Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
10360339 | Pattern Recognition | 2014 | 12 Pages |
Abstract
In the context of unsupervised segmentation of noisy images, a Minimum Description Length (MDL) polygonal active contour technique based on nonparametric modeling of the noise probability density function (pdf) has been proposed in 2011. This approach allows fast and efficient segmentation of an object without a priori knowledge on the intensity fluctuations. Nevertheless, since the object and the background are assumed to be homogeneous, degraded segmentation results are obtained when images present inhomogeneous intensity variations. It is shown in this paper that this constraint of homogeneity can be removed, still with minimizing a MDL criterion without undetermined parameters and adapted to nonparametric modeling of the noise pdf. For that purpose, the spatial inhomogeneity of the intensity is modeled with 2D quadratic functions. Moreover, low computation times can be achieved (approximately 60Â ms on 256Ã256 pixel images) using a two-step optimization strategy. The efficiency and the robustness of this approach are then validated on various synthetic and real images acquired from different sensors.
Related Topics
Physical Sciences and Engineering
Computer Science
Computer Vision and Pattern Recognition
Authors
Siwei Liu, Frédéric Galland, Nicolas Bertaux,