Automated metrology for additively manufactured parts using deep learning and computer vision

Abstract

This paper presents a comprehensive, fully automated pipeline for dimension analysis of machine parts from 2D image data. The pipeline is designed to process high-resolution grayscale scans, leveraging advanced computer vision and deep learning techniques to extract and summarize part dimensions. The first stage employs the Line Segment Detector (LSD) to identify linear features, filter irrelevant segments, and calculate distances between parallel lines. The second stage utilizes a CNN-based Ellipse Detector (ElDet) to detect and measure elliptical and circular features, including concentric ellipses, through a recursive detection strategy for improved accuracy. The proposed method significantly enhances measurement speed and automation over traditional computer vision-based techniques while maintaining accuracy, efficiently handling both simple and complex geometries while reducing the need for manual analysis. Extensive evaluation on real and synthetic datasets highlights the pipeline's robustness and effectiveness in industrial applications.