Scholarlink

<p>Ensuring the freshness of meat is vital for food safety, consumer trust, and waste reduction. Traditional chemical and microbiological tests (e.g., TVB-N, pH, and microbial counts) are reliable but destructive and time-consuming, which limits their practicality for real-time monitoring. In this work, we formulate the problem as RGB-based visual freshness classification into three categories (Fresh, Half-Fresh, and Spoiled) using appearance cues (color and texture) and visual labels, rather than direct estimation of biochemical freshness indices. This study introduces DW–SPPFNet, a lightweight deep learning framework designed for rapid, non-destructive classification of meat freshness from RGB images. The model integrates depthwise separable convolution (DW) to minimize redundant computation and Spatial Pyramid Pooling Feature-lite (SPPF-lite) to enhance multi-scale spatial representation. This combination achieves a superior trade-off between accuracy and efficiency, enabling edge-level deployment. Trained and tested on a dataset of 10,372 labeled pork images, DW–SPPFNet achieved 98.31% test accuracy, 98.32% macro-F1, and Cohen’s <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\kappa \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>κ</mi> </math></EquationSource> </InlineEquation> of 0.9747, surpassing state-of-the-art lightweight backbones such as MobileNetV4-S and EfficientViT. The model operates at 2.72 ms per image with a minimal computational footprint (0.213 GFLOPs, 1.63M parameters), allowing real-time inference on resource-limited devices.</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

Fast and Accurate Meat Freshness Classification Using Depthwise Separable Convolution and SPPF

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Khanh-Duy Cao-Phan,
  • Hoang-Khang Dang,
  • Huyen-Tran Tran-Quynh,
  • Minh-Phuc Lam-Doan,
  • Huyen-Trang Luu,
  • Hong-Quan Bui

摘要

Ensuring the freshness of meat is vital for food safety, consumer trust, and waste reduction. Traditional chemical and microbiological tests (e.g., TVB-N, pH, and microbial counts) are reliable but destructive and time-consuming, which limits their practicality for real-time monitoring. In this work, we formulate the problem as RGB-based visual freshness classification into three categories (Fresh, Half-Fresh, and Spoiled) using appearance cues (color and texture) and visual labels, rather than direct estimation of biochemical freshness indices. This study introduces DW–SPPFNet, a lightweight deep learning framework designed for rapid, non-destructive classification of meat freshness from RGB images. The model integrates depthwise separable convolution (DW) to minimize redundant computation and Spatial Pyramid Pooling Feature-lite (SPPF-lite) to enhance multi-scale spatial representation. This combination achieves a superior trade-off between accuracy and efficiency, enabling edge-level deployment. Trained and tested on a dataset of 10,372 labeled pork images, DW–SPPFNet achieved 98.31% test accuracy, 98.32% macro-F1, and Cohen’s \(\kappa \) κ of 0.9747, surpassing state-of-the-art lightweight backbones such as MobileNetV4-S and EfficientViT. The model operates at 2.72 ms per image with a minimal computational footprint (0.213 GFLOPs, 1.63M parameters), allowing real-time inference on resource-limited devices.