Optimización de la Segmentación Local de Sauvola Aplicada a la Detección de Defectos Superficiales en Escenas con Iluminación No Homogénea

AbstractThe presence of non-homogeneous illumination in real scenesimages is an actual problem that difficult the correct segmentationof these. This paper presents a methodology for optimizing Sauvolalocal segmentation for the detection of superficial defects in non-homogeneous illuminated images by adjusting its parametersthrough genetic algorithms. The methodology consists of thesestages: First, the problem is proposed from the perspective ofgenetic algorithms where each individual in the populationrepresents the values for Sauvola's parameters. Then severalfitness functions are proposed using comparison metrics between aSauvola's segmentation and one performed manually. Eachfunction is evaluated by running the genetic algorithm with it in asubset of images. The best fitness function, according to the resultsof optimization, is used again in a larger sample. Finally, the lastoptimization results are analyzed by a clustering analysis. Theresults show that it is possible to adjust Sauvola's parameters tosuccessfully segment each image but these do not exhibit atendency to a specific point that allow to suggest uniqueparameters to segment all images with a high performance.