Oil palm fresh fruit bunch ripeness classification on mobile devices using deep learning approaches

Video based oil palm ripeness detection model using deep learning

“…Previous oil palm detection studies employed augmentations such as random brightness and Gaussian blur [ 1 ]. The emphasis on geometric augmentation is also carried out in research on the classification of oil palm maturity with additional Gaussian blur augmentation [ 3 ], but applying photometric or geometric augmentation alone is insufficient because objects can be seen in real time conditions from various camera points of view and dynamic lighting conditions. The data augmentation experiment conducted in Table 5 demonstrates that data augmentation successfully overcomes the challenge of detecting oil palm maturity by combining photometric and geometric augmentation, with an indication of improvement from mAP 50 where objects can be properly identified and IoU where the model can do object localization better.…”

“…The key advancement produced in this research is in the form of a video dataset with six classes of oil palm maturity levels, which is more suitable to real-world settings than non-sequential image datasets. This research investigated multi-category video data because the current research has not tested multi-category data, therefore the findings of the existing research are unsatisfactory, and most of the previous research employs a classification model [ 2 , 3 , 5 , 6 ] that cannot recognize several objects in one picture frame. Even so, using video datasets necessitates a large amount of data in order for model detection to perform well.…”

“…Because of the high subjectivity of the manual approach, the determined level of palm oil maturity is inconsistent, resulting in frequent confrontations between the determinants of the maturity level and the customer [ 2 ]. Furthermore, the manual technique incurs high production costs [ 3 ] due to the vast number of human resource requirements and takes a long time to sort oil palm, causing the operations to become inefficient [ 4 ]. Based on the problems described, it is critical to establish an accurate detection of oil palm maturity level utilizing real-time computer vision so that the palm oil industry's production operation becomes more efficient in terms of production costs and time used to sort oil palm.…”

“…Majority of current research on the maturity level of oil palm is conducted utilizing a classification system without object localization. The method used is traditional machine learning [ 5 – 7 ] and convolutional neural network (CNN) [ [2] , [3] , [4] , 8 , 9 ]. According to existing studies, CNN outperforms traditional machine learning in image processing in the form of still images and video [ 10 ].…”

Annotated Datasets of Oil Palm Fruit Bunch Piles for Ripeness Grading Using Deep Learning

“…Several previous studies using a computer vision approach with an input image have been carried out by using the SVM method with 3 classes 18 , namely raw, under-ripe and ripe. Research with deep learning for ripeness detection has been carried out by using EfficientNet 3 with single image datasets. Real time oil palm ripeness detection using YOLOv4 with 3 classes dataset has been proposed 19 for harvesting system and another research of real time ripeness detection at harvesting process has been proposed using YOLOv3 20 .…”