Authors

Abstract

To address the low speed and limited accuracy of tomato maturity detection in complex greenhouse environments characterized by dense distribution, overlap, and occlusion, this study proposes YOLOv11n-SDS, an improved algorithm based on YOLOv11n. The key enhancements include: (1) a Spatial Pyramid Depthwise Separable Convolution (SPD-Conv) module; (2) the integration of a Deformable Large-Kernel Attention (DLKA) mechanism into the backbone C3K2 module; and (3) a Semantic and Detail Injection (SDI) module replacing the Concat operation in the neck network. These improvements enhance the detection of small and low-resolution targets, as well as occluded fruits under challenging lighting and background conditions. Experimental results show that YOLOv11n-SDS improves mAP@0.5 by 2.4%, recall by 1.3%, and precision by 3.2%, while maintaining a low computational cost of 9.1 GFLOPs. Compared with existing models, including RT-DETR, Faster R-CNN, SSD, and other YOLO variants, the proposed model achieves a superior balance between accuracy, efficiency, and practical applicability. Furthermore, the model was deployed and validated on a mobile robotic platform in a greenhouse environment, enabling real-time tomato maturity detection and 3D target localization. These results demonstrate its strong potential for practical applications, such as ripeness monitoring and harvesting-oriented perception.

Abstract in Chinese

针对复杂果园环境下番茄存在密集分布、重叠以及叶片遮挡等情况，使用现有的目标检测算法进行番茄成熟度检测存在速度慢、识别准确率低等问题，本研究提出一种基于YOLOv11n的改进算法YOLOv11n-SDS。其核心改进包括：（1）引入空间金字塔深度可分离卷积（SPD-Conv）模块；（2）在主干网络C3K2模块中加入可变形大核注意力（DLKA）机制；（3）采用语义细节注入（SDI）模块替代颈部网络中的Concat连接。这些改进有效提升了对低分辨率图像和小尺寸番茄的识别能力，显著增强了在复杂光照与背景干扰下对遮挡果实的检测性能。实验表明，YOLOv11n-SDS的mAP@0.5提升2.4%，召回率提高1.3%，精确率增长3.2%，计算量仅为9.1 GFLOPs。与RT-DETR、Faster-RCNN、SSD及YOLO系列等模型相比，本算法在检测精度、计算效率和实际应用性方面取得更优平衡，为番茄自动化采摘技术的发展提供了有效支持。