A First-Order Primal-Dual Algorithm for Convex Problems with Applications to Imaging

Abstract

In this paper we study a first-order primal-dual algorithm for convex optimization problems with known saddle-point structure. We prove convergence to a saddle-point with rate O(1/N) in finite dimensions, which is optimal for the complete class of non-smooth problems we are considering in this paper. We further show accelerations of the proposed algorithm to yield optimal rates on easier problems. In particular we show that we can achieve O(1/N 2) convergence on problems, where the primal or the dual objective is uniformly convex, and we can show linear convergence, i.e. O(1/e N) on problems where both are uniformly convex. The wide applicability of the proposed algorithm is demonstrated on several imaging problems such as image denoising, image deconvolution, image inpainting, motion estimation and image segmentation.

GASPACHO: a generic automatic solver using proximal algorithms for convex huge optimization problems

“…For example, for image deblurring where the blur is assumed to be a result of a circular convolution using L 2 regularization of the unknown image, the estimation of the noise-free image is best done in the discrete Fourier domain: the least squares formulation then leads to a point-wise Wiener filter. Alternatively, many sparse solvers (e.g., based on ADMM, 12 augmented Lagrangrian, 13 primal-dual techniques 14 ) use splitting variables to split the problem into a set of subproblems that are more easy to solve. Then, depending on the structure of the involved matrices and the cost function, a particular solver can be used for each subproblem.…”

Building Scene Models by Completing and Hallucinating Depth and Semantics

“…The convergence [42] of this primal-dual procedure for a convex problem depends on the parameter values τ and σ, which must satisfy τ σ A 2 ≤ 1. For non-convex functional, [37] shows the algorithm generates a bounded solution with empirically convergence.…”

Epigraphical Projection for Solving Least Squares Anscombe Transformed Constrained Optimization Problems

“…Now one can choose a primal-dual splitting algorithm as those proposed in [4,6,11,12,23] to solve this problem. One step in all these algorithms consists of the orthogonal projections onto the epigraphs of ϕ i for all i ∈ {1, .…”

“…Let (max{x, 0}, ζ) ∈ epi ϕ and t 0 := 2 max{x, 0} − z. Let the polynomial p be defined by (6). Then the Newton method for finding a zero of p with initial value t 0 converges (after a finite number of steps) monotonically to the root t + if 4x ≥ z 2 , resp., t − if 4x < z 2 .…”