Development of a Machine Vision-Based Quality Inspection System for Industrial Production Lines

Abstract

Industrial production lines increasingly require high-speed, repeatable, and traceable quality inspection that cannot always be achieved by manual visual checking. This methodology paper presents the development framework for an inline machine vision-based quality inspection system that combines controlled illumination, industrial image acquisition, image pre-processing, convolutional-neural-network-based defect recognition, decision logic, and production-line actuation. The proposed workflow begins with inspection-task definition and imaging trials, followed by dataset construction, defect annotation, transfer-learning-based model development, threshold calibration, and integration with a programmable logic controller or line controller. Performance is evaluated using precision, recall, F1-score, accuracy, false-reject rate, false-accept rate, and inspection cycle time. An illustrative prototype evaluation shows how a model can be validated under realistic variations in lighting, part position, surface texture, and defect size without presenting the values as plant-measured production data. The methodology emphasizes robust data preparation, controlled optics, real-time inference, traceability, and human-in-the-loop handling of low-confidence cases. The framework is intended for scalable application to surface defects, missing components, assembly errors, dimensional appearance checks, and product classification in smart manufacturing environments.