Learning Robust Imaging Models without Paired Data

Explore robust imaging models without paired data in this 56-minute conference talk from the 44th Imaging & Inverse Problems (IMAGINE) OneWorld SIAM-IS Virtual Seminar Series. Delve into Prof. Chenglong Bao's research on combining classical mathematical modeling with deep neural networks to improve interpretability in image processing. Discover how Bayesian inference frameworks can be leveraged to build AI-aided robust models for applications such as image denoising and segmentation. Examine the challenges of collecting paired training data and learn about innovative approaches to overcome this limitation. Gain insights into linear approximation for imaging processes, error effects, and the data bottleneck in deep learning. Analyze quantitative and qualitative results for real noisy images, including examples from Huawei. Understand the probabilistic model for image segmentation and explore unpaired degradation modeling techniques. Conclude with a summary of experimental findings and visual results demonstrating the effectiveness of these novel approaches in practical imaging systems.

Intro
Outline
Linear approximation for imaging process
The error effects
Model based approaches
Deep learning (DL) based approaches
Data bottleneck in DL
Data collection in video superresolution
Goal of the talk
Image denoising
The basic idea
Model formulation
Numerical method
One remark on overfitting issue
Quantitative results for real noisy images
Qualitative results
Latent space verification
Real-world noisy images from Huawei
Image segmentation
Probabilistic model
Examples
Deep CV model
Distributions in latent space
Motivation
The case of unpaired datasets
Unpaired degradation modeling
The idea
The loss function
Inference invariant condition
Synthetic noisy images
Experiments
Visual results
Summary