Overview
Explore new approaches to image and video reconstruction using deep learning in this F8 2019 conference talk. Discover how to improve images and videos, including increasing resolution and colorizing black & white footage, using advanced techniques like generative adversarial networks (GANs). Learn about faster, more efficient PyTorch-based tools developed by fast.ai, the Salk Institute, and DeOldify that can be trained in just hours on a single GPU. Delve into topics such as generative imaging, loss functions, self-attention, and reliable feature detection. Witness demonstrations of these techniques applied to various scenarios, including colorizing old movies and enhancing microscopy images. Gain insights into the challenges and solutions in brain imaging, live fluorescence imaging, and the balance between speed, quality, and data size in image processing.
Syllabus
Introduction
generative imaging
do defy
Golden Gate Bridge
Porcelain
Basic Approach
Units
Loss Function
Gains
Moving Images
Three Key Approaches
Reliable Feature Detection
SelfAttention
Solution
Demonstration
Conclusion
Salk Institute
The eternal triangle of compromise
How to image the brain
The new strategy
Results
Data is cleaner
False positives
Live fluorescence imaging
In conclusion
Jeremy
Taught by
Meta Developers