HoloPhase is a projection system designed to render clear images through dynamically changing turbid media such as fog, where Mie scattering would otherwise blur and distort the projected content. Our approach combines computational optics, a Digital Micromirror Device (DMD)– based wavefront shaper, and a camera-in-the-loop optimization pipeline that iteratively updates the projected wavefront to compensate for scattering. By jointly designing the optical setup, fog chamber, and control software, we can recover higher-contrast, sharper images through fog than conventional, uncorrected projection.
Our software pipeline starts by converting target images to DMD patterns using Computer- Generated and Lee hologram methods. Image quality is evaluated with deep IQA metrics (DISTS + MUSIQ). A PPO-based reinforcement learning agent iteratively optimizes DMD parameters to maximize clarity through fog, comparing optimized vs. baseline projections across varying fog densities. Training is conducted in a physics-based Python simulation environment. Our hardware consists of a laser source with beam-expanding/conditioning optics that illuminates a DMD that projects structured light through a sealed fog chamber. Ultrasonic mist generators regulate fog density, a fan circulates air flow, and a downstream camera captures output images for feedback. An embedded controller (e.g., Raspberry Pi) synchronizes DMD updates and image acquisition for repeatable experiments.
We were able to fully create the optical set up including encoding our image into a DMD pattern and capturing clear and blurry images with our camera sensor. We trained the RL model in the simulated environment and were able to increase the clarity of the image, making it ready for physical deployment with our optical set up.