Authors

Abstract

During field tomato harvesting, challenges such as stem-leaf occlusion, fruit overlap, and difficulties in stem localization significantly hinder the performance of harvesting robots. To address these issues, a joint detection model for fruits and fruit stems, termed YOLO-TRS, is proposed based on the YOLO11n network. First, a novel C3k2-DS module is designed and integrated into the backbone network, enhancing the model’s ability to represent complex structural features of fruit stems. In addition, a CAA module is incorporated into the backbone to improve long-range feature modeling, thereby effectively reducing missed detections of fruits and fruit stems under occlusion conditions. The proposed model is evaluated using a self-constructed dataset. Experimental results show that YOLO-TRS achieves precision, recall, and mAP values of 89.9%, 91.5%, and 94.8%, respectively, outperforming the baseline YOLO11n model by 2.3%, 1.0%, and 2.4%. Compared with other classical object detection algorithms, YOLO-TRS demonstrates clear advantages in both detection accuracy and computational efficiency. These results confirm that the proposed model can effectively support fruit ripeness-related detection and accurately localize stem positions in complex field environments, providing a theoretical basis for intelligent agricultural harvesting.

Abstract in Chinese

田间番茄采收过程中，茎叶遮挡、果实重叠及果柄定位困难等问题严重影响采收机器人的作业性能。为解决上述挑战，本文基于YOLO11n网络提出一种果实与果柄联合检测模型YOLO-TRS。首先，提出并将C3k2-DS模块集成于骨干网络中，增强模型对果柄复杂结构特征的建模能力；此外，在骨干网络中集成CAA模块，提升模型的长距离特征建模能力，进而有效降低遮挡场景下果实与果柄的漏检率；最后，基于自建田间番茄数据集对所提模型进行验证。实验结果表明，改进后的模型精度（Precision）、召回率（Recall）和平均精度均值（mAP）分别达到89.9%、91.5%和94.8%，较YOLO11n模型分别提升2.3%、1.0%和2.4%；与其他经典目标检测算法相比，该模型在检测精度与计算效率方面均展现出显著优势。这些实验结果验证了，YOLO-TRS模型能够在复杂田间环境下有效检测果实成熟度并精准定位果柄位置，为农业智能采摘提供理论支撑。