<![CDATA[Crude palm oil (CPO) is the leading export commodity for countries such as Indonesia and Malaysia. The quality of CPO depends on the raw material and oil palm fresh fruit bunches (FFB). Various sorting and grading methods based on computer vision and machine learning have been developed to assess FFB quality automatically. However, most research has focused on fruit ripeness. In fact, empty bunches, rotten fruit, long stalks, and thorny or spiky bunches are also sorting parameters and are categorized as defective FFBs. This study aims to evaluate the performance of the YOLOv8l-Seg and YOLOv8x-Seg models in detecting and segmenting normal and defective FFBs stacked on a moving conveyor. Stacked FFBs mean there is more than one FFB in a camera field of view (FOV), which is easily found during the real-time sorting process. The dataset consists of five classes: normal, long stalks, thorny, empty, and rotten bunches. Evaluation was conducted using the mean Average Precision (mAP), accuracy, precision, recall, and F1-score metrics. The results show that YOLOv8x-Seg obtained 93% accuracy and 95% mAP. The YOLOv8l-Seg reached 92% accuracy and 93.4% mAP. Therefore, both models have the potential to be applied in real-time automated oil palm FFB sorting and grading systems.]]>
Copyrights © 2026
