In precision agriculture, on-tree fruit detection is a solution for automating many tasks which governs yield estimation, harvesting and quality monitoring. Among all tropical fruits, mangoes are the most widely consumed fruit and is a member of the cashew family. Mangoes have been cultivated in South and Southeast Asia for thousands of years. There are several cultivars of mango, which vary in size, shape, sweetness, skin colour and flesh colour. Mango is India’s most important commercial fruit crop, accounting for over 54% of all mangoes produced globally. This paper is a study conducted to detect the fruit using a simple non-destructive YOLOv7 deep architecture, a highly accurate object detection method with less error rate. A robust dataset, MangoNet of mango images which are annotated using labelImg tool, is used in the study. The detection of on-tree mangoes will be beneficial for the yield estimation of fruits. The YOLOv7 deep model is employed, incorporating transfer learning to enhance model efficiency and accuracy. The performance of the model is evaluated using the metrics mean average precision (mAP) and Intersection over Union (IoU). 99.5% accuracy is met in mAP@0.5. The results show the feasibility of on-tree mango detection with high precision and recall for automated agricultural systems. Python programming language with pyTorch library is used for the transfer learning. This work highlights the potential of YOLO approach to optimize the mango farming practices and contribute to the smart agriculture technologies.
Read full abstract