MangoYOLO5: A Fast and Compact YOLOv5 Model for Mango Detection

Abstract

Detection of fruits in orchards is crucial for agricultural applications like automatic estimation and mapping of yield. The state-of-the-art approaches for this task are based on hand-crafted features and therefore are liable to variations in a real orchard environment. However, the current deep learning-based one-stage object detection methods like the YOLO provide excellent detection accuracy at the cost of increased computational complexity. So this paper presents an improved, fast, and compact YOLOv5s model named as MangoYOLO5 for detecting mangoes in the images of open mango orchards given by the MangoNet-Semantic dataset. The proposed MangoYOLO5 has adopted a few improvements over YOLOv5s. Firstly, the feedback convolutional layer is removed from the BottleneckCSP module of the original YOLOv5s model, reducing the convolutional layers by 11. Secondly, two convolutional layers, one from the focus module and another just after the focus module, are removed to reduce the overall weight of the architecture. It is observed from precision, recall, and mAP metrics of the experimental results that MangoYOLO5 detection performance is $3.0\%$ better than the YOLOv5s, addressing several factors such as occlusion, distance, and lighting variations. In addition, the realized lighter model requires $66.67\%$ less training time as compared to original YOLOv5s, which can significantly affect its real-time implementations.