Scholarlink

<p>Cluttered backgrounds, variable shooting angles, and fluctuating lighting frequently induce missed or false detections of filaments by picking robots, especially for small and imbalanced targets in safflower fields. To address these inherent limitations, we propose SAF-YOLO, a novel detector specifically optimized for robotic safflower harvesting. The architecture integrates three complementary innovations to ensure robust perception in unstructured environments. Firstly, a Visual State Space Model (VSSM)-based VSS-SPPF module is integrated into the backbone to capture global spatial context, which effectively separates filaments from complex background clutter. Secondly, an Asymptotic Feature Pyramid Network (AFPN) adaptively merges multi-scale features to mitigate the scale discrepancies induced by variable shooting angles. Finally, a Super-Resolution Self-Supervised (SRSS) auxiliary branch regularizes backbone learning via reconstruction tasks, driving the model to learn illumination-invariant features that resist lighting fluctuations; this branch operates only during training and is removed at inference to preserve efficiency. Experimental results demonstrate SAF-YOLO achieves 90.1% Precision, 85.9% Recall, and 93.3% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(mAP_{50}\)</EquationSource> </InlineEquation> on the safflower dataset, and 94.1% <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(mAP_{50}\)</EquationSource> </InlineEquation> on the GWHD benchmark. This outperforms the popular YOLO variants, including YOLOv5/v7/v8/v11 (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(mAP_{50}\)</EquationSource> </InlineEquation> +4.1%-8.0%), and mainstream small-object detectors (e.g., DETR, CFINet, CFPT, and InSPyReNet; <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(mAP_{50}\)</EquationSource> </InlineEquation> +7.9%-14.5%). With only 3.0M parameters and 8.3 GFLOPs, the model achieves a real-time inference speed of 38.5 FPS with TensorRT acceleration on the Jetson Nano platform, providing a practical approach to autonomous agricultural harvesting.</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

SAF-YOLO: Super-resolution augmented detection model with visual state space enhancement for safflower filament picking

  • 期刊论文

  • 发表日期：

    2026年4月27日

查看全文
  • Duan Mengyu,
  • Wang Xiaorong,
  • Qiu Linwei,
  • Li Menghao,
  • Chen Jinrong,
  • He Liang

摘要

Cluttered backgrounds, variable shooting angles, and fluctuating lighting frequently induce missed or false detections of filaments by picking robots, especially for small and imbalanced targets in safflower fields. To address these inherent limitations, we propose SAF-YOLO, a novel detector specifically optimized for robotic safflower harvesting. The architecture integrates three complementary innovations to ensure robust perception in unstructured environments. Firstly, a Visual State Space Model (VSSM)-based VSS-SPPF module is integrated into the backbone to capture global spatial context, which effectively separates filaments from complex background clutter. Secondly, an Asymptotic Feature Pyramid Network (AFPN) adaptively merges multi-scale features to mitigate the scale discrepancies induced by variable shooting angles. Finally, a Super-Resolution Self-Supervised (SRSS) auxiliary branch regularizes backbone learning via reconstruction tasks, driving the model to learn illumination-invariant features that resist lighting fluctuations; this branch operates only during training and is removed at inference to preserve efficiency. Experimental results demonstrate SAF-YOLO achieves 90.1% Precision, 85.9% Recall, and 93.3% \(mAP_{50}\) on the safflower dataset, and 94.1% \(mAP_{50}\) on the GWHD benchmark. This outperforms the popular YOLO variants, including YOLOv5/v7/v8/v11 ( \(mAP_{50}\) +4.1%-8.0%), and mainstream small-object detectors (e.g., DETR, CFINet, CFPT, and InSPyReNet; \(mAP_{50}\) +7.9%-14.5%). With only 3.0M parameters and 8.3 GFLOPs, the model achieves a real-time inference speed of 38.5 FPS with TensorRT acceleration on the Jetson Nano platform, providing a practical approach to autonomous agricultural harvesting.