"DISEÑO DE FILTROS DIGITALES PARA EL RECONOCIMIENTO Y LA MULTI-CLASIFICACIÓN DE OBJETOS EN ESCENAS REALES"

Abstract

Abstract Nowadays, the need for object recognition techniques has became increasingly evident in various important applications such as security, automatic industrial inspection, and medical diagnostics, among others. Feature-based object recognition systems are perhaps the most popular, however, these systems may be computationally intensive and, moreover, its overall performance depends on the some ad-hoc decisions made by the designer. Object recognition by correlation ltering represents important alternative to feature-based systems because they have strong mathematical basis and can be e ciently implemented in hybrid opto-digital correlators or in digital processing systems at high speed. The main problem of object recognition is the identi cation, localization, and classi cation of several objects of interest from an input observed scene. This problem can be solved by using a bank of correlation lters which requires at least one lter per each di erent class of objects. However, under this approach the computational complexity of the resultant pattern recognition system can be increased considerably as the number of classes of objects increases. In this thesis, we propose the design of new adaptive composite lters with complex constraints for multi-class pattern recognition in real scenes. The hypothesis consists in proving that it is possible to recognize and classify multiple objects with a single correlation operation, using the intensity and phase distributions at the output complex correlation plane at the lter output, the same stage. n this manner, the computational complexity of the pattern recognition system can be reduced considerably. The proposed adaptive lters are constructed with the help of an iterative algorithm optimizing performance metrics. The impact of the proposal is analyzed in terms of tolerance to additive noise, computational complexity, and recognition and classi cation e ciency, compared to conventional correlation lters.