Lemon Fruit Detection and Instance Segmentation in an Orchard Environment Using Mask R-CNN and YOLOv5

This paper presents the use of models that can identify, localize, and estimate the number of lemons from any commercial lemon orchard. Object detection, separation, and estimation were carried out by using Mask R-CNN and YOLOv5 machine learning models. These models were trained on images collected from different commercial lemon orchards around Bangladesh. According to the Mask R-CNN and YOLOv5 models, the FPR and FNR were 12.7% and 15.8%, 7.2%, and 9.4%, respectively. For the Mask R-CNN and YOLOv5 models, F1 scores were obtained as 85.6% and 91.5% respectively. This estimation is comparable to other fruit detection models available in literature. This study can benefit commercial lemon producers in estimating the production of lemons from the viewpoint of agricultural requirements such as soil management, fertilizer use, and crop nutrition. The main challenges of this project were collecting images from the lemon orchards and annotating images to create the datasets. This study can help the farmers and agricultural corporations to accurately estimate the production of lemons.