Diffusion-Based Multi-Degradation Remover for Fiber Bundle Image Restoration

Due to the specialized imaging environment, the images captured through fiber bundles are usually with low-quality visualization, which may bring great challenges to disease diagnosis. Recently, deep-learning based methods have shown promising performance in solving image degradation task. However, current discriminative models attempt to restore images by learning pixel-to-pixel mappings, which often results in overly smooth outputs that lack detail, especially when a significant amount of valid information is lost from the image. To address this problem, we propose a Diffusion-based Multi-degradation Remover (DMR) to first restore the degraded fiber bundle image to a coarse result, then restore the coarse result to a high-quality image. Specifically, images with large-scale grid-like occlusions and information loss are initially processed using a diffusion model. By leveraging the diffusion model's distribution estimation capabilities, high-fidelity image details are sampled, and this result is further refined through a convolutional neural network to produce the final detailed output. To verify the effectiveness of our model, we design an optical acquisition device to construct a real-world Fiber-bundLe image dataset. The experimental results demonstrate the superiority of our model in image restoration for fiber-optic imaging.