Review of Automated 3D Surface Defect Detection

Abstract

Automated 3D surface defect detection has become essential in modern manufacturing, where high precision, complex geometries, and large production volumes exceed the capabilities of manual inspection. This review presents a focused comparative analysis of three major point-cloud–based inspection approaches: Random Sample Consensus (RANSAC), Moving Least Squares (MLS), and deep learning methods based on 3D-to-2D residual mapping (DentNet). RANSAC provides a fast and interpretable baseline for flat surfaces but struggles with curvature. MLS addresses this limitation through adaptive local surface reconstruction, offering higher accuracy on curved components at the cost of increased computation. Deep learning approaches deliver the highest throughput, enabling real-time inspection and multi-class defect identification, but require representative training data and careful preprocessing. The review synthesizes their strengths, limitations, computational characteristics, and industrial deployment considerations. Overall, method selection should be guided by part geometry, defect size requirements, and throughput constraints. Emerging hybrid geometric–learning pipelines offer promising directions for achieving both robustness and real-time performance in next-generation 3D inspection systems.