基于轻量化SLD-YOLO v8n的温室番茄成熟度检测方法

Tomato Maturity Detection Method in Greenhouse Based on Lightweight SLD-YOLO v8n

摘要:

面向复杂温室环境下番茄成熟度检测中存在的光照波动、枝叶遮挡以及果实密集分布等挑战,本文提出一种轻量化番茄成熟度检测方法。以YOLO v8n为基线模型,构建C2f-SHAB特征提取模块,改善局部细节与全局语义信息的联合建模效果;引入DySample上采样模块,优化多尺度特征融合过程中的高分辨率信息恢复;设计轻量化共享细节感知检测头LSDA-DHead,提升检测端对复杂边界与遮挡区域的特征表达能力;同时结合Focaler-DIoU损失函数优化边界框回归过程。基于自建番茄成熟度数据集开展试验,将番茄划分为未熟期、转色期、转熟期和成熟期4个等级。试验结果表明,改进后SLD-YOLO v8n模型,与YOLO v8n相比,mAP@0.5提高7.11个百分点,参数量、浮点运算量和模型内存占用量分别下降30.6%、25.9%和15.1%。在保证轻量化的同时,有效提升了复杂环境下番茄成熟度检测的精度与鲁棒性,可为智能采摘机器人和边缘端视觉感知系统提供技术支撑。

Abstract:

Aiming to address the challenges of illumination fluctuation, branch and leaf occlusion, and dense fruit distribution in complex greenhouse environments for tomato maturity detection, a lightweight tomato maturity detection method was proposed. Taking YOLO v8n as the baseline model, a C2f-SHAB feature extraction module was constructed to improve the joint modeling of local details and global semantic information. The DySample upsampling module was introduced to optimize high-resolution information recovery during multi-scale feature fusion. A lightweight shared detail-aware detection head, LSDA-DHead, was designed to enhance the feature representation capability for complex boundaries and occluded regions at the detection end. Furthermore, the Focaler-DIoU loss function was incorporated to optimize the bounding box regression process. Experiments were conducted on a self-built tomato maturity dataset, in which tomatoes were classified into four stages: immature, turning-color, turning-ripe, and mature. Experimental results showed that the improved SLD-YOLO v8n model achieved a 7.11 percentage point increase in mAP@0.5 compared with the baseline YOLO v8n, while the number of parameters, floating-point operations, and model memory footprint were reduced by 30.6%, 25.9%, and 15.1%, respectively. While ensuring lightweightness, the proposed method effectively enhanced the accuracy and robustness of tomato maturity detection under complex conditions, providing technical support for intelligent harvesting robots and edge-side visual perception systems.