Abstract: To address the challenges of high-temperature interference, complex defect morphology, and the decoupling of structural and texture information in hot rail surface quality inspection, this paper presents and verifies the core algorithm for the 2D/3D vision collaborative inspection system. In terms of 2D detection, alight weight DFAM-YOLO-Met model is proposed, which can efficiently recognize tiny defects by coupling the attention module and feature enhancement technologies with YOLO. The proposed DFAM-YOLO-Met(Defect-Focused Attention Module YOLO for Metallurgy) achieves 91.6% mAP and 0.89 F1-sccore on the custom dataset. In terms of 3D reconstruction, a three-stage point cloud optimization and segmented curvature fitting strategy is proposed. A thermal response compensation mechanism is also introduced that enables the system to achieve online dimensional precision measurement under high temperatures, with the error reduced from ±0.28 mm to ±0.12 mm, meeting the national standard requirement of ≤0.15 mm. In terms of multimodal fusion, a graph-point mapping and module mutual reference mechanism are constructed to realize the linkage review of 2D and 3D detection results, reducing the system false alarm rate from 2.4% to 1.5%, and improving the recall of complex defects by 11.7%. Online tests show that the system can run stably for 72 hours on a 5 m/s rail production line, with a single-frame inference latency of 11.4 ms and a GPU power consumption of less than 15 W. The proposed core algorithm has good robustness and deployment value, and provides a technical path and practical reference for the metallurgical industry to realize intelligent quality inspection under high temperature and high-speed conditions.