Dictionary optimization for representing sparse signals using Rank-One Atom Decomposition (ROAD)

Abstract

Dictionary learning has attracted growing research interest during recent years. As it is a bilinear inverse problem, one typical way to address this problem is to iteratively alternate between two stages: sparse coding and dictionary update. The general principle of the alternating approach is to fix one variable and optimize the other one. Unfortunately, for the alternating method, an ill-conditioned dictionary in the training process may not only introduce numerical instability but also trap the overall training process towards a singular point. Moreover, it leads to difficulty in analyzing its convergence, and few dictionary learning algorithms have been proved to have global convergence. For the other bilinear inverse problems, such as short-and-sparse deconvolution (SaSD) and convolutional dictionary learning (CDL), the alternating method is still a popular choice. As these bilinear inverse problems are also ill-posed and complicated, they are tricky to handle. Additional inner iterative methods are usually required for both of the updating stages, which aggravates the difficulty of analyzing the convergence of the whole learning process. It is also challenging to determine the number of iterations for each stage, as over-tuning any stage will trap the whole process into a local minimum that is far from the ground truth. To mitigate the issues resulting from the alternating method, this thesis proposes a novel algorithm termed rank-one atom decomposition (ROAD), which intends to recast a bilinear inverse problem into an optimization problem with respect to a single variable, that is, a set of rank-one matrices. Therefore, the resulting algorithm is one stage, which minimizes the sparsity of the coefficients while keeping the data consistency constraint throughout the whole learning process. Inspired by recent advances in applying the alternating direction method of multipliers (ADMM) to nonconvex nonsmooth problems, an ADMM solver is adopted to address ROAD problems, and a lower bound of the penalty parameter is derived to guarantee a convergence in the augmented Lagrangian despite nonconvexity of the optimization formulation. Compared to two-stage dictionary learning methods, ROAD simplifies the learning process, eases the difficulty of analyzing convergence, and avoids the singular point issue. From a practical point of view, ROAD reduces the number of tuning parameters required in other benchmark algorithms. Numerical tests reveal that ROAD outperforms other benchmark algorithms in both synthetic data tests and single image super-resolution applications. In addition to dictionary learning, the ROAD formulation can also be extended to solve the SaSD and CDL problems. ROAD can still be employed to recast these problems into a one-variable optimization problem. Numerical tests illustrate that ROAD has better performance in estimating convolutional kernels compared to the latest SaSD and CDL algorithms.