Automated Detection and Classification of Aluminium Surface Defects Using YOLOv8 and Swin Transformers

Abstract

Ensuring defect-free surfaces in aluminium manufacturing is vital for product quality and reliability. This project introduces a hybrid deep learning framework for automated detection and classification of aluminium surface defects, integrating YOLOv8 and SWIN Transformer models. YOLOv8 delivers high-speed and accurate localization of surface anomalies, while the SWIN Transformer, with its hierarchical attention mechanism, excels in fine-grained classification of defects such as scratches, dents, and discolorations. A custom aluminium surface defect dataset was used to train the system, leveraging transfer learning and data augmentation for enhanced generalization and efficiency. Evaluation using metrics like mean Average Precision (mAP), precision, recall, and F1-score confirms the framework's high performance under diverse industrial conditions. The approach offers a scalable, real-time inspection solution, minimizing human error and aligning with Industry 4.0 automation goals in quality assurance.