|
| 结合MFNet与ShuffleMFNet的食品质量安全多区域智能检验方法研究 |
| Research on Intelligent Inspection Method for Multi-Region Food Quality and Safety Combining MFNet and ShuffleMFNet |
| 投稿时间：2026-01-22 修订日期：2026-05-26 |
| DOI： |
| 中文关键词: 食品质量安全 多区域检测 MFNet 鲁棒特征增强 |
| 英文关键词: Food quality and safety Multi-region detection MFNet Robust feature enhancement |
| 基金项目:安徽省高校自然科学研究项目,中式传统肉制品加工品质控制及保鲜技术的研究 |
|
| 摘要点击次数: 94 |
| 全文下载次数: 0 |
| 中文摘要: |
| 复杂多食品场景下的食品质量安全智能检验对自动化与鲁棒性提出了更高要求,传统及现有食品图像检测方法难以适应多食品区域共存、遮挡重叠及复杂工况条件。因此,研究提出一种结合多分支特征网络(Multi-branch Feature Network,MFNet)与通道混洗多分支特征网络(Shuffle Multi-branch Feature Network,ShuffleMFNet)的食品质量安全多区域智能检验方法,通过构建多区域食品定位与判别特征建模网络实现精准区域切分,并引入卷积调制、局部窗口注意力与通道混洗联合的鲁棒特征增强机制完成区域级质量安全判别。结果表明,研究方法在决策阈值0.4时准确率达到97.42%,高于视觉Transformer基础模型(Vision Transformer,ViT-Base)的93.15%和高效卷积神经网络B4模型(EfficientNet-B4)的91.24%,曲线下面积峰值达到0.98,领先ViT-Base的0.92和EfficientNet-B4的0.90。由此证明,研究方法在多区域定位精度、复杂场景鲁棒性及质量安全判别可靠性方面均具有优势,能够有效支撑复杂用餐与生产环境中的食品质量安全智能检验。 |
| 英文摘要: |
| Intelligent inspection of food quality and safety in complex multi-food scenarios places higher demands on automation and robustness. Traditional and existing food image detection methods are difficult to adapt to the coexistence of multiple food regions, occlusion and overlap, and complex working conditions. Therefore, this study proposes an intelligent multi-region food quality and safety inspection method that combines a multi-branch feature network (MFNet) and a shuffle multi-branch feature network (ShuffleMFNet). By constructing a multi-region food localization and discrimination feature modeling network, accurate region segmentation is achieved. Furthermore, a robust feature enhancement mechanism combining convolutional modulation, local window attention, and shuffle is introduced to complete regional quality and safety discrimination. The results show that the research method achieves an accuracy of 97.42% at a decision threshold of 0.4, which is higher than the 93.15% of the Vision Transformer (ViT-Base) model and the 91.24% of the EfficientNet-B4 convolutional neural network model. The peak area under the curve reaches 0.98, surpassing ViT-Base"s 0.92 and EfficientNet-B4"s 0.90. This demonstrates that the research method has advantages in multi-region positioning accuracy, robustness in complex scenarios, and reliability in quality and safety discrimination, and can effectively support intelligent inspection of food quality and safety in complex dining and production environments. |
|
View Fulltext
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
