GraniTR: A Novel Dataset and Transfer Learning Approach for Multi-Class Granite Tile Classification
IEEE Access, cilt.14, ss.98407-98416, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3705540
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.98407-98416
- Anahtar Kelimeler: CNN, data augmentation, deep learning, granite classification, sustainability
- Erciyes Üniversitesi Adresli: Evet
Özet
Effective granite tile classification is essential for ensuring quality and consistency in various industrial applications. Althoıugh existing public datasets mainly contain polished stone images collected under controlled conditions, they often lack the variability and realism required for industrial deployment. This study introduces a novel GraniTR dataset featuring six distinct classes of granite stones, all sourced from the Bilecik region of Türkiye, a major hub in the natural stone industry. The dataset consists of 934 images captured in both indoor and outdoor environments using multiple camera devices, incorporating variations in lighting, resolution, and surface appearance. By reflecting real-world imaging conditions and authentic regional granite types, the dataset presents a challenging and practical benchmark for computer-vision models. We evaluated this dataset using several deep learning (DL) models, including a basic Convolutional Neural Network (CNN), VGG16, Inception-v3, and MobileNetV2. Our experiments were conducted in three phases: baseline performance assessment, common augmentation techniques, and a sliding window patch extraction (SWPE) method tailored to the GraniTR dataset. The results revealed that VGG16 achieved 99.21% accuracy with the SWPE, demonstrating its effectiveness with augmented data. MobileNetV2, which is notable for its efficiency, maintained a strong baseline performance with 98.32% accuracy, making it suitable for applications requiring speed and real-time processing. These findings underscore the importance of selecting appropriate models based on specific application needs and highlight the value of the GraniTR dataset as a benchmark for robust granite stone classification under real-world industrial conditions.