Detection of Multi-varieties of On-tree Mangoes using MangoYOLO5

Abstract

Automated harvesting and detection of fruits are crucial for agronomic applications like estimation and mapping of yield. Earlier, fruit detection methods were mostly dependent on hand-crafted features and were prone to changes in the actual orchard environment. However, recently deep learning-based methods especially one-stage object detection techniques like YOLO has achieved a higher detection accuracy to detect different fruits including mango in on-tree orchard images. In our previous work, we proposed a lightweight YOLOv5 model named "MangoYOLO5" for the detection of mangoes, and we have achieved an accuracy of 94.4% on one variety. Now, we have created a dataset of seven varieties of on-tree mangoes, with four varieties being publicly available, and the other three varieties from a local mango orchard using a UAV. We have tried detecting these seven varieties using the MangoYOLO5 model and achieved an average accuracy of 92%. It shows that the mango detection performance is 3.4% better than the YOLOv5s, taking into several characteristics like occlusion, distance, and lighting conditions. Additionally, compared to the original YOLOv5s, the achieved lighter model requires 55.55% less training time, which can significantly affect on real-time implementations.