Image Super-resolution Based On Structural Feature Learning Dictionary

Image super-resolution uses one or more low-resolution images to reconstruct high-resolution images with richer details.This technology plays an important role in subsequent image recognition,analysis,and tracking,and has been widely used in satellite remote sensing,public security,medical imaging,and pattern recognition and other fields.Therefore,it is of great theoretical significance and application value to study image super-resolution deeply.At present,the commonly used image super-resolution methods include interpolation-based,reconstruction-based and learning-based methods.Among the learning-based methods,the methods based on sparse representation has become a hot research direction for solving the problem of super-resolution and the key lies in the construction of learning dictionary.In recent years,the nonlocal self-similarity,as a priori information,has been successfully applied to image deblurring,image denoising,and image restoration.Based on that,this paper exploits the image intrinsic nonlocal self-similarity,and the direction edge characteristics,and focuses on constructing learning dictionary with strong expression and robustness,and applies it to the image super-resolution,and achieved good results.This paper mainly includes the following three aspects:1.It learns and studies several commonly used image super-resolution reconstruction methods: interpolation-based,reconstruction-based,and learning-based methods,and implements and compares some of the algorithms.2.This paper proposes direction edge learning dictionary and apply it to image super-resolution.This method firstly designs a pair of direction edge templates to cluster image patches;Then the K-SVD algorithm is used to train the dictionary for each type of image patches,and two pairs of direction edge dictionary are obtained;Finally,sparse coding and the direction edge dictionary are combined to realize image super-resolution.Experimental results prove that this method can recover better edge structure and detailed information.3.This paper proposes new structural dissimilarity learning dictionary and apply it to realize image super-resolution.The constraints of this method is inherent nonlocal self-similarity of the image.First,it uses three different similarity judgment methods to delete similar patches of the training set,and gets a small set of dissimilar samples;And then the small set is trained to obtain a pair of structural dissimilarity learning dictionary;Finally,the image super-resolution reconstruction is realized by using sparse coding.Experiments results demonstrate that,under the premise of ensuring quality of image reconstruction,this method can greatly reduce the number of training samples and improve the efficiency of learning dictionary construction.In a word,focusing on improving the expression ability and construction efficiency of learning dictionary,this paper proposes direction edge learning dictionary and the structural dissimilarity learning dictionary,and apply them to realize image super-resolution successfully and finally obtains better super-resolution results.