A modular neural network applied to image transformation and mental images

Abstract

This paper presents an original modular neural network architecture whose modules are multilayer perceptrons. The modules' inputs are external inputs or hidden layers of other modules, thereby allowing them to be connected in a general manner. Based on this flexible architecture, networks with high numbers of inputs and outputs can be elaborated and properly trained. A suitable application is image transformation, i.e. the transformation of many input pixels into as many output pixels. Some architectural variations are presented; first the localization over a fraction of the network of a specific transformation's training, and then the merging of two input images into a single output image. As a case study, we use the modular network to model mental images and their transformations (mental rotation, mental assemblage). It should eventually prove to be a valuable tool for image processing applications, such as super-resolution, or for the elaboration of complex cognitive systems.