Unsupervised Learning
for Industrial Defect Detection

In automated assembly production lines, defects such as missing components, misalignment, connector errors, or labeling inconsistencies frequently arise and pose critical challenges to manufacturing quality assurance. Conventional inspection methods predominantly rely on manual labor, which is inherently inefficient and susceptible to fatigue-induced errors, subjective inconsistencies, and limited scalability. These shortcomings often result in false detections or overlooked defects, ultimately reducing overall yield and compromising production stability.

To address these limitations, this research introduces a Learning from Normal Samples (LNS)-based computer vision framework. Instead of depending on extensive annotated datasets, the proposed approach leverages exclusively defect-free samples to model the standard structural and logical configurations of assembly machines. Once trained, the system autonomously detects anomalies in newly acquired images by identifying deviations from learned normality. This paradigm not only reduces reliance on costly large-scale annotations but also enhances the efficiency, robustness, and accuracy of defect detection in real-world manufacturing environments.