标题：Registro automático, preciso e robusto de imagens com RANSAC (Random Sample Consensus) adaptado para o descritor SIFT (Scale Invariant Feature Transform)
出版社：Grupo de Pesquisa Metodologias em Ensino e Aprendizagem em Ciências
摘要：Image registration is a common problem in computer vision with several applications which consists of finding the correct transformation between pairs of overlapping images. This work aims to present an automatic and accurate model for image registration using the SIFT descriptor and the adapted RANSAC estimation method. The registration occurs through the estimation of the homography between the pairs of images, which use the point correspondences given by the SIFT descriptor. Matches are classified using a dynamically estimated error threshold. The analysis considers the dispersion of the residual over various error thresholds and adopts the one that minimizes the dispersion and the magnitude of the error. The model is tested with 8 heterogeneous pairs of images divided into two groups: 4 pairs obtained with a professional camera and 4 pairs obtained with a common camera. Due to the high quality of the images in the first group, few iterations of the model are necessary for a good estimate of the correct homography. In the second group, the model showed that it is capable of building mosaics between pairs of images with an overlap of less than 20%, finding exact correspondences between pairs of images regardless of the acquisition method. Furthermore, it was able to handle up to 65% corruption between matches, with a total execution time of a few seconds.
