07-29, 18:10–18:20 (UTC), Red Track
The Dark Channel Prior was introduced by He, et al. as a method to dehaze a single image. Since its publication in 2010, other authors have sought to improve this dehazing method. Using the parameters that other authors have tuned as a guide, we parameterize the Dark Channel Prior dehazing method and utilize Zygote to apply gradient based optimization.
This talk will briefly discuss the Dark Channel Prior implementation and will emphasize how it is possible to parameterize the algorithm and replace key portions with small neural networks. It will also cover gathering useful data and defining a meaningful loss function for measuring the performance of the dehaze algorithm. The hope is that others will see how programming in Julia can allow for easy application of differentiable programming in variety of scientific domains.
Vandy Tombs is an Applied Mathematician in Geographic Data Science group at Oak Ridge National Laboratory. She enjoys applying group theory to develop new machine learning architectures and utilizing differentiable programming to improve existing scientific algorithms.