A Lightweight Industrial Image Anomaly Detection Using Multi-Scale Feature Fusion Çok Ölçekli Öznitelik Füzyon Yöntemiyle Kaynak-Etkin Endüstriyel Görüntü Anomali Tespiti
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636368
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: computer vision, deep learning, industrial anomaly detection, one-class classification, unsupervised learning
- Orta Doğu Teknik Üniversitesi Adresli: Evet
Özet
This study proposes a lightweight and unsupervised method for industrial image anomaly detection using multi scale feature fusion. The method learns a one class embedding from multi level features extracted by a pretrained ResNet-18 and is trained only with normal samples. The architecture is designed for applications requiring low computational cost, low memory usage, and high inference speed. During training, a Deep SVDD based loss is used. The method achieves AUROC scores of 93.5% and 83.2% on the MVTec AD and VisA datasets, respectively. In addition, the proposed model provides a balanced tradeoff between accuracy and efficiency with 15 million parameters, 0.2 GB memory usage, and 26.2 FPS inference speed.