A Deep Learning-Based Cyber-Physical Approach for Visual Inspection of Defects in Pad Printing

Abstract

Industrial quality inspection processes based on manual visual analysis are prone to human error, fatigue, low repeatability, and high operational variability, especially in high-throughput manufacturing environments. This work presents a Cyber-Physical Framework for Intelligent Visual Inspection, an intelligent computer vision system developed for the automated inspection of logos applied via pad printing on air conditioner condenser cabinets. The proposed architecture integrates deep learning techniques, including YOLOv8-based detection, Region of Interest (ROI) extraction, RGB colorimetric analysis, dimensional validation, and the generation of synthetic images using Generative Adversarial Networks (GANs). The system was designed with real-world industry constraints in mind, such as lighting variations, positioning deviations, mechanical vibrations, and takt time limitations. Experiments were conducted using real and synthetic datasets acquired under controlled industrial conditions. The results demonstrated high robustness in defect detection and good operational repeatability, highlighting the feasibility of integrating the solution into industrial environments aligned with Industry 4.0 concepts. In addition, the study discusses the integration of mechanical design, automation systems, and artificial intelligence to ensure operational stability and reliability in real production lines.