Generating Realistic 3D Surface Defects for Training AI-Based Industrial Inspection Systems

Ensuring the surface quality of industrial components requires the detection of small superficial defects—such as cracks, bumps, and peaks—using high-resolution 3D sensors. However, training machine learning algorithms for this task is constrained by the limited availability of annotated 3D defect datasets. In this work, we propose a method for generating synthetic 3D datasets of surface defects using Free-Form Deformation applied to CAD models. The technique allows localized insertion of parametrized defects adapted to the object’s geometry and supports diverse defect types through customizable elevation maps. To simulate realistic sensor acquisition, we replicate the scanning process of a profilometric 3D sensor, including surface and sensor noise. The output consists of labeled depth images that closely resemble real-world sensor data. We validate the approach by training object detection networks on synthetic datasets and evaluating their performance on real samples, demonstrating comparable accuracy. The proposed method reduces reliance on rare and costly real defect data, offering a scalable tool for developing and testing surface inspection systems in industrial contexts.