CITRUS FLOWER, FRUIT, AND SHOOT RECOGNITION BASED ON IMPROVED YOLOv10

Abstract

In the process of agricultural intelligence, precise detection of plant organs serves as the foundation for core tasks such as crop phenotyping analysis and yield prediction. However, in complex field environments, small targets such as citrus flowers and shoots face challenges including scale variation, background interference, and dense occlusion, which severely impact detection accuracy. This study improves the YOLOv10 model by introducing the BAM (Bottleneck Attention Module) attention mechanism and GIoU (Generalized Intersection over Union) loss function, constructing a YOLOv10s-BAM-GIoU model suitable for citrus flower, fruit, and shoot recognition. The BAM attention mechanism enhances the model's feature extraction capability for small target organs under complex backgrounds through parallel channel and spatial attention branches; the GIoU loss function improves the localization accuracy of densely occluded targets by optimizing the geometric alignment between predicted and ground-truth boxes. Validation experiments were conducted on a self-constructed dataset. The experimental results show that the improved YOLOv10s achieves significant advantages in comprehensive detection accuracy, with an mAP50 of 89.1%, representing an improvement of 2.9%~9.5% over the original YOLOv10s and other comparative models. In fine-grained category detection, the model achieves mAP50 of 91.2%, 83.6%, and 92.5% for shoots, flowers, and fruits, respectively. Furthermore, while maintaining high detection accuracy, the model achieves a detection speed of 23.6 ms per frame, meeting real-time detection requirements. The research results demonstrate that the improved YOLOv10s model integrating the BAM attention mechanism and GIoU loss function achieves an optimal balance between accuracy and speed in citrus organ detection tasks, providing a preferred solution for field real-time detection systems.

Plant phenotyping relevance

柑橘の花・果実・シュートという植物器官を画像から検出する改良モデルを開発し、データセットで精度と速度を検証しており、表現型取得手法が中心である。

Code and data availability

The paper describes a self-constructed citrus flower/fruit/shoot dataset (380 images plus augmentation to 1,000) and a field validation set of 200 images, and an improved YOLOv10s-BAM-GIoU model, but no block contains any public deposit, availability statement, or authors' URL for the dataset, code, or trained model. A

No evidence-backed public reproduction asset is currently recorded.