A selective upscaling method via super-resolution validity measure

In this paper, we propose a selective upscaling method which applies either super-resolution or bicubic interpolation according to region characteristics for reducing computational complexity. In order to divide into the two regions, we measure super-resolution validity scores in each local patch. We train Random forests as a regression model for estimating the super-resolution validity score by using three feature vectors, such as gradient magnitude, image residual, and weighted sum of DCT coefficients. Experimental results show that the selective upscaling method reduces processing time while maintaining quantitative performance in comparison with state-of-the-art SR methods.