438. Neural Blind Deconvolution for Poisson Data
Invited abstract in session MC-3: First-order methods in modern optimization (Part II), stream Large scale optimization: methods and algorithms.
Monday, 14:00-16:00Room: B100/4011
Authors (first author is the speaker)
| 1. | Alessandro Benfenati
|
| Environmental and Science Policy, Universita' di Milano | |
| 2. | Ambra Catozzi
|
| Università di Parma | |
| 3. | Valeria Ruggiero
|
| Università di Ferrara |
Abstract
The Blind Deconvolution problem arises in several scientific imaging fields, such as Microscopy, Medicine, and Astronomy. The Point Spread Function (PSF) may only be approximately known: blind deconvolution aims to reconstruct the image when only the recorded data is available. Among the standard variational approaches, Deep Learning techniques have gained attention due to their impressive performance. The Deep Image Prior framework has been employed to solve this task, giving rise to the so-called Neural Blind Deconvolution, where the unknown PSF and image are estimated through two separate neural networks. In this paper, we focus on microscopy images, where the predominant noise is of Poisson type: the objective function to minimize is hence the generalized Kullback-Leibler divergence, coupled with regularization terms for both the PSF and the image. Furthermore, we propose modifying the standard NBD formulation by incorporating an upper bound for the blur kernel, which depends on the optical instrument. A numerical solution is obtained via an alternating Proximal Gradient Descent-Ascent procedure, resulting in the Double Deep Image Prior for Poisson noise algorithm. We evaluate the proposed strategy on both synthetic and real-world images, achieving promising results and demonstrating that the correct choice of loss and regularization functions strongly depends on the specific application.
Keywords
- Data driven optimization
- Large-scale optimization
Status: accepted
Back to the list of papers