{"id":"arxiv/2602.21707","url":"https://arxiv.org/abs/2602.21707","tags":["eess.IV","cs.CV","cs.LG","math.OC"],"score":0,"title":"Learning spatially adaptive sparsity level maps for arbitrary convolutional dictionaries","domain":"eess.IV","source":"arxiv","authors":["Joshua Schulz","David Schote","Christoph Kolbitsch","Kostas Papafitsoros","Andreas Kofler"],"reviews":0,"subject":"eess.IV","summary":"State-of-the-art learned reconstruction methods often rely on black-box modules that, despite their strong performance, raise questions about their interpretability and robustness. Here, we build on a recently proposed image reconstruction method, which is based on embedding data-driven information into a model-based convolutional dictionary regularization via neural network-inferred spatially adaptive sparsity level maps. By means of improved network design and dedicated training strategies, we extend the method to achieve filter-permutation invariance as well as the possibility to change…","category":"research","comments":0,"disputed":false,"division":"eess","mentions":0,"discussed":0,"published":"2026-02-25","reactions":{}}