Authors

Abstract

Improper selection of tomato harvest maturity often leads to uneven ripening, overripening, rotting, or softening damage during transportation, resulting in significant distribution losses. To obtain objective criteria for tomato maturity grading and quantitative statistics, this paper proposes an improved lightweight tomato maturity detection and counting model, CMLE-YOLO, using common tomatoes as the research subject. Building upon YOLOv11, this model incorporates Cross-Phase Fast Mixed Attention (CFMA) modules in both the trunk and neck networks to enhance spatial feature interaction and global context modeling capabilities. Concurrently, a lightweight Quality-Aware Detection Head (LQAD) is designed to improve consistency between classification confidence and localization accuracy while reducing parameter redundancy. Experimental results demonstrate that CMLE-YOLO achieves outstanding performance in detecting immature, semi-mature, and mature tomatoes, with an mAP@50 of 0.8508, significantly outperforming mainstream models such as YOLOv5, YOLOv6, and YOLOv8. while maintaining lightweight characteristics with only 2.13 million parameters and computational complexity of 5.2 GFLOPs, lower than most comparable models. This study confirms the method achieves a favorable balance between detection accuracy and computational efficiency, providing effective technical support for real-time tomato ripeness monitoring and yield assessment.

Abstract in Chinese

番茄采收成熟度选择不当易导致运输过程中出现成熟不均、过熟腐烂或软化损伤等问题，造成显著流通损耗。为获取番茄成熟度分级与数量统计的客观依据，本文以普通番茄为研究对象，提出一种改进的轻量化番茄成熟度检测与计数模型 CMLE-YOLO。该模型以 YOLOv11 为基础，在主干与颈部网络中引入跨阶段快速混合注意力模块（CFMA），增强空间特征交互与全局上下文建模能力；同时设计轻量化质量感知检测头（LQAD），提升分类置信度与定位精度的一致性并降低参数冗余。实验结果表明，CMLE-YOLO 在未成熟、半成熟和成熟三类番茄检测任务中表现优异，其 mAP@50 达到 0.8508，显著优于 YOLOv5、YOLOv6、YOLOv8 等主流对比模型；同时模型保持轻量级特性，参数量仅 2.13 M、计算量为 5.2 GFLOPs，计算复杂度低于多数同类模型。研究证实，该方法实现了检测精度与计算效率的良好平衡，可为番茄成熟度实时监测与产量评估提供有效的技术支持。