Reflection removal using diffusion model based on stochastic differential equation

Abstract

An image with reflections can be defined as a composition of the background and reflection images, where the two layers are mixed together. Reflection removal is a highly challenging and ill-posed problem that involves separating these two layers. To address this issue, our work utilizes a diffusion model as a type of auto-encoder with stochastic differential equations in a dual manner. This approach allows both the background and reflection images to gradually revert to their original states. Typically, a diffusion model goes through a forward process where the input image is subjected to pure Gaussian noise. However, in reflection removal, where both layers are mixed, the forward process is designed to degrade each image into mixed images. This is achieved by employing stochastic differential equations, enabling the background and reflection images to blend together. To revert to their original states, the values added during the forward process are estimated and subtracted through deep learning. This process facilitates a progressive return to the image of the previous timesteps. Through our approach, we successfully restore both the background and reflection images from a single image containing reflections, allowing for the bidirectional restoration of scenes reflected on light-transmitting material surfaces.