Classical Computer Vision for Automated Quality Inspection: A Reproducible Study Integrating SPC and Process Capability

Abstract

Automated visual inspection is often evaluated as an image-classification task, although an industrial quality decision also depends on defect escape risk, dimensional measurement error, process behavior, and the distinction between engineering specifications and statistical control. This work presents a reproducible classical-computer-vision workflow integrating OpenCV-based inspection, dimensional measurement, deterministic pass/fail logic, confusion-matrix analysis, threshold sensitivity, Individuals statistical process control (SPC), and descriptive process capability analysis. The method is evaluated on a frozen synthetic experiment containing 240 part images generated with seed 42, eight ground-truth classes, a nominal width of 50.00 mm, and specification limits of 49.80–50.20 mm. At the frozen visual threshold of 95, the system achieved 88.33% accuracy, 100.00% precision, and 81.58% defect recall, with 28 false accepts and no false rejects. Low-contrast surface defects accounted for 21 of the 28 escapes. An exploratory threshold sensitivity study reduced the observed false-accept rate from 18.42% at the frozen baseline to 4.61% at threshold 150 without introducing false rejects in the synthetic dataset. Vision-based dimensional measurement produced a bias of +0.006 mm, MAE of 0.095 mm, and RMSE of 0.184 mm. Controlled dimensional drift toward the upper specification limit reduced descriptive Cpk from 0.45 to 0.29, followed by recovery to 0.50. The study demonstrates how an interpretable computer-vision inspection pipeline can be evaluated as a quality-engineering system using defect-escape metrics, measurement analysis, SPC, and process capability. Because the experiment uses synthetic imagery and simulated calibration, the results establish reproducibility and controlled ground-truth recovery within the designed case study rather than industrial validation.