Scholarlink

Purpose <p>The objective of this study was to address the challenge of reliable backfat thickness estimation in pig carcasses, often hindered by muscle coverage, fat deposits, and bloodstains, by proposing a robust and efficient dual-stage deep learning network.</p> Methods <p>An enhanced U-Net architecture was proposed, integrated with chest cavity–guided localization to form a dual-stage network. A custom image acquisition system was developed to collect carcass images under real slaughterhouse conditions for dataset construction. The baseline U-Net was enhanced with a dual-attention mechanism (DANet) and residual connections to improve structural edge perception of key anatomical regions. For precise localization, multiple edge detection methods (Roberts, Sobel, Laplacian, and Canny) were compared to extract the chest cavity contour and identify the terminal point of the sternum, defining the optimal measurement site. A secondary segmentation stage, guided by the localized chest cavity, was then employed to focus on critical regions and enhance detection accuracy.</p> Results <p>The proposed ARU-PCSN network achieved an average IoU of 97.12% on 100 pig carcass images, with a mean measurement error of 0.61&#xa0;mm and a processing time under 0.3&#xa0;s per image.</p> Conclusion <p>The results confirm the effectiveness and practical potential of the proposed dual-stage network for intelligent and precise backfat thickness estimation in pig carcasses.</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 U-Net with Chest-Cavity–Guided Localization: A Dual-Stage Network for Accurate Backfat Thickness Estimation in Pig Carcasses

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Liukui Duan,
  • Wude Yang,
  • Liuqian Gao,
  • Wenyu Zhang,
  • Siyuan He,
  • Gaobin Li,
  • Hao Yuan,
  • Yang Liu,
  • Xu Zhang,
  • Huihui Wang

摘要

Purpose

The objective of this study was to address the challenge of reliable backfat thickness estimation in pig carcasses, often hindered by muscle coverage, fat deposits, and bloodstains, by proposing a robust and efficient dual-stage deep learning network.

Methods

An enhanced U-Net architecture was proposed, integrated with chest cavity–guided localization to form a dual-stage network. A custom image acquisition system was developed to collect carcass images under real slaughterhouse conditions for dataset construction. The baseline U-Net was enhanced with a dual-attention mechanism (DANet) and residual connections to improve structural edge perception of key anatomical regions. For precise localization, multiple edge detection methods (Roberts, Sobel, Laplacian, and Canny) were compared to extract the chest cavity contour and identify the terminal point of the sternum, defining the optimal measurement site. A secondary segmentation stage, guided by the localized chest cavity, was then employed to focus on critical regions and enhance detection accuracy.

Results

The proposed ARU-PCSN network achieved an average IoU of 97.12% on 100 pig carcass images, with a mean measurement error of 0.61 mm and a processing time under 0.3 s per image.

Conclusion

The results confirm the effectiveness and practical potential of the proposed dual-stage network for intelligent and precise backfat thickness estimation in pig carcasses.