3D Vision-Based Inspection Using Multi-View Reconstruction and Depth Estimation for Industrial Quality Control

Abstract

Industrial inspection remains difficult when defects are subtle, weakly textured,
or partially occluded, since 2D vision provides limited geometric evidence.
This paper presents a compact 3D inspection framework that combines multi-view reconstruction, depth estimation using the pretrained MiDaS v3.1 DPT-Hybrid model, and Open3D-based geometric analysis for OK/NOK classification and conformity assessment. The pipeline integrates feature matching, pose estimation, reconstruction, dense depth inference, and point-to-reference deviation analysis in one decision process. On a multi-view dataset of mechanical and metallic parts, the method achieves 95.1% accuracy, 94.2% precision, 93.5% recall, 0.46 mm mean geometric error, and 54 ms average processing time per part. Relative to RGB-only and depth-only baselines, the fused framework is more robust to weak-texture and geometry-driven defects while remaining compatible with practical deployment.