Neural Network Image Scaling Using Spatial Errors
Carl Staelin, Darryl Greig, Mani Fischer, Ron Maurer · 2003
We propose a general method for gradient-based training of neural network (NN) models to scale multidimensional signal data. In the case of image data, the goal is to fit models that produce images of high perceptual quality, as opposed to simply a high peak signal to noise ratio (PSNR). There have been a number of perceptual image error measures proposed in the literature, the majority of which consider the behavior of the error surface in some local neighborhood of each pixel. By integrating such error measures into the NN learning framework, we may fit models that minimize the perceptual error, producing results that are more visually pleasing. We introduce a spatial error measure and discuss in detail the derivative computations necessary for backpropagation. The results are compared to neural networks trained with the standard sum of squared errors (SSE) function, as well as a state of the art scaling method.