Generalization of raster images containing patterns featuring stochastic repetitiveness
Artur Rataj · Machine Graphics & Vision International Journal archive · 2007
The issue of reconstruction of missing or unreliable parts of an image is one of the basic problems in image processing. For example, there are a number of methods for texture generation on the basis of a small sample. This paper presents a method that 'bottlenecks' an image processing feedforward neural network so that only some basic traits of the image are preserved. These basic traits are in turn used to generalize the image, thus filtering out any unusual parts of the image. The ability of neural networks and several other learning machines to generalize is based on the premise of smoothness of the generalizing function. Thus, in order to detect advanced patterns that exhibit complex traits like repetitiveness, instead of training these machines directly with raw data, transforms of the patterns like the Fast Fourier Transform are sometimes performed. In this paper it is shown, that a simple feedforward neural network, without any pre-processing of the training data, using the described 'bottleneck' architecture, can properly predict a stochastically repetitive pattern in a raster image.