Model-based learning of co-sparse representations for image processing applications
Martin Kiechle · 2019
In this thesis, approaches for unsupervised learning of image representations are developed with the goal of reducing training complexity over fully learned and improving accuracy over expert-crafted models. To that end, application-specific knowledge of image formation is incorporated into sparsity-based, unsupervised learning models for low-level image representations. The designed numerical algorithms achieve state-of-the-art results in uni- and multi-modal image reconstruction, alignment, and segmentation benchmarks.