Research and selected publications.

Automated visual inspection of linear and areal surface defects: A metrological survey

Henri Vennikas, Olev Märtens, Tavo Kangru, and Yannick Le Moullec

Abstract

Automated visual inspection commonly reports detection and segmentation performance, while comparatively few studies provide geometric defect measurements in physical units with sufficient measurement evidence. This survey examines 49 surface-defect quantification papers published from 2020 to 2025, comparing reported width, length, and area results through their reference, scale, dispersion, uncertainty, and traceability information. It identifies recurring reporting gaps and proposes a lightweight framework for treating machine-learning-assisted inspection as a measuring system.

doi:10.1016/j.measurement.2026.122522

MetroDef: Metrologically Grounded Synthetic Surface-Defect Data

Henri Vennikas, Olev Martens, and Yannick Le Moullec

Abstract

MetroDef is a geometry-first synthetic defect generator for measurement-oriented automated visual inspection. It stores defect ground truth as explicit metric geometry with declared measurands rather than tying it to a raster image at one resolution. The same reference geometry can be serialized, rasterized, printed, captured, and measured through different pipelines while its underlying values remain fixed. This makes it possible to study how resolution, sensing, processing, and measurement conventions affect reported defect dimensions.

doi:10.1109/AI4IM69129.2026.11558236

Modeling Rough Painted Metal Surfaces for Automated Visual Inspection

Henri Vennikas, Olev Martens, and Yannick Le Moullec

Abstract

This paper presents a physically grounded, renderer-independent model for synthesizing high-resolution images of standardized metal test panels with matte camouflage coatings. Measurements and observations define panel tolerances, placement variation, layered paint structure, and spatial texture generated through fractal Brownian motion and Voronoi noise. Comparison with transillumination scans shows that the model reproduces relevant roughness and material variation while retaining exact geometry for photorealistic images and pixel-accurate annotation masks.

doi:10.1109/IST66504.2025.11268402