From Defect Detection to Defect Type Recognition: A Vision Transformer-Based Hybrid Framework for Magnetic Tile Surface Defect Inspection

Abstract

Defects such as cracks, blowhole, fray, uneven, breaks, and surface irregularities that occur during the production process of magnetic tiles directly and negatively affect the performance and lifespan of electric motors. Therefore, early and accurate inspection of magnetic tiles on the production line is critically important for industrial quality control processes. However, traditional visual inspection methods have significant limitations such as high cost, susceptibility to human error, and low consistency. In this study, a two-stage deep learning framework based on Vision Transformer (ViT-Base Patch16, 224×224) is proposed for the automatic detection of magnetic tile surface defects and detailed classification of defect types. In the first stage of the proposed system, magnetic tile images were subjected to binary classification as "defective" and "free". In this stage, performance evaluation was carried out using different machine learning (ML) classifiers after ViT-based feature extraction. Experimental results showed that the highest performance was obtained by the ViT+ Multilayer Perceptron (MLP) hybrid model with 97.77% accuracy, 97.30% F₁-score, and 97.30% Area Under Curve (AUC) value. In the second stage, the goal was to recognize different defect types (cracks, blowhole, fray, uneven, and breaks) in a multi-class manner on the images identified as defective. The most successful results in this stage were again obtained with the ViT+MLP model, achieving 94.27% accuracy, 92.43% F-score, and 96.42% AUC. These findings demonstrate that the ViT-based global context learning capability provides high success in both detecting defects and discriminating defect types in magnetic tile surfaces. In conclusion, the proposed two-stage ViT-based hybrid approach is considered to offer an effective solution for fast, reliable, and low-cost automated surface inspection systems in magnetic tile production.