GeoDyn-Stereo: a geometry-guided dynamic fusion network for disparity estimation

Stereo vision, as a core optical method for three-dimensional perception, is critically important for in situ monitoring in high-precision manufacturing. However, its application in dynamic processes such as laser machining is severely constrained by the inherent trade-off between reconstruction accuracy and computational speed. We present a geometry-guided dynamic fusion network (GeoDyn-Stereo) designed to resolve this conflict. The network architecture explicitly integrates geometric constraints to fortify the cost-volume representation, while a dynamic fusion mechanism enhances robustness in optically challenging regions. A subsequent GRU-based module refines the disparity map, preserving fine edge details. To empirically validate the method for optical metrology, we implemented a calibrated binocular vision system and constructed a corresponding dataset simulating laser processing scenarios. Comprehensive evaluations on standard benchmarks (Scene Flow, KITTI, and Middlebury) and our dataset confirm the advancements. The proposed model reduces the end-point error (EPE) and disparity outlier rate (D1-error) by 10.96% and 6.45%, respectively, and achieves a 22% speed improvement, increasing the frame rate from 0.37 FPS to 0.45 FPS. These results demonstrate that the method successfully meets the dual requirements of real-time feedback and sub-millimeter spatial resolution for in-process optical inspection, offering a practical vision-based solution for laser machining and analogous industrial applications.