructed, thus face image supper-resolution is needed. Face image super-resolution, by certain methods and without the need to increase hardware, can reconstruct a given lowresolution image to a high resolution image. This technique is also named face hallucination(FH) and can improve the face image resolution and has great theoretical and practical significance.This paper first presents the development of face image super-resolution technology, then the existing face super-resolution techniques are summarized. More Specifically, position-patch based face hallucination in recent years are carried out in details in this paper as well as the analysis of their problems.Position-patch based face hallucination methods aim to reconstruct the highresolution(HR) patch of each low-resolution(LR) input patch independently by the optimal linear combination of the training patches at the same position. Most of current approaches directly use the reconstruction weights learned from LR training set to generate HR face images, without considering the structure difference between LR and the HR feature space. Therefore, in this paper, we propose to use high-resolution image feature space to optimize the reconstruction weights and the concept of high-resolution reconstruction weights. Unlike conventional methods, our algorithm herein better use high-resolution training samples and spatial information. Secondly, weights optimization is solved by linear regression method to build high resolution images. Finally, experiment and analysis of the algorithm in database are conducted as well as compared with the existing methods to evaluate the performance. Moreover, the algorithm is generalized and extended to refine a given result from other FH methods. The experiments conducted on several algorithms demonstrate its performance.Meanwhile, deep learning begins to be applied in image processing, including image super-resolution. Among them, convolutional neural network(CNN), due to its superior performance, has won the attention of many researchers. This paper conducted in-depth research and analysis on related work. Currently, CNN is designed for general image super-resolution and appropriate network can achieve great results, but after research and experiments, we found that for face images, which are highly structured, the reconstruction effect, compared to the methods of using the position information, is very general. How to fully use deep learning to achieve facial image reconstruction needs to be studied on the next stage.