关键词: 深度学习, 草莓检测, YOLOv8, 注意力机制, 数据增强

Abstract: In order to achieve fast and accurate detection of strawberries in complex greenhouse environments, a multi‑stage strawberry detection algorithm based on an improved YOLOv8 was proposed. Firstly, the strawberry dataset, collected in greenhouse environments, was initially annotated by using LabelImg. Subsequently, in order to address issues such as the small size of strawberries and the complexity of the environment, a BiFormer dynamic attention mechanism was integrated into the backbone network. This integration allowed for more flexible computational allocation and feature perception, focusing the network model more on small object detection and enhancing its fruit detection capabilities in complex environments. Finally, a VanilaNet module was introduced in the Neck component to reduce the computational complexity of the model and further improve its strawberry recognition accuracy. Experimental results demonstrated that the improved YOLOv8, in comparison to the traditional YOLOv8, increased the mAP by 4.6%, reaching 93.8%. The improved YOLOv8 not only has higher detection accuracy, but also performs well  in small target detection, which can provide support for the subsequent real‑time small target detection of picking robots.

Key words: deep learning, strawberry detection, YOLOv8, attention mechanism, data enhancement