Self-learning: Faster, Smarter, Simpler

Abstract

Self-learning, introduced in the previous chapter, uses only one image to achieve super-resolution with high magnification factors. But, as the image resolution increases, the number of patches in the dictionary also increases dramatically, and makes the SR reconstruction computationally prohibitive. It employs l₁-minimization to exploit the sparsity inherent within natural images, which further compounds the time-complexity problems associated with the technique. One way to mitigate this problem is to reduce the dictionary size in a meaningful way by employing some kind of learning algorithms (e.g., K-SVD, K-means). However, such learning algorithms come with their own computational burden, and may ultimately prove to be of little advantage when time is of the essence. Another way to find the best-mapped patch for a particular LR patch is to restrict the search space within a neighborhood of the LR patch in the coarser-resolution image. But this may not always result in the best-mapped patch for that particular LR patch.