Comparative Analysis of Illumination Levels in the Performance of Neural Networks for Defect Detection in Metallic Surfaces

Abstract

Defect detection on metallic surfaces using artificial vision systems is highly sensitive to environmental disturbances such as reflections, shadows, and color temperature variations. These factors directly affect the performance of artificial neural networks (ANNs) during training and inference. This study evaluates the impact of three illumination ranges—optimal (150–300 Lux), moderate (350–500 Lux), and inadequate (550–1200 Lux)—on the performance of YOLO V5, YOLO V8, and Roboflow 3.0 for metallic surface defect detection. A dataset was collected from local metalworking workshops, and a spiral methodology was implemented to iteratively evaluate model behavior. Results show that optimal illumination conditions significantly improve performance, reaching a maximum mAP of 99.5% for iron surfaces and 88.7% for aluminum. The findings demonstrate that illumination control plays a critical role in ensuring reliable defect detection, particularly for highly reflective materials.