Scholarlink

<p>Detecting and classifying tiny objects in cluttered scenes is vital for both industrial automation and ecological monitoring. However, the weak textural cues and low contrast between targets and backgrounds pose significant challenges for accurate recognition. To address these issues, we propose a robust deep detection framework that integrates a <i>Dynamic Multi-scale Cross-Attention (DMCA)</i> module, a <i>Star Feature Fusion Module (SFFM)</i>, and a shape-sensitive <i>Intersection over Union (IoU)</i> loss. These components collaboratively enhance fine-grained feature representation and improve boundary localization of small targets. Experimental results demonstrate that our method achieves mean average precision scores of 70.9% and 92.1% on two challenging datasets, surpassing recent state-of-the-art detectors by a remarkable margin. The code is available at <a href="https://github.com/Lou510/InsectDet">https://github.com/Lou510/InsectDet</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

Enhanced feature fusion for precise small object detection in complex environments

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Chunzhi Li,
  • Weizhen Lou,
  • XiaoHua Chen,
  • Yanfen Fan,
  • Jiaxin Li

摘要

Detecting and classifying tiny objects in cluttered scenes is vital for both industrial automation and ecological monitoring. However, the weak textural cues and low contrast between targets and backgrounds pose significant challenges for accurate recognition. To address these issues, we propose a robust deep detection framework that integrates a Dynamic Multi-scale Cross-Attention (DMCA) module, a Star Feature Fusion Module (SFFM), and a shape-sensitive Intersection over Union (IoU) loss. These components collaboratively enhance fine-grained feature representation and improve boundary localization of small targets. Experimental results demonstrate that our method achieves mean average precision scores of 70.9% and 92.1% on two challenging datasets, surpassing recent state-of-the-art detectors by a remarkable margin. The code is available at https://github.com/Lou510/InsectDet.