SIParCS 2024 - Jefferson Boothe

Jefferson Boothe

Jefferson Boothe, University of Pittsburgh

Distributed Holographic Image Processing with Neural Networks

Recorded Talk

This project aims to improve the performance and scalability of a neural network processor for holographic images of cloud particles obtained using the Holographic Detector for Clouds (HOLODEC). The HOLODEC sensor is a cloud particle imager housed at NCAR RAF that captures holographic images of liquid and ice cloud particles. The holograms can be computationally refocused and subsequently passed through a “U-Net” style neural network to recognize and detect particles at a large range of depths. Computer vision based post-processing steps are then leveraged to obtain physical characteristics and distributions of the particles detected. Due to the extremely large quantity of data, it is necessary that the pipeline be scalable across many GPU’s on the Derecho supercomputer. We demonstrate improvements to the processing pipeline as well as the ability to process holograms at large scale. We also successfully calculate the liquid density to be within 93% and effective particle radius to be within one pixel on synthetic test data.

Mentors: John Schreck, Matthew Hayman

Slides and poster