Sparsity aware coding for single photon sensitive vision using Selective Sensing
Optical coding is widely used in computational imaging systems and is a good approach for designing vision systems. However, most coding methods are developed assuming additive Gaussian noise, while modern optical imaging systems are mainly affected by Poisson noise. Previous studies have highlighted the significant differences between these noise models and proposed coding optimization algorithms for image recovery under Poisson noise. They concluded that the compressibility arising from data variance is crucial for image recovery under Poisson noise. This makes a strong case for the design of end-to-end vision systems that avoid image formation, since the data-driven vision tasks, typically downstream of imaging, is more compressible than imaging itself. In this project, we propose a coding strategy by jointly optimizing an entire vision system, including measurement and inference, using the classification accuracy as a metric. We demonstrate the importance of incorporating Poisson noise in optimizing even the simplest vision systems and propose an approach to achieve it.
PDF Abstract
MNIST
Indian Pines
