Explore computational imaging systems and techniques in this webinar from the IEEE Signal Processing Society's SPACE seminar series. Delve into the computational imaging pipeline, focusing on mask-based cameras and the DiffuserCam concept. Learn about multiplexing in computational cameras, compressed sensing for 3D imaging, and neural activity tracking. Examine challenges in object-dependent resolution and model mis-matching, along with proposed solutions. Compare physics-based and deep learning-based image reconstruction methods, and understand the philosophies behind inverse problem solving in computational imaging. Gain insights from Laura Waller of UC Berkeley on cutting-edge developments in the field, including unrolled physics-based algorithms and physics-based learning for improved speed and quality in image reconstruction.
Computational Imaging Systems: From DiffuserCam to Neural Activity Tracking - Seminar 2
IEEE Signal Processing Society via YouTube
Overview
Syllabus
Intro
Computational Imaging Systems
Computational imaging pipeline
Lenses map points to points
Mask-based cameras multiplex
DiffuserCam: stick a scatterer on a sensor
Traditional cameras take direct measurements
Computational cameras can multiplex
DiffuserCam forward model is a convolution
Video from stills with rolling shutter
Point spread function shifts and scales with posit
Single-shot 3D is difficult
Compressed sensing to the rescue! solves under-determined problems via a sparsity prior
3D neural activity tracking
Neural activity tracking with flat DiffuserScope
Improved diffuser for low light
Keeping the objective lens is good
Resolution is more uniform
Tiny microscope version
Single focal length MLA
Multi-focal length MLA
Challenge: object-dependent resolution
Solution?: use condition number of sub-proble
Challenge #2: model mis-match
Solution #2: Local convolution model
Image reconstruction is nonlinear optimizatior
Physics-based image reconstruction
Deep learning based reconstruction
Inverse Problem Philosophies
Unrolled physics-based algorithm makes efficient ne
Physics-based learning improves speed + quali
Taught by
IEEE Signal Processing Society
