Akademik Veri Yönetim Sistemi
2025 Innovations in Intelligent Systems and Applications Conference, ASYU 2025, Bursa, Türkiye, 10 - 12 Eylül 2025, (Tam Metin Bildiri)
This study presents a comprehensive anomaly detection framework utilizing a hybrid dataset composed of both publicly available and custom-collected images. The inclusion of curated custom data addresses existing gaps in public datasets, enhancing the representation of real-world anomaly scenarios and improving model generalizability. The final dataset comprises over 17,000 images, each meticulously annotated and exported in COCO format, with corresponding mask images generated through format conversion for segmentation-based training. Five deep learning models - YOLOv8, DeepLabV3+, U-Net, FPN, and PSPNet - were employed to detect six distinct anomaly types. While YOLOv8 provided aggregated detection metrics, the other four models generated class-specific results using three backbone architectures (ResNet, MobileNetV2, EfficientNet). In total, 72 segmentation-based models were trained and evaluated. Experimental findings reveal that model and backbone performance varies across anomaly types, with no universally superior configuration. Moreover, the results indicate a positive correlation between dataset size per anomaly and detection performance, emphasizing the significance of data volume in model accuracy.