Ripeness recognition of citrus fruit by combining attention mechanism with YOLOv5

Accurate and real-time detection of fruit ripeness under an orchard environment is a critical technique for many smart orchard applications, such as automatic harvesting. This study is dedicated to developing a deep learning-based citrus ripeness recognition model. Firstly, YOLOv5, YOLOv7, YOLOX, and YOLOv8 were investigated respectively by training with our selt-built citrus fruit data set, in which the ripeness of citrus was divided into immature, semi-mature and mature. Experiments show that the recognition model trained with YOLOv5 has the highest recognition rate, and its mAP@0.5 and mAP@0.7 reach 94.73% and 93.73%, respectively. Secondly, five attention mechanisms (SE, EVC, CA, SimAM, and CBAM) were added to the YOLOv5 model to further improve the recognition effect. The experimental results showed that YOLOv5s combined with the CBAM attention mechanism could get the best recognition performance, with 96.07% on mAP@0.5 and 95.40% on mAP@0.7.