Scholarlink

<p>Efficient object detection is essential for recognition algorithms deployed on mobile devices with limited computational resources. To address this challenge, we propose a lightweight object detection algorithm incorporating a Dynamic Task-Aligned Feature Pyramid Network (DyTA-FPN). Our approach introduces three key innovations: (1) a Ghost-Shuffle Hierarchical Block (GSHB) introduces a main-branch and auxiliary-branch design within a hierarchical graph structure, which preserves both local details and global semantics while effectively reducing redundant computations; (2) A Dynamic Decomposition and Modulation Module (DyDeMod) separates and optimizes the classification and regression tasks through task decomposition mechanism, avoiding optimization conflicts and effectively improving detection accuracy; and (3) a Multiscale Hybrid Fusion (MSHF) module adopts a segmentation-and-cross-fusion strategy on feature maps, which effectively reduces information redundancy and computational overhead. Evaluated on the VOC dataset, our method surpasses YOLOv8n, achieving a 37.5% reduction in parameters and 5.20&#xa0;ms of inference time, while improving precision by 2.1% and mAP50:95 by 1.4%. These results highlight the effectiveness of our lightweight algorithm for resource-constrained 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

Efficient object detection with dynamic task-aligned feature pyramids for resource-limited mobile devices

  • 期刊论文

  • 发表日期：

    2025年7月25日

查看全文
  • Yuqing Chen,
  • Shiwen Xie,
  • Qingxin Li,
  • Huosheng Hu

摘要

Efficient object detection is essential for recognition algorithms deployed on mobile devices with limited computational resources. To address this challenge, we propose a lightweight object detection algorithm incorporating a Dynamic Task-Aligned Feature Pyramid Network (DyTA-FPN). Our approach introduces three key innovations: (1) a Ghost-Shuffle Hierarchical Block (GSHB) introduces a main-branch and auxiliary-branch design within a hierarchical graph structure, which preserves both local details and global semantics while effectively reducing redundant computations; (2) A Dynamic Decomposition and Modulation Module (DyDeMod) separates and optimizes the classification and regression tasks through task decomposition mechanism, avoiding optimization conflicts and effectively improving detection accuracy; and (3) a Multiscale Hybrid Fusion (MSHF) module adopts a segmentation-and-cross-fusion strategy on feature maps, which effectively reduces information redundancy and computational overhead. Evaluated on the VOC dataset, our method surpasses YOLOv8n, achieving a 37.5% reduction in parameters and 5.20 ms of inference time, while improving precision by 2.1% and mAP50:95 by 1.4%. These results highlight the effectiveness of our lightweight algorithm for resource-constrained environments.