Research On Learning-based Image Super-Resolution Reconstruction Algorithms

With the rapid development of computer vision applications,Single image super-resolution reconstruction algorithm has become a research hotspot.Super-resolution reconstruction technology refers to obtaining higher quality images by adopting image processing without changing the imaging method,this technology improves the image resolution from the algorithm level without changing the existing hardware conditions,it has been applied in many fields and has great development prospects.This paper mainly studies image super-resolution reconstruction algorithms based on sparse learning and deep learning.The main research contents and contributions are as follows:1.Firstly,this paper analyzes the principle of super-resolution reconstruction technology,application field,mainstream super-resolution reconstruction technology including super-reconstruction problems,and discusses the evaluation criteria of reconstructed images.2.Then,we analyze the image super-resolution reconstruction based on sparse learning,and propose a dictionary algorithm based on image texture information for the problem that the results of single redundant dictionary are not clear enough and the reconstruction time is long.This method constructs feature vectors based on texture information for classifying image blocks.We divide the image blocks of the same structure into the same class and perform individual training to make the obtained sub-dictionaries more specific.The experimental results show that the improved dictionary reconstruction is better and the reconstruction time is shorter.3.Finally,we introduce the concept of deep learning and the image super-resolution reconstruction algorithm based on deep learning.It is difficult to meet the requirements of shallow network reconstruction.We use residual learning combined with dilated convolution and standard convolution to obtain a residual mixed convolution module,which is to improve the image reconstruction effect by building a new network structure.This paper also uses data enhancement to solve the multi-scale problems that exist in image training.