# Advancing workpiece dimension measurement: Integrating AI-based edge detection with machine vision and coordinate measuring systems

## TL;DR

This paper explores using AI-based edge detection with machine vision and coordinate measuring systems to improve workpiece dimension measurement accuracy.

## Contribution

A standardized methodology integrating deep learning with industrial inspection systems is proposed for precise workpiece measurement.

## Key findings

- The CNN achieved 100% classification accuracy in identifying interference regions and predicting circularity.
- The CNN matched CMM precision (r = 1.000) and outperformed vision systems in multi-hole measurements.
- The method showed cross-material scalability to Drelin after retraining on 200 images.

## Abstract

Image preprocessing and edge detection are critical in industrial machine vision for workpiece dimension measurement. Challenges arise from interference regions on workpiece surfaces, complicating edge detection and roundness assessment. This paper investigates the application of AI-based detection methods within the industrial image analysis framework of coordinate measuring machines. Initially, two models with varying hole sizes and counts were designed in SolidWorks, fabricated using a Prolight 3-axis CNC milling machine, and analyzed. A transfer learning approach mitigated overfitting on the limited dataset of model surface features. The study employed a Convolutional Neural Network (CNN) to identify interference regions and predict circularity, enhancing measurement accuracy. Validated with a testing dataset, the CNN achieved 100% classification accuracy, confirmed by a Confusion Matrix. Fine-tuning of the CNN with specific training data leveraged image preprocessing to enhance features via multi-layer convolution, pooling, and detailed analysis through fully connected layers. Comparative diameter analysis across Models 1–2 showed all methods maintained ≤0.05 mm deviation from actual values, with CNN exhibiting minor variations at Model 1’s points 3,7,9 while matching CMM precision (r = 1.000) and outperforming vision systems in Model 2’s multi-hole measurements, supported by ANOVA-confirmed discrimination (F = 34,514,683, p < .001) and cross-material scalability to Drelin via 200-image retraining. The results underscore the effectiveness of integrating deep learning techniques into industrial inspection, contributing a standardized methodology for precise workpiece dimension measurement. This research highlights the potential of combining machine vision, deep learning, and coordinate measuring systems to advance industrial measurement processes.