Super-Resolution Reconstruction Of Face Images Under Small Sample Training Set

Image super resolution technology is a method that takes a series of low resolution images as input, and uses signal processing technology to obtain high resolution image. Image super resolution technology can break hardeware constraints to achieve better image quality without the need for upgrading existing hardware conditions, So, it has been widely used for video surveillance, medical images and satellite images etc.This paper proposes a super resolution algorithm of face images based on small sample training set. Adopting a patch-based Markov Network as the fundamental mechanism, our method can get more similary result by using interpolation, and improve the efficiency of matching process to get correct patches by using visual compatibility feature as a matching criteria, and combine the weight mechanism into image reconstruction process to make training set can get fully used. Analysis and experiments show that the approach can effectively prevent the blocking effect and the partial block distortion in the process of image restoration and have better performance in the final results.This paper clearly sets forth the detail on how to get patches from images, how to generate the suitable training data and which size should we prefer for patches. This paper also focuses on the several factors that can influence the reslut, we design many experiments for each factor in order to estimate the effect that factors bring along. We also give our suggests about how to achieve better result.This paper also designs a human face super resolution prototype system based on Markov Network model. The system can use our method to make super resolution processing for face images, and it also compare the effect among different factors and afford us the data that we need.