Loose Oil Palm Fruitlet Detection Using YOLOv8 Nano Model By Layer Freezing Training for Better Accuracy with Small Dataset

Palm oil is a vital commodity for Malaysia, contributing 30% of global vegetable oil production. Beyond fresh fruit bunch (FFB) harvesting, loose fruitlets play a crucial role in palm oil production. Automating their detection is essential for sustainable agricultural robotics in addressing ongoing labour shortages. Traditional image processing methods rely on manually engineered features, which may not accurately represent fruitlets and it is very time-consuming to implement. This study evaluates the use of layer freezing of YOLOv8 nano for loose fruitlet detection, using mean average precision (mAP50-95) as the performance metric. Results show a 0.251 difference in mAP50-95 between large and small datasets, indicating potential overfitting with smaller datasets. To mitigate this, pre-processing and dynamic augmentation were applied to a 280-image dataset. Models were trained using the AdamW optimizer, with the YOLOv8-Nano model achieving the highest mAP50-95 of 0.75 on the smaller dataset. A training-validation gap of approximately 0.1 for DFL loss and 0.2 for box loss over 100 epochs suggests reduced overfitting.