Deep learning-based Mango Fruit Detection and Counting

Mango is an agricultural product that is of important economic significance in Thailand, requiring yield estimation in order to plan for the harvest stage effectively. This research presents Mango fruit detection and counting on a canopy tree using a deep learning-based object detection model, Mask R-CNN and Cascade R-CNN. To acquire an image of a mango tree in the current circumstances, partial occlusion of mangoes usually occurs. We enhance the model learning by combining the Random Erasing method for image augmentation to improve performance in dealing with partial occlusion issues. This paper compares the performance of the different backbone layer for feature extraction. We utilize the most optimal backbone, RestNext-101-32x8d-FPN is an essential component of the model. The Mask R-CNN with Random Erasing outperforms the Cascade R-CNN achieving the highest performance with a precision score of 0.922, a recall score of 0.786, an F1-score of 0.849, a mAP@50 score of 0.883, and counting performance MAE score of 1.106 fruits per image. The best model can be deployed in real-world applications.