<![CDATA[Sweetness is one of the main determinants of pineapple quality, yet its conventional measurement through the TSS/TA ratio requires cutting the fruit open, making it unsuitable for non-destructive, large-scale sorting. Prior work by our group has explored non-destructive classification approaches, including a multi-view attention-based fusion model. As part of the iterative model development process toward that solution, this paper reports and analyzes the performance of a simpler baseline: a single-view YOLO object detection model trained directly to localize and classify pineapples into three sweetness categories-Asam (sour), Manis Ideal (ideal), and Sangat Manis (very sweet)-from a single RGB image per fruit. The model was trained for 50 epochs and evaluated using standard object detection metrics. The baseline achieved an overall mAP@0.5 of 0.555 and mAP@0.5:0.95 of 0.460, with the best F1-score of 0.58 reached at a confidence threshold of 0.183. Per-class analysis shows that the Asam category was the easiest to detect (mAP@0.5 = 0.695), while Manis Ideal (0.505) and Sangat Manis (0.465) were considerably weaker. Confusion matrix analysis at the default confidence threshold reveals that only 32-64% of ground-truth instances per class were correctly classified, notably lower than the recall trend suggested during training, and that the Sangat Manis class-the smallest in the dataset-was most frequently confused with its visual neighbor, Manis Ideal. These findings indicate that a single viewpoint, without any imbalance handling, is not yet sufficient to reliably separate boundary categories, providing empirical grounds for the multi-view and attention-based refinements explored in the continuation of this research.]]>
Copyrights © 2026
