Research On Learning-based Face Superresolution Reconstruction

With high-resolution images and videos being popular, the low-resolution images and videos output by old image acquisition devices are more and more difficult to satisfy people's needs. As well as some specific fields always have a higher demand for resolution, for example, face recognition, criminal investigation and medical diagnostics etc. And face, as is often an area of concern, the superresolution algorithm aiming at face becomes critical. To solve this problem, this thesis take a deep research on the face super-resolution reconstruction algorithms in real image sequences, from the perspective of the image super-resolution reconstruction, study and improve the existing algorithms, and also explore the application of this technology in real scenarios. The specific contents are as follows:Firstly, we study and improve the face reconstruction algorithm which based on weighted sparse representation. The shortcoming of this algorithm is that the best regularization parameter is changing all the time while the input face's noise level changes, and the tuning of this parameter is timeconsuming and inefficient. Therefore this thesis takes a pre-reconstruction process before the real reconstruction, which could automatically estimate the noise level and scale parameter according to Maximum A posterior criterion. Finally through experiments we show the automatic regularization parameter estimation algorithm has full feasibility and effectiveness.Secondly, for the high intensity noise, we propose a face super-resolution method based on similarity selection and representation. Because of the reconstruction results of existing method is not good enough when noise appears. So we use similarity instead of the sparse feature and the whole framework consists of two parts, the first part is the face selection method based on global similarity. The experiments based on 2D-PCA indicating that it can fully guarantee the reconstruction effect while dramatically reduce the number of face in the database. The second part is the super-resolution method based on local similarity. By considering the neighbor patches and parallel the algorithm structure, this method have the ability to withstand high intensity noise to a certain extent.Thirdly, we introduce machine learning methods to solve the problem that the methods above have limited capacity of restoration under real scenarios. On the top of Two-Step face hallucination methods, we propose the face super-resolution method based on extreme learning machine. The real environments experiments show that this algorithm has the strongest recover ability, compared to other methods.Finally, we explore the role of this technology in practical applications. We make the face calibration and extraction standard and study the Poisson image editing method, and on top of this we propose a complete face restore scheme. The experiments are taken under real street surveillance videos, and the randomly selected pedestrians are tested and achieve a very good splicing effect.