Intelligent detection system for micro-nano defects on complex curved surfaces based on 3D laser scanning and deep learning

In modern manufacturing, precise measurement of components with complex curved geometries remains challenging, especially for quantitative size measurements at the micro scale of 5–500 μm. Conventional two-dimensional imaging methods suffer from limited depth-measurement capabilities and cannot capture the complete three-dimensional surface topology with metrological accuracy. This paper presents an integrated intelligent measurement framework that combines multi-line 3D laser profilometry with a deep convolutional neural network for the automated quantitative characterization of micro-scale surface anomalies on non-planar surfaces. After compression and feature-preserving smoothing based on a multi-resolution octree, the raw point cloud data is fed into an enhanced U-Net architecture improved with an attention mechanism. The network simultaneously performs defect segmentation and dimensional measurement across five categories: micro-cracks, pinholes, scratches, contaminant particles, and surface roughness anomalies. Experimental validation on pharmaceutical capsule surfaces, precision optical components, and semiconductor wafers demonstrates a measurement accuracy of 97.3% for dimensional anomalies as small as 5 μm. Within the critical 5–20 μm measurement range, detection recall rates reach 78–91%, with false-positive rates below 2.1%. Measurement uncertainty in high-curvature regions (κ > 0.25 mm−1) contributes to the lower bound of this recall range, as quantified in our failure mode analysis. The average measurement throughput achieves 850 mm2/s, representing a 4–7 × improvement over commercial profilers, while the per-part measurement latency is maintained at 245 ms, making the framework suitable for inline metrology applications.