Multi-Class SVM Learning using Adaptive Code

Classification of objects in computer vision is done mostly without any knowledge about multiple class memberships. In this master thesis several learning algorithms based on Support Vector Machines and similar approaches regarding multiple class memberships are explored. Recognition performance and robustness of the algorithms are tested with small quantities of training objects, making all learning difficult for standard and new learning algorithms. New algorithms are real multiple class membership algorithms based on a innovative class separation, but also some one-vs-rest SVM based algorithms. Most approaches improve slightly the performance compared to classical SVM, but no important changes can be found. SVM-inlärning av multipla klasser med adaptiv kodning