Aprendizado não supervisionado de métricas via redução de dimensionalidade para reconhecimento de dígitos numéricos

Abstract

Character recognition is a task that has occurred since the last century and is still a problem to be solved nowadays. One of the challenges is that the characters to be classified are images, therefore, they have many dimensions (each pixel is a dimension in a gray image). In this paper were used two different databases that contain handmade characters. Working with a database with many dimensions is a challenge, due to issues like the curse of dimensionality, and to try to solve this complication, dimensionality reduction is usually an attempt to mitigate the problem. In this work were used several supervised classification algorithms and were used non-linear dimensionality reduction algorithms: ISOMAP, LLE and Laplacian Eigenmaps with the purpose of learning a metric (a distance function) better suited than the Euclidean distance among the input images. Finally, the Wilcoxon test was used to compare statistically the dimensionality reduction algorithms.