Scholarlink

<p>Robotic fruit harvesting requires accurate and real-time visual perception under severe occlusion, illumination variation, dense fruit clustering, and resource-constrained edge deployment. Existing lightweight detectors often suffer from unstable low-level feature extraction, limited long-range contextual reasoning, and localization jitter under blurred or partially occluded boundaries. To address these challenges, we propose AppleDet, a lightweight and robust apple detection framework built upon YOLOv11n. AppleDet integrates a reparameterized RepStem module for stable early visual encoding, a C3k2-MambaOut module for long-range contextual feature modeling, and an efficient detection head with distributional bounding-box regression for uncertainty-aware localization. On a natural orchard dataset collected from Fuji apple orchards in Aksu, Xinjiang, AppleDet achieved an mAP@0.5 of 95.81%, improving YOLOv11n by 8.2 percentage points while reducing parameters from 5.22M to 4.06M and GFLOPs from 6.48 to 4.78. Deployment experiments on the RK3588 edge platform further demonstrated the feasibility of real-time orchard perception under close-range, long-shot, backlight, and complex-background conditions. The source code, trained weights, annotation protocol, and benchmark split will be released at <a href="https://github.com/xfl-521/AppleDet">https://github.com/xfl-521/AppleDet</a>.</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

AppleDet: lightweight and robust apple detection in natural orchard environments via reparameterized and state space visual modeling

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Ruimin Qi,
  • Guodong Zhang,
  • Xuchao Yang

摘要

Robotic fruit harvesting requires accurate and real-time visual perception under severe occlusion, illumination variation, dense fruit clustering, and resource-constrained edge deployment. Existing lightweight detectors often suffer from unstable low-level feature extraction, limited long-range contextual reasoning, and localization jitter under blurred or partially occluded boundaries. To address these challenges, we propose AppleDet, a lightweight and robust apple detection framework built upon YOLOv11n. AppleDet integrates a reparameterized RepStem module for stable early visual encoding, a C3k2-MambaOut module for long-range contextual feature modeling, and an efficient detection head with distributional bounding-box regression for uncertainty-aware localization. On a natural orchard dataset collected from Fuji apple orchards in Aksu, Xinjiang, AppleDet achieved an mAP@0.5 of 95.81%, improving YOLOv11n by 8.2 percentage points while reducing parameters from 5.22M to 4.06M and GFLOPs from 6.48 to 4.78. Deployment experiments on the RK3588 edge platform further demonstrated the feasibility of real-time orchard perception under close-range, long-shot, backlight, and complex-background conditions. The source code, trained weights, annotation protocol, and benchmark split will be released at https://github.com/xfl-521/AppleDet.