Reconhecimento facial aplicado a imagens infravermelhas: Um estudo comparativo utilizando abordagens holísticas.

Abstract

Modern facial recognition systems have such an extraordinary performance when applied to high definition and well illuminated pictures. Nevertheless, these are not the conditions found in real life situations where applications of this kind are needed. In most cases, recognition rates tend to degrade when exposed to changes in pose, lighting and facial expressions, common situations which the system has to deal with, especially in uncontrolled environments. For these reasons, the use of infrared imagery for facial recognition has become one of the most promising approaches, since they tend to be more robust to those problems that affect facial recognition systems. This work introduces the development of an application capable of applying face recognition methods on images of the visible and infrared spectrum, and latter, the system was used in a study comparing both results. For this task, a data set containing multispectral pairs of images from 30 individuals was used. Each set has variations in pose, light and expression, making it possible to measure their impact on recognition rates and verify how robust each kind of image is for multiple scenarios. For recognition, Eigenfaces and Fisherfaces, two well known and studied methods were implemented. For classification, the k-nearest neighbors algorithm was applied, switching between Euclidian, Manhattan, Chebyshev and Mahalanobis distances, and varying the values of k and principal components for dimensionality reduction and image projection. The results confirm the hypothesis that infrared imagery is more robust to illumination changes, preserving recognition rates even in the most unfavorable conditions, which was not the case for visible specter images. For pose and facial expressions, the tests were inconclusive due to the small amount of images available. The research shows that facial recognition field is in need of more complete databases, so more precise tests can be done. Despite of this, the results suggests that image fusion in a multispectral system could be the best option, taking advantage of the benefits from each king of image.