轻量级模型同时识别番茄成熟度与采摘点，助力温室机器人精准采收。
YOLO26-RipeLoc Lite: A lightweight architecture for tomato ripeness detection and picking point localization in greenhouse robotic harvesting
- 融合深度可分离卷积的轻量特征金字塔，高效处理多尺度信息
- 成熟度感知注意力模块提升颜色纹理区分能力，准确率达95.2%
- 集成抓取点回归分支，直接输出机器人抓取位置，适合部署于小型农机
在温室番茄生产中，自动化采收需精确检测成熟番茄、分类成熟度并精确定位采摘点以指导机械臂。本文提出YOLO26-RipeLoc Lite，基于YOLO26的轻量级深度学习架构，实现番茄检测、成熟度分类与中心点定位的联合任务。模型引入三项改进：(1) 轻量级特征金字塔网络（LFPN）采用深度可分离卷积实现高效多尺度融合；(2) 成熟度感知注意力模块（RAAM）结合双池化与可学习成熟度偏置向量，增强颜色-纹理判别能力；(3) 紧凑检测头（CDH）共享卷积并集成中心点回归分支，支持直接抓取规划。在阿布扎比SILAL温室采集的自定义数据集上评估，共含1,500张图像、6,227个实例（其中3,566个成熟，2,661个未成熟），模型在[email protected]达92.9%（成熟类95.2%，未成熟类90.6%），精度最高为95.2%，仅需2.38M参数。训练后进行30%的BatchNorm剪枝，参数降至约1.8M，精度损失可忽略。消融实验表明，温室感知的HSV增强带来最大提升（+2.02 pp mAP@50），冻结主干网络达到最高精度（93.8%），三阶段渐进解冻获得最优定位效果（mAP@50:95为64.6%）。相比YOLOv8n/s、YOLO11n/s、YOLO12n/s和YOLO26s，该模型在精度与效率平衡上表现更优：较YOLO12n精度高2.9个百分点，参数减少7.0%，且集成中心点回归功能，直接支持机械臂末端执行器引导。
原文摘要 · Abstract (English)
In greenhouse tomato production, automated harvesting requires accurate detection of ripe tomatoes, ripeness classification, and precise picking-point localization for robotic end-effectors. This paper proposes YOLO26-RipeLoc Lite, a lightweight deep learning architecture based on YOLO26 for simultaneous detection, ripeness classification, and center-point localization of greenhouse tomatoes. The model introduces three modifications: (1) a Lightweight Feature Pyramid Network (LFPN) with depthwise separable convolutions for efficient multi-scale fusion, (2) a Ripeness-Aware Attention Module (RAAM) with dual pooling and a learnable ripeness bias vector for enhanced color-texture discrimination, and (3) a Compact Detection Head (CDH) with shared convolutions and an integrated center-point regression branch for direct grasp planning. The model is evaluated on a custom dataset of 1,500 images with 6,227 instances (3,566 ripe, 2,661 unripe) from the SILAL greenhouse, Abu Dhabi, UAE. YOLO26-RipeLoc Lite achieves [email protected] of 92.9% (95.2% ripe, 90.6% unripe) with the highest precision (95.2%) among all evaluated architectures using only 2.38M parameters. Post-training BatchNorm pruning at 30% reduces parameters to ~1.8M with negligible accuracy loss. Ablation studies confirm that greenhouse-aware HSV augmentation provides the largest improvement (+2.02 pp mAP@50), backbone freezing achieves peak precision (93.8%), and 3-phase progressive unfreezing yields the best localization quality (mAP@50:95 of 64.6%). Comparisons with YOLOv8n/s, YOLO11n/s, YOLO12n/s, and YOLO26s confirm superior accuracy-efficiency: 2.9 pp higher precision than YOLO12n with 7.0% fewer parameters and integrated center-point localization for robotic end-effector guidance.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。