Research On High-Performance Super Resolution Reconstruction Algorithm

The real-time acquired high resolution image information plays a very important role in many fields such as medical treatment, video monitoring, remote sensing satellite imaging, etc. However, due to the various factors which include noise, imaging environment and so on, the imaging equipment is always affected in the process of image acquisition and results in collecting the poor quality and low resolution images. And the traditional method to improve image resolution via improving the hardware equipment will generate a series of new problems such as increased equipment size, increased cost of manufacturing and the difficulty of processing. Therefore, people want to improve the image resolution by the software method without changing the hardware device. The super resolution image reconstruction technology arises at the historic moment.However, in a variety of super resolution algorithms, whether single frame or multiple frames method, each of them has certain deficiencies. The single frame method runs faster than multiple frames but has a bad result, and multiple frames method has a better effect than single frame but its processing time is quite slow. When reconstructing a high quality image efficiently, especially for the larger size images, the two types of algorithms could not balance the resolution improvement and the reduction of processing time well. In order to improve the resolution and reduce the processing time of algorithm, the specific super resolution methods have been researched, and an effective high-performance super resolution reconstruction algorithm which is a hybrid of multiple frames Variational Bayesian reconstruction and single frame Dictionary Learning reconstruction method is proposed to reconstruct a high resolution image in this article. Firstly, the multiple frames Variational Bayesian algorithm is studied deeply, and the unknown high resolution image, the acquisition process, the motion parameters and the unknown model parameters are built together in a single mathematical model. Then all the parameters are estimated precisely to reconstruct a high resolution image from multiple low resolution images. For purpose of processing time improvement, the ideal of partitioned management is introduced into the algorithm process. The input of the low resolution image is blocked, and every image block information is estimated using Canny technology. If the information is more, it is reconstructed via the Variational Bayesian method, otherwise reconstructed via the bicubic interpolation. Secondly, the single frame Dictionary Learning algorithm is researched as well, and the principal component analysis(PCA) and K-mean and singular value decomposition algorithm are utilized to train the high and low resolution dictionary, and then the orthogonal matching pursuit algorithm is applied to reconstruct a high resolution image. A new Kernel based Principle Component Analysis algorithm is utilized to replace the PCA for the dictionary training process so as to compress data and improve the processing speed effectively. In addition, a specific filter is used for feature extraction to obtain more high frequency information which represents more image resolution. Finally, in order to improve image resolution and running speed simultaneously, especially for the larger size images, an effective high-performance super resolution reconstruction algorithm which combines multiple frames Variational Bayesian with single frame Dictionary Learning method is developed in this article; that is to say, an high resolution image is reconstructed via multiple frames Variational Bayesian method at first. Then by taking the above high resolution image as input, a higher resolution image can be rebuilt utilizing single frame Dictionary Learning method. In each Reconstruction step, the input image is blocked and extracted the edge information. Every image block is judged by the above information, and the block is reconstructed via the Variational Bayesian or Dictionary Learning method when its information is more, otherwise it is reconstructed via the bicubic interpolation. The high-performance super resolution reconstruction algorithm is able to improve image resolution and running speed simultaneously.This article has done a lot of experiments on the Variational Bayesian method, the Dictionary Learning method and the high-performance method, and carried on multiple sets of image data experiments and data comparisons. In this experiments, the reconstruction results and experimental data proved that the proposed algorithms have a good validity and practicability and put forward an useful ideas to solve the difficulty of the existing super resolution algorithm.