Core Concepts
Leveraging the Sliding Frank-Wolfe algorithm for efficient line recovery in degraded images.
Abstract
The article introduces a novel approach using the Sliding Frank-Wolfe algorithm to address line recovery challenges in degraded images. It focuses on two models: blurred line deconvolution and ridge detection of linear chirps in spectrogram images. By optimizing over the space of measures, the authors aim to improve parameter recovery precision beyond the Rayleigh limit. The paper discusses atomic norm minimization, convex optimization problems, and the use of conditional gradient methods for sparse inverse problems. The proposed Alternating Conditional Gradient Method (ACGM) enhances line estimation by incorporating non-convex search steps. Experimental results demonstrate improved accuracy in line detection and parameter estimation compared to previous methods.
Stats
N = 65 for Gaussian Lines experiment.
K = 3 lines with amplitudes equal to 1 in Exp. 1.
σ1 = σ2 = 1 for blurred lines in Exp. 1.
N = 256 for Chirp Lines experiment.
K = 2 chirps with equal amplitudes α1 = α2 = 1.
Quotes
"The Sliding Frank–Wolfe algorithm provides a very good quality of line estimation even in the presence of strong degradations."
"By working directly in the parameter space, we no longer need to operate in two steps: denoising and deconvolution via the atomic norm."
"The method consistently outperforms previous approaches in terms of accuracy for Gaussian Lines detection."