Designing Constrained Projections for Compressed Sensing: Mean Errors and Anomalies with Coherence

Dhruv Shah ([email protected]), Alankar Kotwal and Ajit Rajwade

[Project Page] [Supplemental Material]

---

This repository contains the authors' implementation for the paper "Designing Constrained Projections for Compressed Sensing: Mean Errors and Anomalies with Coherence" submitted to IEEE Global Conference on Signal and Information Processing 2018.

• coherence-opt/: Our implementation of average coherence-based projection design, described in section 3 of the paper.
• datasets/: Test natural images drawn from the Berkeley Segmentation Data Set (BSDS500) and the INRIA Holidays Data Set for testing the algorithms and generating results. These images were not a part of the training data.
• designed-matrices/: Sample matrices designed using the various algorithms discussed in the paper provided for use.
• gmm-train/: Unoptimized implementation sourced from MATLAB File Exchange, courtesy Mo Chen (downloaded 2018-01-19). A sample GMM trained on natural image patches from BSDS500 can be found as gmm-train/results/trained_model_25.mat.
• misc/: Miscellaneous files useful for reconstruction and file handling. This includes original implementations of l1-magic, SPGL1, our implementation of the piecewise-linear decoder and other scripts.
• mmse-opt/: Our implementation of the MMSE-based projection design algorithm proposed in section 4.3 of the paper.
• results/: Compare different reconstruction methods (and matrices) and visualize results. Results from the paper can be replicated using the scripts provided.