Detection of Oil Palm Fresh Fruit Bunches (FFBS) with Computer Vision Models

Abstract

Fresh Fruit Bunches (FFBs) are an important agricultural asset for countries such as Malaysia as they are used to obtain oil palm which contributes to a large portion of the countries' GDP. However, the current FFBs harvesting process requires a significant amount of manual labor and may cause injuries or damage to the FFBs which must be avoided to ensure the oil produced does not get affected. Therefore, an automatic harvesting system needs to be developed to help avoid such issues. An important part of the harvesting system would be a vision system for the detection and localization of the fruits and involves the usage of computer vision models. In this paper, a comparison is performed between four different models: Support Vector Machine (SVM) which uses the Histogram of Gradients (HOG), Faster R-CNN, YOLOv4 and YOLOv5. The models were tested by the metrics of mean average precision (mAP), precision, recall and F1 score and their detection speed. The results showed that the SVM + HOG model had the worst performance with a mAP of 89% and the YOLOv4 model had the best performance with a mAP of 98.5%. However, the YOLOv5 model provided the best results in terms of speed with a detection speed of 11.4 ms and its mAP was 96.4% making it the best model for use in a harvesting system for the detection of FFBs as it provides both speed and accuracy.