Authors

Abstract

The detection of tomatoes for automatic picking is challenging due to the dense distribution of fruit and severe occlusions. To address this, a dataset is developed using tomato images captured in a greenhouse environment, and an enhanced model for tomato fruit maturity detection based on YOLOv8n is proposed, which incorporates the EMA attention mechanism and the C2f-Faster module for multi-scale feature fusion. These additions not only improve detection accuracy but also enhance detection speed, thereby boosting the model's robustness and generalization ability. Experimental results demonstrate that the proposed ECF-YOLOv8n model achieves detection accuracies of 93.8%, 94.7%, 92.5% and 94.1% for immature, nearly mature, ripe tomatoes and mean average precision in a greenhouse setting, respectively. The model's size is 4.7 MB, with GFLOPs of 6.5G. Compared to advanced models like RT-DETR, YOLOv5 and YOLOv7, the ECF-YOLOv8n model outperforms them in both detection accuracy and speed. This work provides valuable insights for the research, development and optimization of tomato picking robots.

Abstract in Chinese

针对目前番茄自动化采摘目标检测中因果实密集、遮挡严重等导致目标检测难度大的问题，本研究基于温室大棚环境下的番茄图像，构建了数据集，提出了一种基于YOLOv8n的番茄果实成熟度检测的改进模型，并添加引入了EMA注意力机制和C2f-Faster模块，以实现多尺度特征融合，在保证检测精度较高的情况下，有效提高了番茄果实检测速度，从而进一步提高了模型的鲁棒性和泛化能力。试验结果表明：提出的ECF-YOLOv8n模型对温室大棚环境下未成熟、将要成熟、成熟番茄检测精度和均值平均精度分别为：93.8%、94.7%、92.5%和94.1%，模型大小为4.7 MB，GFLOPs为6.5G，与RT-DETR、YOLOv5、YOLOv7等先进模型的比较，该模型实现了较高的检测精度和更快的检测速度，本研究可为番茄采摘机器人的研发和优化提供重要参考。