Image Super-resolution Technology Research

Image super resolution technology is to recover the original high resolution image from low resolution image.However,super resolution problem is extremely ill-posed due to the complicated image degradation model,therefore can not be directly solved.Since the learning-based methods have outstanding advantages in solving ill-posed problem,we will mainly do research on it and make further analysis and improvements.First,this paper proposes an image super resolution method using sparse representation based on structure-texture decomposition.For structure,we proposes a self-driven dictionary learning method based on self-similarity and for texture,we directly use the external dictionary for reconstruction.This framework can efficiently reduce the complexity of image pattern,as well as the demand for dictionary size and training samples and indirectly improve the representation ability of dictionary.Thus,our method can provide better super resolution performance and can handle more complex images.Second,this paper does research on manifold-regularized sparse support method and proposes the modified manifold-regularized collaboration support method,which uses the anchored neighborhood strategy to construct the objective function and uses collaboration representation coefficients instead of sparse representation coefficients to construct the support samples.This method can make the preserved structure of manifold space smoother and make the projection between high and low resolution patch space more accurate.Part of the calculation process of collaboration representation coefficients is transferred to offline,therefore can improve the super resolution performance while reducing the running time.Finally,this paper studies on anchored neighborhood regression based method and proposes the enhanced version,which uses K-means clustering analysis instead of KSVD to obtain the dictionary,therefore maintains the consistency between dictionary training and online atom searching.Moreover,we use absolute correlation as the measure in neighborhood searching process to further increase the density of neighborhood.Thus,our method can improve both the performance and efficiency of super resolution reconstruction.Experimental results show that the proposed image super resolution algorithms can provide sharper edges,finer details,less artifacts and faster running speed,which has great application value.