Development of Deep Learning-based Approaches for Solving Inverse Problems in Imaging and Comparative Performance Evaluation

Project Coordinator: Assoc. Prof. Figen S. Oktem

Project Type: TÜBİTAK 1001 Scientific and Technological Research Projects Funding Program

Project Budget: 869.686 TL

Project Duration: 36 Months

Project Start Date: 15 March 2021

Funded Personnel: 1 PhD student (full-time), 2 MSc students (full-time)

Project Summary: 

Computational imaging, a rapidly evolving interdisciplinary field, enables new forms of visual information in various applications in natural sciences. In a computational imaging system, an inverse problem has to be solved to reconstruct an image from the acquired raw data. For the solution of these high-dimensional inverse problems, commonly used fast direct inversion methods are not robust to noise. On the other hand, regularization-based methods can offer better reconstruction quality but with higher computational cost.

Recently deep-learning based approaches have been developed to achieve high accuracy with fast reconstruction. In this project, by using a general framework for various inverse problems in imaging, we will develop deep learning-based fast techniques that enable unprecedented reconstruction quality. Moreover, the advantages and disadvantages of the developed techniques will be investigated not only through numerical simulations but also experimentally. For this purpose, inverse problems will be studied by considering the following three general categories separately: a) two-dimensional (2D) linear problems, b) two-dimensional nonlinear problems, c) three- or more-dimensional linear problems. Because the existing deep learning-based methods are mostly developed for two-dimensional linear problems, this project focuses on the development of new methods mostly for the two-dimensional nonlinear problems, and three or more-dimensional linear problems. Likewise, deep learning-based techniques will be studied by considering the following general categories: i) learned direct reconstruction, ii) reconstruction with learned prior, iii) learned iterative reconstruction. Different type of techniques will be developed separately, and then will be compared with each other and commonly used analytical inversion methods. By demonstrating the pros and cons of each approach, it is expected to understand which approaches perform better for different inverse problems and measurement settings. 

Project Team Members: 

  • Süleyman Ayazgök
  • Okyanus Oral
  • Deniz Külbay
  • Ercihan Kara

Past Members: 

  • İrfan Manisalı
  • Utku Gündoğan
  • Can Deniz Bezek
  • Onurcan Kaya
  • Adem Deniz Pişkin

Publications

M.S. Theses:

  • C. D. Bezek, Deep learning-based unrolled reconstruction methods for computational imaging, 2021
  • İ. Manisalı, Deep Learning-based Reconstruction MEthods for Near-field MIMO Radar Imaging, 2022

Journal Papers:

Conference Proceedings:

  • O. Oral and F. S. Oktem, “Plug-and-Play Reconstruction with 3D Deep Prior for Complex-Valued Near-Field MIMO Imaging.” 31st European Signal Processing Conference (EUSIPCO 2023), Helsinki, Finland, 4–8 Sept. 2023.
  • I. Manisali and F. S. Oktem, “Deep Learning-Based Reconstruction for Near-Field MIMO Radar Imaging.” 31st European Signal Processing Conference (EUSIPCO 2023), Helsinki, Finland, 4–8 Sept. 2023.
  • C. D. Bezek and F. S. Oktem, “Unrolling-Based Deep Reconstruction for Compressive Spectral Imaging” In 2021 OSA Imaging and Applied Optics Congress, Virtual Meeting, 19-23 July 2021.

Invited Talks and Seminars:

  • F. S. Oktem, “Computational Imaging and Deep Learning: Making the Invisible Visible.” Invited talk given at the 7th International Symposium on Multidisciplinary Studies and Innovative Technologies, Ankara, Turkey, Oct. 26, 2023.