Scholarlink

<p>Particle morphology is an important origin for shaping complex granular soil behavior, yet traditional measurement methods are inefficient for field-scale characterization of densely packed particles. While deep learning-based image segmentation automates morphology extraction, these approaches require large model sizes with extensive computational burdens, making them impractical for deployment on mobile devices. This study proposes a lightweight object detection model, YOLOv8-PMQ, capable of end-to-end prediction of particle shape and size distribution without image segmentation or subsequent geometric computation. By introducing a morphology parameter regression branch, the model simultaneously outputs particle bounding boxes and multiple descriptors, including roundness, sphericity, equivalent projected circular diameter, and Feret diameters. Additionally, to alleviate data scarcity, an image synthesis framework based on the OASIS generative adversarial network (GAN) was constructed to generate diverse particle images for model training. Experiments show that YOLOv8-PMQ trained with synthetic data achieves a bounding box mean average precision (b_mAP) of 54.13% and a total mean absolute error (MAEtotal) of 0.1097, outperforming training on existing data alone. Furthermore, an offline smartphone application based on this model was developed, enabling rapid on-site particle morphology analysis with better practical utility.</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

Quantifying particle morphology via lightweight object detection and generative data augmentation

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Ziyang Liu,
  • Wenju Zhu,
  • Tongming Qu,
  • Jian Gong,
  • Keyu Zhao,
  • Guang Chen,
  • Junyi Yang,
  • Yu Jiang,
  • Feng Wang,
  • Li-E. Yan

摘要

Particle morphology is an important origin for shaping complex granular soil behavior, yet traditional measurement methods are inefficient for field-scale characterization of densely packed particles. While deep learning-based image segmentation automates morphology extraction, these approaches require large model sizes with extensive computational burdens, making them impractical for deployment on mobile devices. This study proposes a lightweight object detection model, YOLOv8-PMQ, capable of end-to-end prediction of particle shape and size distribution without image segmentation or subsequent geometric computation. By introducing a morphology parameter regression branch, the model simultaneously outputs particle bounding boxes and multiple descriptors, including roundness, sphericity, equivalent projected circular diameter, and Feret diameters. Additionally, to alleviate data scarcity, an image synthesis framework based on the OASIS generative adversarial network (GAN) was constructed to generate diverse particle images for model training. Experiments show that YOLOv8-PMQ trained with synthetic data achieves a bounding box mean average precision (b_mAP) of 54.13% and a total mean absolute error (MAEtotal) of 0.1097, outperforming training on existing data alone. Furthermore, an offline smartphone application based on this model was developed, enabling rapid on-site particle morphology analysis with better practical utility.