Machine vision for detection of defective oil palm fresh fruit bunches using YOLO Algorithm

Abstract In the last couple of years, machine vision with object detection methods has been applied in agriculture, especially using deep learning with the You Only Look Once (YOLO) algorithm. It has gained attention for rapid in-line sorting machine vision, where efficiency in time and labor cost is crucial. This study designed and built a real-time machine vision system and a YOLOv4 tiny algorithm to identify defective oil palm fresh fruit bunches (FFBs) moving in a roller conveyor. Oil palm FFBs are the source of crude palm oil (CPO), which is the export commodity for countries of Indonesia and Malaysia. Sorting FFB is a daily process when an oil palm FFB transporting truck arrives at an oil palm refinery mill. This process is usually done manually and traditionally by two experienced workers who sort for unqualified and defective FFBs, such as long stalk, rotten, empty, and thorny or spiky FFBs. We developed a two-class model to detect normal and defective FFBs. The model has obtained the average precision (AP) of 98% for normal and 96% for defective oil palm FFBs. The accuracy validation using the confusion matrix resulted in an mAP of 99.7%. The results exhibited the potential of machine vision and YOLOv4 tiny to detect defective FFBs.