Scholarlink

<p>As a core research direction in real-time object detection, lightweight models aim to achieve efficient and accurate object recognition and localization with limited computing resources. This paper proposes a lightweight 2MCD-YOLO11 gesture recognition model, aiming to provide high-precision, low-latency inference for resource-constrained edge devices and enable an efficient, real-time intelligent vision screening system. By adopting MobileNetV4 as its backbone instead of the original architecture, the proposed model enhances dynamic feature extraction capability and improves computational efficiency. The neck network is based on Cross-scale Channel Feature Fusion (CCFF), with depthwise separable convolution (DWConv) integrated to optimize the detection head’s path, reducing complexity, parameter count, and inference latency. Minimum Point Distance Intersection over Union (MPDIoU) enhances detection stability via boundary point distance constraints and a direction-aware mechanism. Experiments on the self-built UDLRU-DAT gesture dataset demonstrate that the 2MCD-YOLO11 algorithm achieves a detection accuracy of 87.3% with 0.98&#xa0;M parameters and an inference speed of 159.7 FPS. Deployed on mobile devices, it retains considerable detection accuracy and real-time performance of 61 FPS, verifying the technical feasibility of lightweight models in edge computing environments.</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

2MCD-YOLO11: a lightweight algorithm based on cross-scale feature fusion for gesture recognition

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • Yueqiang Feng,
  • Jiayi Qian,
  • Wenxing Zuo,
  • Yutong Tan,
  • Lulu Li,
  • Zhengwei Shui

摘要

As a core research direction in real-time object detection, lightweight models aim to achieve efficient and accurate object recognition and localization with limited computing resources. This paper proposes a lightweight 2MCD-YOLO11 gesture recognition model, aiming to provide high-precision, low-latency inference for resource-constrained edge devices and enable an efficient, real-time intelligent vision screening system. By adopting MobileNetV4 as its backbone instead of the original architecture, the proposed model enhances dynamic feature extraction capability and improves computational efficiency. The neck network is based on Cross-scale Channel Feature Fusion (CCFF), with depthwise separable convolution (DWConv) integrated to optimize the detection head’s path, reducing complexity, parameter count, and inference latency. Minimum Point Distance Intersection over Union (MPDIoU) enhances detection stability via boundary point distance constraints and a direction-aware mechanism. Experiments on the self-built UDLRU-DAT gesture dataset demonstrate that the 2MCD-YOLO11 algorithm achieves a detection accuracy of 87.3% with 0.98 M parameters and an inference speed of 159.7 FPS. Deployed on mobile devices, it retains considerable detection accuracy and real-time performance of 61 FPS, verifying the technical feasibility of lightweight models in edge computing environments.