Cómputo evolutivo como enfoque en la descripción del contenido de la imagen aplicado al reconocimiento de objetos

Abstract

The ability to analyse and describe the image content from real or designed scenarios is a challenge and an attractive task for computer vision. This work presents three evolutionary algorithms which analyse the image content for three particular high level tasks such as: image segmentation, object recognition and image mosaicing. On the one hand, the image content is analysed using statistical descriptors in a Gray Level Co-ocurrence Matrix (GLCM) in order to achieve good image segmentations. On the other hand, new local descriptor operators are proposed using genetic programming. These operators describe the image content in order to recognize objects localized within indoor and outdoor scenarios presenting different image transformations. First, we present our EvoSeg algorithm, which uses knowledge derived from texture analysis to identify how many homogeneous regions exist in the scene without a priori information. EvoSeg uses texture features derived from the GLMC and optimizes a fitness measure, based on the minimum variance criteria, using a hierarchical GA. Later, we include interaction within the EvoSeg optimization process obtaining a new algorithm named I-EvoSeg. This algorithm complements the chosen texture information with direct human interaction in the evolutionary optimization process. Interactive evolution helps to improve results by allowing the algorithm to adapt using the new external information based on user evaluation. On the other hand, we describe a genetic programming methodology using a canonical and multi-objective approach, that synthesizes mathematical expressions that are used to improve a well known local descriptor algorithm named SIFT (Scale Invariant Feature Transform). It follows the idea that object recognition in the cerebral cortex of primates makes use of features of intermediate complexity that are largely invariant to change in scale, location, and illumination. These local features have been previously designed by human experts using traditional representations that have a clear, preferably mathematically, well-founded definition. However, it is not clear that these same representations are implemented by the natural system with the same representation. Hence, the possibility to design novel operators through genetic programming represents an open research avenue where the combinatorial search of evolutionary algorithms can largely exceed the ability of human experts. Hence, we provide evidence that genetic programming is able to design new features that enhance the overall performance of the best available local descriptor. Experimental results confirm the validity of the proposed approach using a widely accept testbed for evaluating local descriptors. Finally, we used our evolved descriptors in object recognition application and image mosaicing.