Using Convolutional Neural Networks for Material Surface Quality Inspection and Classification

Abstract

In modern industries, undetected material surface defects lead to increased scrap rates and costly rework, primarily due to the limitations of manual inspection done in a slow and inconsistent process with poor small-defect identification. This lack of high-speed inspection solutions for multistage quality control creates critical gaps in production efficiency and product reliability. Hence, the research introduces a Hybridized Convolutional Neural Network for Surface Quality Control model that integrates U-Net with a ResNet34 backbone for precise defect localization and EfficientNet-B4 for defect classification, enhanced by stroboscopic illuminant preprocessing to optimize defect visibility. The research is validated on the Metal Surface Defect Dataset containing 147,824 high-resolution images capturing eight critical industrial defect types. The research results provide 98.2% classification accuracy, 96.5% defect localization precision, minimizes false alarms, and 98.2% recall for incoming material inspection, preventing defective inputs for industrial quality inspection. By integrating these innovations, the research helps manufacturers with a unified, scalable quality inspection platform that reduces human inspection workload by 12% while operating at production line speeds of 20.6 frames/sec and achieves 83.2 fps. The research model delivers a production-ready quality inspection system, which leads to maintaining a significant leap forward in automated surface quality assurance for Industry 4.0 applications.