INSTITUTO DE INVESTIGACIONES EN ENERGIA NO CONVENCIONAL
Unidad Ejecutora - UE
congresos y reuniones científicas
An invariant descriptor for pattern classification
G. G. ROMERO; A. C. MONALDI; D. A. VITULLI; A. V. BLANC
Conferencia; IX Conference RIAO / OPTILAS 2016; 2016
Object recognition, irrespective of orientation, size or position in an image is an ability that humans take for granted. However, for a computer, an object that has been moved, scaled or rotated in an image represents a completely different object. For a computer to recognize two objects as the same or to classify different objects special algorithms have to be developed, whose response is robust to changes in scale, rotation or translation. Fourier-Mellin Transform (FMT) emerges as an alternative for the design of an invariant fingerprint, since its magnitude is invariant under rotations and scale changes. This work aims to perform classification of objects immersed in binary images, which are interpreted as a superposition of weighted rectangle functions. The main advantage of this assumption is that allows the analytical evaluation of the FMT. The designed invariant fingerprint encompasses a complete description of the object. Nevertheless, for pattern classification it suffices a partial description of the object by means of a set of descriptors obtained as a special case of the invariant fingerprint. These descriptors have been tested on the classification of alphabet character of different sizes and orientation placed in random positions within the image. Results show a good performance of the system on character classification and encourage to incorporate a larger set of descriptors emerged from the same invariant fingerprint improving the object description for automatic recognition of complex objects.