Optimally Spaced Autoencoder
Journal
International Conference on ICT Convergence
ISBN
9798350313277
Date Issued
2023-01-01
Author(s)
Liou, Daw Chih
Abstract
This paper presents a method for image restoration that uses a new object function for a multilayer perceptron (MLP) network. The training algorithm of the MLP aims to maximize the separation between patterns from different classes, while minimizing the distances between patterns from the same class. The trained MLP serves as a transformation encoder, mapping the pattern space into a new space where patterns from different classes are distinctly separated. This encoder accomplishes the 'optimally spaced coding' directly. It will enable effective resolution of challenging classification problems and restoration of severely corrupted images.
Subjects
character recognition | classification | image restoration | optimally spaced codes | pattern recognition | vision
SDGs
Type
conference paper