Leveling cartoons, texture energy markers, and image decomposition.
Petros Maragos, Georgios Evangelopoulos · Biblioteca Digital da Memória Científica do INPE (National Institute for Space Research) · 2007
Keywords: The variational u + v model for image decomposition aims at separating the image into a ‘cartoon component ’ u, which consists of relatively flat plateaus for the object regions surrounded by abrupt edges, and a ‘texture component ’ v, which contains smaller-scale oscillations plus possibly noise. Exploiting this model leads to improved performance in several image analysis and computer vision problems. In this paper we propose alternative approaches for u+v decomposition based on levelings and texture energy. First, we propose an efficient method for obtaining a multiscale cartoon component using hierarchies of levelings based on Gaussian scale-space markers. We show that this corresponds to a constrained minimization driven by PDEs and link the leveling cartoons with total variation minimization. Second, we extract the texture component from levelings of the residuals between the image and its multiscale levelings. Further, we employ instantaneous nonlinear operators to estimate the spatial modulation energy in the most active texture frequency bands and use this as a new type of texture markers that yield an improved texture component from the leveling residuals. Finally, we provide experimental results that demonstrate the efficacy of the proposed image decomposition methods.