Equivariant Learning for Unsupervised Image Dehazing

TL;DR

This paper tackles unsupervised image dehazing for scientific imaging by introducing Equivariant Image Dehazing (EID), which leverages transformation invariances and haze consistency to recover clear images without ground-truth data. It combines a learnable dehaze network with a learned pseudo-haze model $G_h$ to approximate the unknown haze operator $\mathcal{H}$ via adversarial and cycle-consistency losses, and enforces an equivariance constraint $f_\theta(\mathcal{H}(T_g x)) = T_g f_\theta(x)$. The method demonstrates state-of-the-art performance on medical endoscopy (Cholec80-Haze) and cell microscopy (Cell97), as well as natural images (RESIDE datasets), with ablations highlighting the necessity of jointly optimizing haze-consistency and equivariant losses. By unifying physics-informed modeling with equivariant self-supervision, EID offers a practical unsupervised solution for haze removal in diverse scientific imaging contexts, potentially enabling more accurate downstream analyses. The work also points to future extensions including domain priors, efficiency improvements, and applications to other inverse-imaging problems.

Abstract

Image Dehazing (ID) aims to produce a clear image from an observation contaminated by haze. Current ID methods typically rely on carefully crafted priors or extensive haze-free ground truth, both of which are expensive or impractical to acquire, particularly in the context of scientific imaging. We propose a new unsupervised learning framework called Equivariant Image Dehazing (EID) that exploits the symmetry of image signals to restore clarity to hazy observations. By enforcing haze consistency and systematic equivariance, EID can recover clear patterns directly from raw, hazy images. Additionally, we propose an adversarial learning strategy to model unknown haze physics and facilitate EID learning. Experiments on two scientific image dehazing benchmarks (including cell microscopy and medical endoscopy) and on natural image dehazing have demonstrated that EID significantly outperforms state-of-the-art approaches. By unifying equivariant learning with modelling haze physics, we hope that EID will enable more versatile and effective haze removal in scientific imaging. Code and datasets will be published.

Figures (10)

• Figure 1: Performance comparison of our proposed EID with existing image dehazing approaches for scientific imaging and natural image restoration.
• Figure 2: Overview of EID. (a) Motivation: scientific or natural images often show invariances to certain transformations (e.g. rotation), as seen in the real endoscopic examples above. (b) Training without ground truth: given haze images $\{y\}$, the dehazing model $f_\theta$ first outputs coarse dehazed images $f_\theta(y)$, which are then fed back into the haze model to minimize haze consistency. At the same time, $f_\theta(y)$ is transformed by group action $T_g$ and perform haze and dehaze again to achieve system equivariance. By continuing the training loop, $f_\theta$ learns the fine and clean dehazed image. (c) Testing: the EID-trained $f_\theta$ can be used directly to dehaze new, unseen haze images.
• Figure 3: (a) Adversarial training of the generator $G_h$ for modelling the hazing physics. (b) The proposed architecture of $G_h$.
• Figure 4: Comparison of endoscopy image dehazing performance between EID and DCP. We use the Dense Prediction Transformer dpt-depthpredict-sm to predict the depth map and add haze following Eq.\ref{['haze physics']}. PSNR $\uparrow$ and SSIM $\uparrow$ values are displayed below each image.
• Figure 5: Comparison of natural image dehazing performance between EID and NLP berman-ncp-prior2016-intro. We utilize provided depth map in RESIDE dataset and follow Eq.(\ref{['haze physics']}) to add haze. PSNR $\uparrow$ and SSIM $\uparrow$ values are shown below each image.