Rede Neural por Convolução
para Reconstrução Estéreo

Júlio M. Otuyama


Abstract

In order to model neural networks for applications of translation invariant problems, weight sharing restrictions are used. Neural networks with such restrictions are called Shared Weight Neural Networks. Weight sharing corresponds to the convolution operation, allowing for the use of several techniques that accelerate its calculation. This equivalence is the reason why such neural networks are also known as Convolution Neural Networks. This project implements the Convolution Neural Network model with an adequate language that allows new topologies to be created in a flexible, simple and fast way. Allegedly, the reconstruction of three-dimensional information occurs in the first stages of the visual cortex, based on various aspects, among which are the stereoscopic vision, shape from shading, object overlapping, texture, perspective and movement. This project uses the Convolution Neural Network approach for the purpose of stereo reconstruction (stereoscopic vision), using all specic characteristics of the problem to reach an adequate topology. A second line of this project presents an extension of the Convolution Neural Network model, using specic characteristics of the convolution calculation - the Fourier transform. Such extension of the Convolution Neural Network model represents weights on the frequency domain.

Keywords: Neural Networks, Translation Invariance, Shared Weights;

Download: PDF, source code;

Online information at UFSC


Back to home page