An Oil Palm Loose Fruits Image Detection System using Faster R -CNN and Jetson TX2

Abstract

Palm oil is an essential world vegetable oil. The oil palm is reported to have the highest oil yield compared to other major world crops. Palm oil can be extracted from two main sources that are fresh fruit bunches and loose fruits. Although the main focus of the oil producer is to collect fruit bunches, the loose fruit has relatively higher oil content. However, the condition of scattered loose fruits at the farm wastes the farmers' time and energy in the loose fruit collecting process, and it leads to the lower back pain and spine fatigue of the workers. The conventional methods for collecting loose fruit such as manual hand-picked or a roller-type fruit collector are still less efficient. Hence, this work proposes a system for oil palm loose fruits detection using Faster R-CNN, a deep learning algorithm and NVIDIA Jetson TX2 hardware. In this study, 500 images of loose fruits were collected from an oil palm farm at Bukit Bangkong, Selangor during the harvesting process. The data were pre-processed using few techniques such as image resizing, cropping, data augmentation and data labelling. Faster R-CNN, a deep learning algorithm is used to train the model of the detection system by using 400 images from the acquired sample. The trained model was validated and tested with the remaining 100 images from the sample. The model performance showed that the loose fruit detection system is built successfully when it achieved accuracy for about 94 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> , 94 % and 91 % for intersection-of-union thresholds equal to 0.5, 0.7 and 0.9, respectively. The result showed that the developed system is able to detect oil palm loose fruits accurately and has the potential in contributing to the development of oil palm loose fruit automatic harvesting system.