Subject overview

Images, bursts and videos acquired by smartphones, action cameras, and even professional digital cameras are affected by blur of different sources. Even though some of these (intrinsic) sources can be calibrated, other (extrinsic) blur sources like motion, and defocus are unknown and need to be guessed from the degraded image itself. This blind deblurring problem is extremely ill-posed, and despite recent progress thanks to deep learning and generative AI techniques, the problem is still wide open for future improvements.

During the last three years, our team has significantly improved the state of the art performance in blind image deblurring. On one hand [Laroche2023] achieves reconstructions of unprecedented quality, but at a computational cost that is still not well-suited for real-time or embedded applications. On the other hand [Carbajal2023; Laroche2022] provide more flexibility at a lower cost but can only handle blur kernels of relatively small size. The aim of this thesis is to significantly improve the state of the art in blind image and video deblurring, in terms of image quality, real world applicability and computational speed by leveraging both mathematical models and cutting edge deep learning technologies.

Detailed PhD subject

http://up5.fr/2024-phd-gopro-map5

Contact

Andrés Almansa (see email in detailed description)

Supervising Team

Academic Advisor: Andrés Almansa (CNRS & Université Paris Cité)
Industrial Advisor: Eva Coupeté (GoPro)

Funding & work environment

This PhD position is funded by GoPro Research Paris, via a CIFRE agreement with Université Paris Cité.
As a PhD student you will be part of both GoPro’s research team in Boulogne, and of MAP5 lab in Paris 6ème arrondissement.

In addition to the academic advisors you will be part of an international collaboration team.

Candidate

The successful candidate is expected to hold a Master’s degree in Applied Mathematics, Computer Science, Electrical Engineering, or a closely related discipline. They should have good programming skills in deep learning and scientific computing environments (*e.g.* Python, NumPy, PyTorch, TensorFlow, JAX).

In addition, the following qualifications are desired:

• Solid background in some of the following areas: Computational Imaging, Image Restoration, Variational Methods in Imaging, Bayesian Methods for Inverse Problems.
• Fluent in either French or English (written and spoken) and good communication and interpersonal skills.

Application

Candidates should apply by filling in the following form including at least

• CV and motivation letter
• Grades from your M2, M1
• One or two academic references (and their recommendation letters, if available)

Online form: https://forms.gle/5aQR5TTRBWiB3cVP6