Neural Network Image Scaling Using Spatial Errors

Abstract—We propose a general method for gradient-basedtraining of neural network (NN) models to scale multi-dimensional signal data. In the case of image data, the goal is to fitmodels that produce images of high perceptual quality, as opposedto simply a high peak signal to noise ratio (PSNR). There havebeen a number of perceptual image error measures proposed inthe literature, the majority of which consider the behavior ofthe error surface in some local neighborhood of each pixel. Byintegrating such error measures into the NN learning framework,we may fit models that minimize the perceptual error, producingresults that are more visually pleasing. We introduce a spatialerror measure and discuss in detail the derivative computationsnecessary for backpropagation. The results are compared toneural networks trained with the standard sum of squared errors(SSE) function, as well as a state of the art scaling method.Index Terms—image error measures, image scaling, imageinterpolation, super-resolution, neural networks.