Scholarlink

<p>To address the challenges of complex backgrounds, dense small-object detection, and hardware limitations in coal coking operation (CCO) tasks, this paper proposes CCO-DETR, a lightweight, multi-scale object detection model. The model incorporates the faster efficient multi-scale attention block (FEB) module in the backbone network, achieving a balance between lightweight design and improved performance. In the feature interaction stage, the efficient additive attention (EAA) mechanism replaces the traditional multi-head self-attention (MHSA) mechanism, enhancing computational efficiency without sacrificing accuracy and improving deployment capability on mobile devices. Additionally, a novel feature fusion network is proposed, which integrates a shallow small-object feature (SSOF) layer with scale-sequential feature fusion (SSFF) and triple feature encoding (TFE) modules, improving small-object detection capabilities. To address the lack of datasets in this domain, a dedicated CCO dataset is created, and both comparative and ablation experiments are conducted. Results demonstrate that CCO-DETR outperforms existing object detection models. Compared to the baseline model RT-DETR-r18, it achieves a 4.81% increase in mean average precision at intersection over union threshold 0.5 (mAP@0.5), a reduction of 7.86&#xa0;M parameter. Notably, the AP for small-object detection, such as barrels, improved by 8.85%. This research provides an effective solution for CCO detection tasks, supporting safe and unmanned production.</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

CCO-DETR: a lightweight multi-scale object detection model for coal coking operations

  • 期刊论文

  • 发表日期：

    2025年4月2日

查看全文
  • Longteng Yi,
  • Jingjing Wu,
  • Rongyang Wu,
  • Haiyang Li

摘要

To address the challenges of complex backgrounds, dense small-object detection, and hardware limitations in coal coking operation (CCO) tasks, this paper proposes CCO-DETR, a lightweight, multi-scale object detection model. The model incorporates the faster efficient multi-scale attention block (FEB) module in the backbone network, achieving a balance between lightweight design and improved performance. In the feature interaction stage, the efficient additive attention (EAA) mechanism replaces the traditional multi-head self-attention (MHSA) mechanism, enhancing computational efficiency without sacrificing accuracy and improving deployment capability on mobile devices. Additionally, a novel feature fusion network is proposed, which integrates a shallow small-object feature (SSOF) layer with scale-sequential feature fusion (SSFF) and triple feature encoding (TFE) modules, improving small-object detection capabilities. To address the lack of datasets in this domain, a dedicated CCO dataset is created, and both comparative and ablation experiments are conducted. Results demonstrate that CCO-DETR outperforms existing object detection models. Compared to the baseline model RT-DETR-r18, it achieves a 4.81% increase in mean average precision at intersection over union threshold 0.5 (mAP@0.5), a reduction of 7.86 M parameter. Notably, the AP for small-object detection, such as barrels, improved by 8.85%. This research provides an effective solution for CCO detection tasks, supporting safe and unmanned production.