Scholarlink

<p>In response to the growing issue of meat adulteration involving the illegal mixing of pork with beef in Indonesia—an act with profound implications for halal integrity—this study introduces FMEAT-Net, a lightweight convolutional neural network (CNN) architecture designed for efficient, image-based classification of beef, pork, and adulterated meat. Leveraging mixed-precision training (MPT) and data augmentation techniques, FMEAT-Net demonstrates superior performance over the conventional VGG16 model, achieving a test accuracy of 90.83% and a weighted F1 score of 91%, while reducing computational demands. The model was rigorously tested on a curated dataset enhanced with domain-specific augmentations, such as adjusted lighting tones to simulate real-world variability. A comparative analysis of FP16, FP32, and FP64 precision formats also reveals the trade-offs between computational efficiency and sensitivity to subtle adulteration patterns. The findings underscore FMEAT-Net's potential as a scalable and cost-effective tool for meat authentication, offering a meaningful contribution toward enhancing food transparency and supporting halal certification systems in resource-constrained environments. The source code is publicly available at <a href="https://github.com/fikrisyahid/fmeat-net">https://github.com/fikrisyahid/fmeat-net</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

FMEAT-Net: an efficient CNN architecture for image-based identification of beef, pork, and adulterated meat

  • 期刊论文

  • 发表日期：

    2025年12月19日

查看全文
  • Muhammad Fikri Syahid,
  • Toto Haryanto,
  • Hendra Rahmawan

摘要

In response to the growing issue of meat adulteration involving the illegal mixing of pork with beef in Indonesia—an act with profound implications for halal integrity—this study introduces FMEAT-Net, a lightweight convolutional neural network (CNN) architecture designed for efficient, image-based classification of beef, pork, and adulterated meat. Leveraging mixed-precision training (MPT) and data augmentation techniques, FMEAT-Net demonstrates superior performance over the conventional VGG16 model, achieving a test accuracy of 90.83% and a weighted F1 score of 91%, while reducing computational demands. The model was rigorously tested on a curated dataset enhanced with domain-specific augmentations, such as adjusted lighting tones to simulate real-world variability. A comparative analysis of FP16, FP32, and FP64 precision formats also reveals the trade-offs between computational efficiency and sensitivity to subtle adulteration patterns. The findings underscore FMEAT-Net's potential as a scalable and cost-effective tool for meat authentication, offering a meaningful contribution toward enhancing food transparency and supporting halal certification systems in resource-constrained environments. The source code is publicly available at https://github.com/fikrisyahid/fmeat-net.