Scholarlink

Abstract <p>As a high-value economic crop, the cultivation scale of greenhouse tomatoes has been continuously expanding, driving the development of automated harvesting technology. As the foundation of automated harvesting, tomato fruit detection algorithms directly determine the efficiency of the harvesting process. Although You Only Look Once (YOLO) series models have been widely applied in fruit detection due to their outstanding performance, their effectiveness in complex greenhouse environments is often affected by challenges such as fruit occlusion, overlap, background similarity, and small targets. To address these challenges, we propose TMS-YOLO (Texture-aware, Multidimensional Collaborative, and Small-target Enhanced YOLO), an enhanced YOLO11n-based framework that incorporates three key innovations to enhance detection capability: a TAC3k2 module with a Texture-Aware Attention (TAA) module to tackle occlusion and background confusion, an MCC3k2 module with a Multidimensional Collaborative Attention (MCA) mechanism to filter out irrelevant features, and a dedicated small-target detection layer to capture small-sized tomatoes. Comprehensive experiments on the Greenhouse tomatoes dataset confirm that our proposed model achieves 81.1 ± 0.5 %, 79.1 ± 0.3 %, 85.8 ± 0.1 %, 48.6 ± 0.3 % and 80.0 ± 0.2 % in terms of precision, recall, mAP50, mAP50-95, and F1 score, respectively, showing significant improvements compared to the baseline model. The model exhibits particularly strong robustness in challenging scenarios involving occlusion, small targets, and cluttered backgrounds, demonstrating its potential for reliable tomato automated harvesting.</p> Graphic abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

TMS-YOLO: a texture-aware attention-enhanced approach for tomato detection in greenhouses

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Jie Ji,
  • Zhengtong Dong,
  • Yining Feng,
  • Chuanming Song,
  • Jianjun Wang,
  • Zumin Wang

摘要

Abstract

As a high-value economic crop, the cultivation scale of greenhouse tomatoes has been continuously expanding, driving the development of automated harvesting technology. As the foundation of automated harvesting, tomato fruit detection algorithms directly determine the efficiency of the harvesting process. Although You Only Look Once (YOLO) series models have been widely applied in fruit detection due to their outstanding performance, their effectiveness in complex greenhouse environments is often affected by challenges such as fruit occlusion, overlap, background similarity, and small targets. To address these challenges, we propose TMS-YOLO (Texture-aware, Multidimensional Collaborative, and Small-target Enhanced YOLO), an enhanced YOLO11n-based framework that incorporates three key innovations to enhance detection capability: a TAC3k2 module with a Texture-Aware Attention (TAA) module to tackle occlusion and background confusion, an MCC3k2 module with a Multidimensional Collaborative Attention (MCA) mechanism to filter out irrelevant features, and a dedicated small-target detection layer to capture small-sized tomatoes. Comprehensive experiments on the Greenhouse tomatoes dataset confirm that our proposed model achieves 81.1 ± 0.5 %, 79.1 ± 0.3 %, 85.8 ± 0.1 %, 48.6 ± 0.3 % and 80.0 ± 0.2 % in terms of precision, recall, mAP50, mAP50-95, and F1 score, respectively, showing significant improvements compared to the baseline model. The model exhibits particularly strong robustness in challenging scenarios involving occlusion, small targets, and cluttered backgrounds, demonstrating its potential for reliable tomato automated harvesting.

Graphic abstract