Few-Shot Deep Learning for Granite Texture Classification: Adapting to New Classes with Minimal Data
30th IEEE Jubilee International Conference on Intelligent Engineering Systems, INES 2026, Budapest, Macaristan, 2 - 04 Temmuz 2026, ss.687-690, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/ines69513.2026.11661076
- Basıldığı Şehir: Budapest
- Basıldığı Ülke: Macaristan
- Sayfa Sayıları: ss.687-690
- Anahtar Kelimeler: convolutional neural networks, deep learning, few-shot learning, granite, incremental learning, industrial inspection, texture classification
- Bilecik Şeyh Edebali Üniversitesi Adresli: Evet
Özet
Automated classification of granite tiles is important for industrial quality control and construction material management. Conventional deep learning approaches require large labeled datasets, limiting their adaptability when new granite types are introduced or when the availability of samples is constrained. This paper presents a few-shot deep learning framework for granite texture classification, enabling accurate recognition with minimal training samples per class. A pre-trained convolutional neural network and metric-based prototype learning strategies were used to evaluate the model's ability to generalize to unseen granite classes using only a handful of labeled images. Experiments were conducted using a six-class granite dataset, GraniTR, including both full-size images and 224×224 patches extracted from the originals, to assess the effect of input resolution on the classification performance. The results demonstrate that the few-shot model achieves competitive accuracy across one-shot to fifteen-shot scenarios, supports the incremental introduction of new granite types without retraining the backbone, and maintains stable performance across base and newly added classes. These findings highlight the potential of few-shot learning for scalable and flexible granite classification systems in industrial environments.