<![CDATA[Tempeh is a traditional Indonesian fermented food whose quality is influenced by fermentation and environmental conditions. Quality assessment is still commonly performed manually, leading to subjectivity and inconsistency. This study compares three modern object detection models—YOLOv5sM, YOLOv8, and YOLOv11—for digital image–based tempeh quality recognition. A dataset of 1,000 images (500 good and 500 defective) was collected using a Logitech C270 camera under controlled lighting conditions. YOLOv5sM was trained with data augmentation (Mosaic, flip, rotation), while YOLOv8 and YOLOv11 were trained without augmentation to isolate architectural differences. All models were trained for 100 epochs using identical hyperparameters and evaluated on a 10% test set. Results show that YOLOv11 achieved the highest accuracy (98%), outperforming YOLOv8 (94%) and YOLOv5sM (88%). Although mAP@0.5 reached 99.5% across models, stricter evaluation using mAP@0.5:0.95 revealed performance differences (96.2%, 96.9%, and 97.0%, respectively). The superior performance of YOLOv11 is attributed to its C3K2 and C2PSA modules, which enhance fine-grained feature extraction and localization precision. These findings indicate that YOLOv11 is the most suitable architecture for automated tempeh quality inspection]]>
Copyrights © 2026
