Teoria e aplicação de support vector machines à aprendizagem e reconhecimento de objetos baseado na aparência.

Abstract

Support Vector Machines (SVM) is a machine learning technique derived from two solid backgrounds: Statistical Learning Theory and Mathematical Optimisation. SVM has recently been applied with success to a variety of problems, ranging from character recognition to appearance based object recognition. Some of the reasons for this success are related to the fact this technique exhibits good generalisation performance on many real-life data sets, is well-founded theoreticaly, the training process eliminates the possibility of local minima, there are few free paramets to adjust and the architecture does not have to be found by experimentation. However, since this is a relatively new approach, text books and papers are usually in a language that is not easily acessible to Computer Scientists. Therefore one of the objectives of this dissertation is to provide an introduction to SVM that presents the essential concepts and theory behind the technique and that is more didatic. Appearance-based object recognition strategies appear to be well-suited for the solution of recognition problems in which geometric models of the viewed objects can be difficult to obtain, although they are not naturally tolerant to occlusions. Some machine learning techniques have been applied in this problem like, Principal Component Analysis (PCA), Probably Approximately Correct (PAC) and Neural Networks, but none posed as promising as SVM. Within this context, this dissertation aims to investigate the application of SVM to appearance-based object recognition. It presents practical results of classification initially using a small dataset and then exploring the full power of the technique on a relatively large dataset. It also presents experimental results using different variations of the technique and compares the recognition performance of SVM with the performance of Multilayer Percep tron Backpropagation Neural Networks.