EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Edge Feature and Particle Focusing Network for Holographic Particle Detection Under Diffraction Blur and Interference

Leheng Li, Chen Wang, Zhanghao Ni, Pengxiang Wang, Xiangchao Zhang

IEEE Access · 2026 · 14 · 167681-167695

DOI: 10.1109/access.2026.3745957 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Leheng Li und 4 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Natural Science Foundation of Shanghai, EURAMET Project.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Holographic particle detection remains challenging under diffraction blur, low contrast, and dense overlap. We propose the Edge Feature and Particle Focusing Network (EFPFNet), which integrates edge feature convolution (EFC) and a particle focusing mechanism (PFM) into YOLOv12. EFC strengthens blurred boundaries through a center-difference response, while PFM suppresses interference through pixel-level reward-penalty gating. The benchmark contains 500 simulated holograms, split into 400 training and 100 validation images, and 1,580 captured holograms. Before model selection, the captured data were divided by acquisition group into 316 validation images and 1,264 untouched test images. Configuration selection uses the real validation set, and checkpoint selection uses the simulation validation set. Across three independent runs, EFPFNet achieves mean average precision (mAP) of $(73.1\pm 0.5)\%$ at an intersection over union (IoU) threshold of 0.50 and $(29.0\pm 0.4)\%$ averaged over IoU thresholds of 0.50–0.95, compared with $(66.3\pm 0.4)\%$ and $(21.2\pm 0.3)\%$ for YOLOv12 under matched training settings. The model has 3.73 million parameters and requires 4.10 billion floating-point operations per image. Zero-shot evaluation on the independent 300-image IHPD-300 dataset yields 68.9% mAP@50 and 25.6% mAP@50–95; a separate 2,100-image study shows consistent gains across controlled acquisition conditions. EFPFNet achieves 52.3 frames per second on an NVIDIA Jetson Orin Nano with $416\times 416$ input, supporting real-time inference on the tested device.

Förderung

  • National Natural Science Foundation of China
  • Natural Science Foundation of Shanghai
  • EURAMET Project
  • Open Research Project of National Key Laboratory of Crystal Materials, Shandong University

Zitieren

APA

Li, L., Wang, C., Ni, Z., Wang, P., & Zhang, X. (2026). Edge Feature and Particle Focusing Network for Holographic Particle Detection Under Diffraction Blur and Interference. IEEE Access, 14, 167681-167695. https://doi.org/10.1109/access.2026.3745957

Vancouver

Li L, Wang C, Ni Z, Wang P, Zhang X. Edge Feature and Particle Focusing Network for Holographic Particle Detection Under Diffraction Blur and Interference. IEEE Access. 2026;14:167681-167695. doi:10.1109/access.2026.3745957

BibTeX
@article{Li2026,
  author = {Li, Leheng and Wang, Chen and Ni, Zhanghao and Wang, Pengxiang and Zhang, Xiangchao},
  title = {Edge Feature and Particle Focusing Network for Holographic Particle Detection Under Diffraction Blur and Interference},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {167681--167695},
  doi = {10.1109/access.2026.3745957}
}
Kostenlos starten →