BiT‐MoFE: Unified illumination correction in capsule endoscopy

Abstract As a prevalent non‐invasive screening technique, Wireless Capsule Endoscopy is often hindered by poor image quality, including under‐/overexposure and low light condition. While illumination correction based on diffusion modeling or frequency‐domain decomposition has shown effectiveness, existing methods often (1) underexploit structural information, and (2) lack adaptive strategies for varying illumination degradations, leading to suboptimal restoration and unnecessary computation. To this end, we propose Brownian Bridge Diffusion Transformer‐Mixture‐of‐Experts (BiT‐MoFE), a unified adaptive framework that integrates the merits of the two paradigms for endoscopic illumination correction. We adopt a Brownian Bridge Diffusion framework, in which an efficient Transformer serves as the backbone network, and design a frequency‐decomposed MoFEs module to explicitly handle illumination and image structure simultaneously. By dynamically selecting the most suitable experts conditioned on exposure cues and diffusion timesteps, our framework achieves a strong balance between restoration fidelity and computational efficiency. Extensive experiments on multiple public datasets demonstrate that BiT‐MoFE achieves state‐of‐the‐art performance on both exposure correction and low‐light enhancement tasks.