Scholarlink

<p>In the era of pre-packaged foods, accurate mold detection in baked goods is essential for food safety and public health. To address the lack of specialized datasets, this study constructs the MNBF dataset, comprising 14,936 images of Chinese and Western bakery products, to support efficient and non-destructive surface-based detection. However, variations in color, texture, and complex production environments remain major challenges. To overcome these, we propose a novel detection model, the suppression-of-interference multi-scale adaptive optimized YOLOv11. This model incorporates a feature-enhancement layer that reduces background interference, a multi-scale attention mechanism that improves learning across different object sizes, and an adaptive fusion module that strengthens the integration of spatial information. These improvements collectively enhance detection performance when targets are partially occluded or vary widely in scale. Experimental results show that optimized model outperformed five mainstream detectors, achieving 82.3% precision on the MNBF dataset, with a 3.6% improvement over the baseline. It also maintained an average detection time of approximately 31 ms per image across 3,734 test samples. The findings confirm the model’s effectiveness in detecting mold contamination across diverse bakery categories, providing technical support for intelligent quality inspection while reducing spoilage-related losses, lowering foodborne risks, and enhancing consumer trust.</p> Graphical abstract <p></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

Deep learning-based non-destructive mold detection for baked food quality control

  • 期刊论文

  • 发表日期：

    2026年3月11日

查看全文
  • Linzhi Li,
  • Yan Xiang,
  • Yuxuan Deng,
  • Shuiwang Li,
  • Ruxin Guo,
  • Xiaolan Xie

摘要

In the era of pre-packaged foods, accurate mold detection in baked goods is essential for food safety and public health. To address the lack of specialized datasets, this study constructs the MNBF dataset, comprising 14,936 images of Chinese and Western bakery products, to support efficient and non-destructive surface-based detection. However, variations in color, texture, and complex production environments remain major challenges. To overcome these, we propose a novel detection model, the suppression-of-interference multi-scale adaptive optimized YOLOv11. This model incorporates a feature-enhancement layer that reduces background interference, a multi-scale attention mechanism that improves learning across different object sizes, and an adaptive fusion module that strengthens the integration of spatial information. These improvements collectively enhance detection performance when targets are partially occluded or vary widely in scale. Experimental results show that optimized model outperformed five mainstream detectors, achieving 82.3% precision on the MNBF dataset, with a 3.6% improvement over the baseline. It also maintained an average detection time of approximately 31 ms per image across 3,734 test samples. The findings confirm the model’s effectiveness in detecting mold contamination across diverse bakery categories, providing technical support for intelligent quality inspection while reducing spoilage-related losses, lowering foodborne risks, and enhancing consumer trust.

Graphical abstract