Scholarlink

<p>Correctly identifying beef cuts is crucial for consumer trust, food safety, culinary standards, economic value, and government compliance. However, this task is a fine-grained computer vision challenge because different cuts often exhibit only subtle visual differences, making them far harder to distinguish than typical object recognition problems. Traditional non-destructive techniques such as ultrasonic imaging and Raman spectroscopy can provide detailed structural information, but they are limited to controlled laboratory environments, require expensive instrumentation, and are therefore impractical for retail or consumer-level use. To bridge this gap, we propose an affordable deep learning–based approach for non-destructive beef cut classification using standard RGB images, which can be easily captured with consumer-grade devices such as smartphones. We collected a real-world dataset of 5585 images across five beef cut categories (plate, round, short loin, sirloin, and tenderloin) under uncontrolled lighting, varied camera angles, and complex backgrounds, ensuring conditions representative of retail and supply-chain environments. Our pipeline consists of two stages: (1) YOLOv11 is used to identify regions of interest (ROI), and (2) a variety of deep learning models are used to categorize these regions. A hybrid model that integrated pre-trained Global Context Vision Transformer (GCViT-B) features with a support vector machine (SVM) outperformed end-to-end CNN and ViT architectures, achieving an accuracy of 89%. The transformer-based model alone was 86% accurate. These findings show that low-cost RGB imaging combined with&#xa0;deep learning provides an affordable and precise alternative to existing methods for fine-grained beef cut classification in dynamic, consumer-level scenarios.</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

Two-stage hybrid deep learning approaches for non-destructive beef cut classification

  • 期刊论文

  • 发表日期：

    2025年12月24日

查看全文
  • Sait Alp

摘要

Correctly identifying beef cuts is crucial for consumer trust, food safety, culinary standards, economic value, and government compliance. However, this task is a fine-grained computer vision challenge because different cuts often exhibit only subtle visual differences, making them far harder to distinguish than typical object recognition problems. Traditional non-destructive techniques such as ultrasonic imaging and Raman spectroscopy can provide detailed structural information, but they are limited to controlled laboratory environments, require expensive instrumentation, and are therefore impractical for retail or consumer-level use. To bridge this gap, we propose an affordable deep learning–based approach for non-destructive beef cut classification using standard RGB images, which can be easily captured with consumer-grade devices such as smartphones. We collected a real-world dataset of 5585 images across five beef cut categories (plate, round, short loin, sirloin, and tenderloin) under uncontrolled lighting, varied camera angles, and complex backgrounds, ensuring conditions representative of retail and supply-chain environments. Our pipeline consists of two stages: (1) YOLOv11 is used to identify regions of interest (ROI), and (2) a variety of deep learning models are used to categorize these regions. A hybrid model that integrated pre-trained Global Context Vision Transformer (GCViT-B) features with a support vector machine (SVM) outperformed end-to-end CNN and ViT architectures, achieving an accuracy of 89%. The transformer-based model alone was 86% accurate. These findings show that low-cost RGB imaging combined with deep learning provides an affordable and precise alternative to existing methods for fine-grained beef cut classification in dynamic, consumer-level scenarios.