Akademik Veri Yönetim Sistemi
15th International Congress of the Innovative Agricultural Technologies, IAT 2025, Antalya, Türkiye, 15 - 19 Ekim 2025, cilt.805 LNCE, ss.186-194, (Tam Metin Bildiri)
Verifying the exact drying route (microwave power level and whether ultrasound (US) pretreatment was applied) in finished jujube (Ziziphus jujuba) products is essential for traceability and quality assurance; however, this process is typically labor-intensive and operator-dependent. This work reframes end-of-line process verification as a visual, multi-factor classification problem and evaluates deep learning models for discriminating eight drying combinations (100, 200, 300, 600 W; with/without US). A balanced dataset of 1,760 images (220 per class) of dried jujube slices acquired under controlled imaging is split into training (1,232 images), test (352 images), and validation (176 images) sets in a 7:2:1 ratio, preprocessed to 224 × 224 px and modeled with four transfer-learning CNNs (ResNet-18, DenseNet-121, ConvNeXt-Tiny, ConvNeXt-Base) in the Fastai/PyTorch stack. Performance was assessed using accuracy, precision, recall, F1-score, and AUC. A Flask-based REST API was implemented for real-time inference with CPU-optimized settings (model.eval, torch.no_grad) to facilitate practical deployment. DenseNet-121 achieved the highest overall accuracy (~99%) with consistently high precision/recall across classes; confusion-matrix and ROC/AUC analyses confirmed robust separation among the eight drying classes, including US-assisted conditions. The proposed pipeline enables rapid, non-destructive discrimination of multi-factor drying conditions in jujube, thereby advancing traceability and quality control with a lightweight model and a standardized imaging protocol suitable for near-real-time use. The approach enables end-of-line verification without additional chemical assays and is readily extensible to other products and drying recipes.