Variational Depth Estimation for Endoscopic Images: A Lightweight Unsupervised Learning Approach

초록

Monocular depth estimation plays a critical role in minimally invasive surgery (MIS) by enabling 3D scene understanding using a single endoscopic image, supporting compact design and lower system costs. However, existing unsupervised monocular depth estimation (UMDE) methods struggle with photometric inconsistencies and lack robustness in complex endoscopic environments due to the absence of ground truth depth data. In this work, we propose an efficient UMDE framework tailored for endoscopic imaging, featuring an optimized MobileNetV2-based encoder integrated with a First-order Variation Layer (V-layer) to enhance texture and edge representation. Additionally, we incorporate structural pruning to compress the model, reducing its parameters from 5M to 1.25M (4.35 x compression), while preserving accuracy. Experimental results demonstrate that our method reduces Absolute Relative (Abs Rel) loss by 27.6% compared to the baseline, enabling accurate and lightweight depth estimation suitable for real-time 3D reconstruction in minimally invasive surgery, as demonstrated on the SCARED dataset.

키워드