Monocular depth estimation for endoscopic surgery using swin transformer and conditional random fields

Monocular depth estimation in endoscopic surgery is challenging due to weak tissue textures, uneven illumination, specular reflections, and frequent smoke or instrument interference. This paper proposes CRFDepth, a pseudo‑supervised framework for endoscopic monocular depth estimation. A Swin Transformer encoder extracts hierarchical global features under weak‑texture conditions. The proposed CRFDepth framework effectively addresses key challenges in endoscopic monocular depth estimation, achieving superior accuracy and robustness. A Pyramid Scene Parsing module aggregates multi‑scale contextual information to improve structural perception in distorted regions. A NeWCRF decoder performs cross‑scale structured inference for boundary refinement and detail recovery. A RangeAwareLoss is introduced to stabilize optimization within the narrow endoscopic depth range. The network is trained with stereo-derived pseudo-ground-truth supervision using binocular image pairs, while only monocular input is required during inference. Experiments were conducted on the da Vinci laparoscopic dataset. CRFDepth achieves 44.27 dB PSNR and 0.9942 SSIM, outperforming baseline methods, and demonstrates strong robustness to occlusion and smoke interference.