Scholarlink

<p>During production, edible oils may acquire microscopic impurities from raw materials or processing environments, such as organic matter and sediment, necessitating effective detection. Since oil is usually more viscous and less fluid than liquids with low permeability such as water and wine, conventional impurity detection methods are inapplicable to edible oils. This paper proposes DM-FDA, a dual-modal feature detection algorithm for microscopic impurities. Specifically, an event camera firstly captures event images and visible light images of oil products, respectively. Following this, event images are preprocessed to remove noises. The DM-FDA module subsequently performs the detection. It uses BiFPN to fuse features from dual-modal images on the backbone. Then the SPPCSPC module is used to connect the feature maps of different scales. IF-MPDIoU loss function which is inspired by Inner IoU and Focal IoU is used for border regression. To evaluate the algorithm’s performance, we spiked five edible oils with simulated impurities, including raw material fragments, metal fragments, animal hairs, and variously sized tin beads. The algorithm was evaluated on a comprehensive dataset of 14,520 images. It demonstrated a perfect 100% detection accuracy for hairs, while achieving accuracies of 89.40% for metal fragments, 76.10% for raw material fragments, and 69.30% for tin beads. Compared to the original dual-modal YOLOv8 network, the proposed model demonstrated substantial accuracy improvements across all impurity types, with gains of 25.90% for hairs, 25.00% for raw material fragments, 15.30% for metal fragments, and 8.00% for tin beads. The minimum detectable impurity size was 0.2&#xa0;mm.</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

DM-FDA: a dual-modal feature detection algorithm for microscopic impurities

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Ming Li,
  • Xu Yan,
  • Ying Zhao,
  • Jiandun Li,
  • Huiyu Wang

摘要

During production, edible oils may acquire microscopic impurities from raw materials or processing environments, such as organic matter and sediment, necessitating effective detection. Since oil is usually more viscous and less fluid than liquids with low permeability such as water and wine, conventional impurity detection methods are inapplicable to edible oils. This paper proposes DM-FDA, a dual-modal feature detection algorithm for microscopic impurities. Specifically, an event camera firstly captures event images and visible light images of oil products, respectively. Following this, event images are preprocessed to remove noises. The DM-FDA module subsequently performs the detection. It uses BiFPN to fuse features from dual-modal images on the backbone. Then the SPPCSPC module is used to connect the feature maps of different scales. IF-MPDIoU loss function which is inspired by Inner IoU and Focal IoU is used for border regression. To evaluate the algorithm’s performance, we spiked five edible oils with simulated impurities, including raw material fragments, metal fragments, animal hairs, and variously sized tin beads. The algorithm was evaluated on a comprehensive dataset of 14,520 images. It demonstrated a perfect 100% detection accuracy for hairs, while achieving accuracies of 89.40% for metal fragments, 76.10% for raw material fragments, and 69.30% for tin beads. Compared to the original dual-modal YOLOv8 network, the proposed model demonstrated substantial accuracy improvements across all impurity types, with gains of 25.90% for hairs, 25.00% for raw material fragments, 15.30% for metal fragments, and 8.00% for tin beads. The minimum detectable impurity size was 0.2 mm.