Convolutional Analysis Operator Learning by End-To-End Training of Iterative Neural Networks

Research article accepted for the Proceedings of the International Symposium on Biomedical Imaging (ISBI) 2022

Andreas Kofler (PTB) published together with Christian Wald (Charité–Universitätsmedizin Berlin) Markus Haltmeier (University of Innsbruck), Tobias Schäffter (PTB) and Christoph Kolbitsch (PTB) an article about a method for image reconstruction.

The article describes a new method for obtaining convolutional sparsifying filters which can be used for regularization of image reconstruction problems. The method involves the construction of a neural network which corresponds to an alternating algorithm of finite length for reconstructing images using a set of sparsifying filters. The filters can then be learned in a supervised manner by end-to-end training of the network. The study demonstrates that for a certain class of problems, neural networks can also be used as physics-aware training algorithms for classical learning-based methods yielding interpretable regularization methods

The article is available at …. . An implementation of the code can be found at https://github.com/koflera/ConvSparsityNNs.