Dimensioning Defects with Monocular Vision in Automated Optical Inspection

Abstract

Automated optical inspection (AOI) systems are common tools for quality control in industrial manufacturing. AOI systems use robotic systems to load components, take images, and detect defects, often also characterizing the defects by size or class. Among various approaches to this machine vision, monocular systems are popular because they are cheap and simple to integrate while offering intuitive visualization. However, monocular vision alone lacks depth resolution and struggles to accurately dimension defects on 3D surfaces, especially if the imaged component’s pose is ambiguous. This paper presents a transparent, open-sourced, end-to-end image processing pipeline for dimensioning surface defects on industrial components using RGB images. The pipeline estimates component pose through a 2D-3D correspondence, segments defects with machine learning or image comparison techniques, then projects the component’s CAD mesh into the image to calculate the lengths of segmented defect instances. The pipeline was developed on a 3D-printed test object and demonstrated with each of three segmentation methods, yielding defect dimensions with average error between 0.6-1.2mm.