Authors

Abstract

Accurate perception of the maturity of individual truss tomato fruits and in-situ counting in unstructured protected horticulture scenarios are critical prerequisites for driving selective automated harvesting, yield prediction and improving the level of refined management. Affected by factors such as the small fruit size and the complexity of the natural growth environment, vision-based maturity detection still faces considerable challenges. This paper proposes an improved method for maturity detection and counting of individual truss tomato fruits based on YOLO11n. Under the principle of maintaining a lightweight design, a fine-grained feature stream based on the P2 layer is developed, and the DySample operator is integrated to optimize the quality of feature fusion. Combined with the Selective Feature Refinement Module (EMA) and a four-head detector with full-scale coverage, the proposed method aims to maximize the model's representation capability in capturing small long-distance targets and in dense occluded environments. A dataset of individual truss tomato fruits with three maturity labels (immature, turning, ripe) is constructed, and systematic comparative experiments are conducted on the original YOLO11n and the improved model. The experimental results show that the improved YOLO11n-DRE outperforms the original YOLO11n model in maturity detection accuracy, with P, R and mAP@0.5 increased by 0.65%, 0.63% and 1.07% respectively, and the model parameters reduced by 8.5%. This method demonstrates excellent detection performance, and provides a reference model and technical prerequisite for the maturity detection and yield estimation of individual truss tomato fruits.

Abstract in English

在非结构化设施园艺场景中，实现串番茄单果成熟度的精准感知与原位计数，是驱动选择性自动化采收、产量预测及提升精细化管理水平的关键前提。受果实尺度小以及自然生长环境复杂等因素影响，基于视觉的成熟度检测仍面临较大挑战。本文提出了一种基于YOLO11n改进的串番茄单果成熟度检测与计数方法。在维持轻量化设计原则下，开发基于P2层的细粒度特征流，集成DySample算子以优化特征融合质量。配合选择性特征精炼模块（EMA）与全尺度覆盖的四头检测器，旨在最大化模型在远距离小目标捕捉及密集遮挡环境下的表达能力。构建了包含三类成熟度标签（未熟、转色、成熟）的串番茄单果数据集，对原版YOLO11n与改进模型进行了系统对比实验。实验结果表明，改进后的YOLO11n-DRE在成熟度检测精度方面优于原始YOLO11n模型，在P、R、mAP@0.5等分别提升了0.65%、0.63%、1.07%，模型参数减少了8.5%。该方法展现出卓越的检测性能，为串番茄单果成熟度的检测和产量估算提供了参考模型和技术前提。