Multi-Label Segmentation via Residual-Driven Adaptive Regularization
This work addresses segmentation challenges in computer vision, offering an incremental improvement over existing methods for researchers and practitioners in image processing.
The paper tackles multi-label image segmentation by introducing a novel adaptive regularization scheme that adjusts based on spatial statistics of the residual, minimizing bias at convergence, and achieves improved segmentation accuracy on synthetic and real images with concrete performance gains reported.
We present a variational multi-label segmentation algorithm based on a robust Huber loss for both the data and the regularizer, minimized within a convex optimization framework. We introduce a novel constraint on the common areas, to bias the solution towards mutually exclusive regions. We also propose a regularization scheme that is adapted to the spatial statistics of the residual at each iteration, resulting in a varying degree of regularization being applied as the algorithm proceeds: the effect of the regularizer is strongest at initialization, and wanes as the solution increasingly fits the data. This minimizes the bias induced by the regularizer at convergence. We design an efficient convex optimization algorithm based on the alternating direction method of multipliers using the equivalent relation between the Huber function and the proximal operator of the one-norm. We empirically validate our proposed algorithm on synthetic and real images and offer an information-theoretic derivation of the cost-function that highlights the modeling choices made.