Explore the latest advancements in self-supervised deep learning for image recovery in this SIAM-IS Virtual Seminar Series talk. Delve into Hui Ji's research from the National University of Singapore, focusing on a dataset-free approach to solving ill-posed inverse problems in image reconstruction. Learn about Bayesian deep networks and their application in compressive sensing, comparing the performance of this self-supervised method to traditional non-learning techniques and supervised deep learning approaches. Gain insights into data augmentation, self-supervised loss functions, and the stability of training in this cutting-edge field. Examine experimental results demonstrating the method's effectiveness in reconstructing images from limited and noisy measurements, and understand its potential to reduce the cost and complexity of creating labeled training datasets for image recovery tasks.
Self-supervised Deep Learning for Image Recovery - SIAM-IS Virtual Seminar
Society for Industrial and Applied Mathematics via YouTube
Overview
Syllabus
Intro
Inverse problem: Image deblurring
Image recovery/reconstruction
Regularization methods for image recovery
Non-local self-similarity prior of images
Deep learning for linear inverse problem
An example: Unrolling half-quadratic splitting scheme
Dataset-dependence of supervised learning methods
A self-supervised approach to general image recovery proble
Part 1: Data augmentation & Self-supervised loss function
Data augmentation via Bernoulli random sampling
Recap: Self-supervised loss function
Testing with Bayesian neural network
Experiments on removing Gaussian white noise from images
Stability of training
Visual inspection
Taught by
Society for Industrial and Applied Mathematics
