Clustering Oriented Scale Invariant Dictionaries for Single Image Super-Resolution

This paper presents a novel methodology based on single image super-resolution problem by a unique scale invariant image feature. This unique feature is used to classify the image patches into separate unique classes. For that aspect, a pair of joint dictionaries are trained for every class. For each class a separate mapping matrix is also learned. By doing so, we are able to enrich the demonstration capacity for super-resolution. The joint dictionary training and a separate mapping matrix learning helps in making the approximate invariant high resolution and low resolution sparse coefficients further similar. The proposed algorithm is a patch based sparse representation algorithm. Instead of a single dictionary pair for whole image, this paper proposed unique scale invariant class dependent dictionaries in the training. In the testing, each patch is first confirmed to a specific class by utilizing the proposed scale invariant feature, afterward, the dictionary pair and the mapping matrix of that class are selected and utilized for its rebuilding. By using the above proposed methodology, the worth of demonstration is enhanced by getting the scale invariant features properly. The testing results display that the proposed algorithm is more efficient in accomplishing the task of super-resolution.