A semi-supervised method for detecting female cucumber flowers in greenhouses based on unmanned aerial vehicle images

Abstract

Flower quantity and the flowering time often serve as reliable indicators of the growth and nutritional status of the cucumber plants. In particular, the number of female flowers reflects the yield potential of the cucumber crops. Therefore, timely identification and accurate counting of female flowers are important measures in cucumber cultivation. In order to satisfy the practical detection requirements, especially in the scenario of greenhouses with a high planting density, we have presented an efficient and accurate cucumber flower recognition framework MYTS (Mamba-YOLO based Teacher-Student model) in this research. A mini UAV (unmanned aerial vehicle) was used to enhance the convenience and efficiency of obtaining cucumber images. After data augmentation, we employed a three-stage deep learning pipeline to train our model, i.e. supervised learning, self-distillation, and semi-supervised learning. Experimental results demonstrate that the female flower detection of MYTS achieves precision, recall, and mAP values of 90.6%, 89.7%, and 95.2%, respectively. The mAP value outperforms leading models such as YOLO and Mamba-YOLO by 5.7% and 5.1%. The ablation experiments reveal that multiple attention mechanisms significantly improve the model performance during the stage 1 by 3.5% in mAP value. Additionally, compared to the original supervised model, the three-stage pipeline enhances the mAP values of all investigated models by 1.4% to 6.2%. Specifically, the MYTS model shows a 1.6% improvement over its supervised counterpart. In the future, further exploration should be conducted to apply the MYTS model in cucumber yield estimation.