Scholarlink

<p>To address the challenge of balancing computational complexity with detection accuracy in real-time sign language recognition under complex backgrounds, this study presents a multi-module collaborative optimization framework based on YOLOv5. The proposed method introduces a triple-attention enhancement mechanism that jointly recalibrates spatial and channel features through parameter-free operations, effectively suppressing background interference while sharpening the network’s focus on critical sign language regions. The original C3 module is redesigned into a lightweight C3_DWR variant by integrating deformable weight re-parameterization with depthwise separable convolutions and residual connections, optimizing feature representation while reducing model complexity. To preserve multi-scale gesture semantics, a Global Feature Pyramid Network (GFPN) is developed to enable cross-layer feature fusion for capturing hierarchical patterns at multiple granularities. Evaluations on public sign language benchmarks demonstrate that the optimized architecture achieves mAP50 and mAP50-95 scores of 95.2% and 67.2%, respectively, corresponding to improvements of 4.6 and 5.9 percentage points over the baseline YOLOv5. With only 2.5 million parameters and 5.7 GFLOPs, the proposed framework reduces computational costs by approximately 42% compared to similar architectures while outperforming larger models such as YOLOv5m in accuracy. These results confirm that dynamic weight re-parameterization combined with cross-layer fusion effectively reconciles efficiency with precision, offering a practical solution for low-power, real-time sign language detection on edge computing platforms.</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

High-precision YOLOv5 object detection technology based on multi-module collaborative optimization

  • 期刊论文

  • 发表日期：

    2026年2月1日

查看全文
  • Yong Yang,
  • Tianci Wan,
  • Menglu Zhang,
  • Lingling Li

摘要

To address the challenge of balancing computational complexity with detection accuracy in real-time sign language recognition under complex backgrounds, this study presents a multi-module collaborative optimization framework based on YOLOv5. The proposed method introduces a triple-attention enhancement mechanism that jointly recalibrates spatial and channel features through parameter-free operations, effectively suppressing background interference while sharpening the network’s focus on critical sign language regions. The original C3 module is redesigned into a lightweight C3_DWR variant by integrating deformable weight re-parameterization with depthwise separable convolutions and residual connections, optimizing feature representation while reducing model complexity. To preserve multi-scale gesture semantics, a Global Feature Pyramid Network (GFPN) is developed to enable cross-layer feature fusion for capturing hierarchical patterns at multiple granularities. Evaluations on public sign language benchmarks demonstrate that the optimized architecture achieves mAP50 and mAP50-95 scores of 95.2% and 67.2%, respectively, corresponding to improvements of 4.6 and 5.9 percentage points over the baseline YOLOv5. With only 2.5 million parameters and 5.7 GFLOPs, the proposed framework reduces computational costs by approximately 42% compared to similar architectures while outperforming larger models such as YOLOv5m in accuracy. These results confirm that dynamic weight re-parameterization combined with cross-layer fusion effectively reconciles efficiency with precision, offering a practical solution for low-power, real-time sign language detection on edge computing platforms.