Deep Learning-Based Optimal Compressed Sensing
- Type of project: Research Project/Hauptseminar/Thesis (literature review + coding)
- Contact: han.wang@izfp.fraunhofer.de
Background
Compressed sensing is a signal processing technique that allows the reconstruction of a sparse or compressible signal from a small number of acquired measurements, enabling efficient data acquisition directly during the measurement process. Compression matrices tailored to a specific task can offer better performance than random ones. Recently, softmax neural networks have been shown to provide a feasible approach for the design of compression matrices that perform subsampling, an otherwise difficult combinatorial problem.
Task
- Literature research in frequency subsampling, deep unfoldings, and softmax neural networks
- Design of a task-based softmax neural network architecture that learns subsampling matrices
- Training and testing of the network when applied to frequency bin subsampling of synthetic ultrasound NDT data
Reference
[1] Kirchhof, J. et al. Frequency Subsampling of Ultrasound Nondestructive Measurements: Acquisition, Reconstruction, and Performance. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 68.10 (2021): 3174-3191
[2] Diamantaras, K. et al. Sparse Antenna Array Design for MIMO Radar Using Softmax Selection. arXiv preprint arXiv:2102.05092 (2021).
Supervisor
- Dr. Ing. Florian Römer, M. Sc. Han Wang
- Email: han.wang@izfp.fraunhofer.de