Hierarchical sparse representation with adaptive dictionaries for image super-resolution

This paper presents an image hierarchical super-resolution (SR) method with adaptive dictionaries, based on signal sparse representation. It can not only improve image detail quality but also reduce computational cost. Research on the human visual system suggests that our eyes are mainly sensitive to high-frequency contents. Inspired by this observation, we implemented a hierarchical process where an image was decomposed into a detail layer and a base layer. The detail layer is reconstructed through an over-complete dictionary while the base layer is interpolated by bi-cubic. Through these, we can keep the HR details better. Next is how to accelerate while keeping good quality. In our method, adaptive dictionaries are trained by feature clustering. Firstly, we train low dimension sub-dictionaries to reduce time complexity. Secondly, then we apply overlapping feature clustering to the training. Thus dictionaries can be adaptive and more complete. All these can also prevent sub-dictionaries with over strong independence but less compatibility. Besides, initializing the sparse coefficients also plays an important role in our acceleration. Experimental results validate that ours are competitive or even superior in quality than those produced by other methods and our test data indicates substantial reduction in processing time over other similar SR methods.