Enhancing Object Clarity In Single Channel Night Vision Images Using Deep Reinforcement Learning

This paper implements a system of enhancing single channel night vision images using reinforcement learning approach and optimizing pixel prediction using q-table. We implemented some models to learn and process a small static images dataset using a reward bias q-table in a reinforcement learning architecture thus optimizing computational complexities and requirements of large dataset with the help of q-table. It also outperformed with respect to existing CNN models like SRCNN. Where SRCNN is observed to generate a PSNR of 24. S13 on average at 256 batch size. Our system generated a PSNR of 24.1 on average with results in a 10.29% increase of relative efficiency at 3000 epoch. It has shown a 10.39% and 10.36% increase of efficiency with respect to VDSR (at 12S batch size) model and DRCN (at filter number 16) model respectively.