Aprendizaje automático a partir de formas = Machine Learning from Shapes

Abstract

In this work the problem of learning from images to perform grouping and classification of shapes is addressed. The key idea of the shape recognition approach is to encode the instances available for learning in the form of directional data, that will be used to characterize those instances and perform comparisons among them. The objects to study are thus 2 and 3 dimensional shapes, that will be characterized by the distribution of the direction of the normal vectors to the tangent hyperplanes at the boundary of the shape. In two dimensions, this boundary is a contour, and these directional data will in fact form a curve, that manipulated as functional data can be used to encode the shapes in two discrete representations: a normalized histogram and a kernel density estimation for the probability function. These representations are used to extract characteristics based on metrics defined in the space of circular distributions, categorize the encoded shapes and finally compare them. These characterization and comparison techniques will be later embedded in some clustering and classification algorithms, applying them in a simple shape recognition problem and a real world problem of clustering and classification with fish otolith shapes.