Project
Dirichlet-field segmentation with uncertainty estimation
Variational and neural segmentation methods that combine Dirichlet uncertainty fields with anisotropic spatial regularization.
This project formulates image segmentation as an optimization problem in which predictive uncertainty is represented by Dirichlet concentration parameters and regularized with structure-aware spatial terms.
The work aims to improve calibration, preserve edges, and retain computational efficiency by avoiding repeated stochastic inference wherever a principled closed-form uncertainty calculation is possible.
Research areas
Dirichlet uncertainty models for imaging
Scale-consistent variational regularization
Participants
Evgeny Yuryevich Shchetinin
Andrey Andreyevich Shevchuk
Related publications
Semantic Segmentation with Uncertainty Estimation Based on the Dirichlet Model and Anisotropic Regularization
A Computational Method for Image Segmentation Based on a Dirichlet Field and an Analysis of the Asymptotic Accuracy of Spatial Regularizer Discretization