Cal Poly PI Improves Microscope Imaging with ML

Researchers at Cal Poly have developed a new method for improving the quality of microscope images taken under extreme low-light conditions.  Cal Poly Professor Jonathan Ventura led the research team of undergraduate researchers from several institutions and a faculty collaborator at University of Colorado, Colorado Springs.  The method uses machine learning to teach the computer how to "de-noise" low-light images, without ever showing the machine the corresponding images taken under high light exposure.  Instead, the system learns to predict the denoised image and estimate the noise level from the noisy images alone, in a scheme called self-supervised learning.

The researchers extended recent work on this topic to handle the Poisson-Gaussian noise which is inherent in microscope images, and also improved the training procedure using a novel uncalibrated approach.  A paper about the method, with three undergraduate students authors including first author Wesley Khademi, a computer science student at Cal Poly, will appear at the IEEE Winter Conference on Applications of Computer Vision (WACV) in January.  The manuscript, related code, and example images are available at https://jonathanventura.github.io/self-supervised-poisson-gaussian/.