Viterbi Faculty of Electrical Engineering, Technion
A Local Block Coordinate Descent Algorithm for Convolutional Sparse Coding
The Convolutional Sparse Coding (CSC) model has recently gained considerable traction in the signal and image processing communities. By providing a global, yet tractable, model that operates on the whole image, the CSC was shown to overcome several limitations of the patch-based sparse model while achieving superior performance in various applications. Contemporary methods for learning the CSC dictionary often rely on the Alternating Direction Methods of Multipliers (ADMM) in the Fourier domain for the computational convenience of convolutions, while ignoring the local characterizations of the image. A recent work by Papyan et al. suggested a different approach, named Slice Based Dictionary Learning (SBDL), for solving the convolutional sparse pursuit and training the convolutional filters. The SBDL algorithm operates locally on image patches and demonstrates better performance compared to the Fourier-based methods, albeit still relaying on the ADMM, introducing further auxiliary variables and an extra parameter that needs tuning. In this work we maintain the localized strategy but we propose a new approach based on the Block Coordinate Descent (BCD) algorithm for solving the CSC pursuit without the need for auxiliary variables nor parameter tuning. Furthermore, we introduce a novel Local Stochastic Gradient Descent (LSGD) method for training the convolutional filters. The LSGD leverages the benefits of stochastic optimization on nonconvex problems, while still being applicable to a single training image. We demonstrate the advantages of the proposed algorithm for image inpainting and multi-focus image fusion, achieving state-of-the-art results.
* MSc seminar under supervision of Prof. Ron Kimmel.
On 28/08/2018 at 11:00
in 401, Taub Building