Denoising With a Compass: Geometry-Guided Diffusion for Robust Low-Quality Image Segmentation

In the field of image segmentation, foundation models such as the Segment Anything Model (SAM) demonstrate remarkable zero-shot generalization capabilities on high-quality images. However, their performance significantly deteriorates when applied to low-quality images commonly encountered in real-world scenarios, such as those corrupted by noise, blur, or low-light conditions. Although existing diffusion-based image enhancement methods can improve visual quality, they often introduce "hallucinated" details that are geometrically inconsistent with the input, thereby compromising the structural accuracy essential for segmentation tasks. To address this issue, this paper proposes a geometry-guided diffusion model (G²D), which, for the first time, explicitly incorporates geometric consistency as a constraint into the enhancement process. G²D leverages pre-trained geometric prior extractors (e.g., depth or edge estimation models) as a structural "compass" to guide the denoising process of the diffusion model via a cross-attention mechanism. Moreover, we innovatively propose a geometric consistency loss that forms a self-supervised closed loop, enforcing the enhanced images to preserve the original scene layout. Experiments on challenging benchmarks such as ImageNet-C and ACDC demonstrate that G²D, as a plug-and-play pre-processing module, significantly improves the performance of segmentation models like SAM on low-quality images without requiring fine-tuning. By effectively restoring visual appearance while preserving geometric structure, our work provides a new direction for task-oriented robust image enhancement.