The 16th Meeting on Image Recognition and Understanding High-frequency Restoration using Deep Belief Nets for Super-resolution

Super-resolution techniques are generally divided into two approaches: example-based methods and statistical methods. Example-based methods [1] simply use (or select in sparce coding [2]) pairs of low-resolution and high-resolution patches for the reconstruction. In this approach, a lowresolved input image is decomposed into patches, each of which is compared with the patches in the database and replaced with the corresponding high-resolved patch. Although this approach produces relatively less-deteriorated images, it is not based on any statistical models and lacks versatility. For the statiscical approach, various methods have been proposed so far: the eigen-space BPLP [3], the MRF-based approach [4], a GMM-based approach [5], and so on. Some of these statistical approaches rely on the training of the correspondence relationships between lowresolved/high-resolved images. Therefore, if one wants to enlarge an image with the desired scale, the relationships between the low and high resolution with that scale need to be trained beforehand. Meanwhile, Hinton et el. introduced an effective training algorithm of Deep Belief Nets (DBNs) in 2006 [6], and the use of DBNs rapidly spread in the field of signal processing with great success. DBNs are probabilistic generative models that are composed of multiple layers of stochastic latent variables, and have a greedy layer-wise unsupervised learning algorithm. DBNs are not only used for classification tasks, but also for the completion of an image or for collaborative filtering. Eslami et el. adopted a type of DBNs (called ShapeBM) to complete the missing region in an image [7]. Salakhutdinov et el. used 2-layer DBNs (i.e., Restricted Boltzmann Machines; RBMs) for collaborative filtering [8], which has the benefit of the DBNs dealing with missing data. In this paper, we propose a novel super-resolution method using DBNs to restore the missing high-frequencies (Fig. 1), motivated by the above-mentioned characteristics of DBNs. In our approach, a low-resolved image is first scaled up to the prescribed size by using bicubic interpolation, and