Scholarlink

<p>Cross Scattering Polarization (CPS) is an optical technique that produces diffraction images free from illumination artifacts, yielding signals determined exclusively by the intrinsic optical and geometrical properties of bacteria. In this work, we evaluate the use of a VGG11 convolutional neural network for bacterial classification from CPS images generated through numerical simulations, and we investigate training strategies aimed at enhancing robustness to realistic noise sources, including photon shot noise, dark current, and readout noise. We systematically analyze the effect of spatial down-sampling, L2 regularization, and alternative pixel-scaling strategies (linear vs. logarithmic) on classification accuracy. Our results show that dimensionality reduction significantly improves robustness to noise, whereas L2 regularization has an inconsistent impact on generalization. Logarithmic scaling stabilizes training but reduces noise tolerance, while linear scaling provides the opposite behaviour. Finally, we demonstrate that noise in the high-intensity regions contributes minimally to model performance, whereas it dominates in the low-intensity (idle) regions the VGG11’s decision process. These results highlight the discriminative power of CPS and identify practical strategies to increase the robustness of deep-learning models trained on synthetic optical datasets.</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

Optical bacteria recognition: VGG-11 neural algorithms for pattern recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月22日

查看全文
  • Riccardo Pepino,
  • Francesco Calabrò,
  • Paolo Fella,
  • Hamed Tari,
  • Alessandro Bile,
  • Arif Nabizada,
  • Eugenio Fazio

摘要

Cross Scattering Polarization (CPS) is an optical technique that produces diffraction images free from illumination artifacts, yielding signals determined exclusively by the intrinsic optical and geometrical properties of bacteria. In this work, we evaluate the use of a VGG11 convolutional neural network for bacterial classification from CPS images generated through numerical simulations, and we investigate training strategies aimed at enhancing robustness to realistic noise sources, including photon shot noise, dark current, and readout noise. We systematically analyze the effect of spatial down-sampling, L2 regularization, and alternative pixel-scaling strategies (linear vs. logarithmic) on classification accuracy. Our results show that dimensionality reduction significantly improves robustness to noise, whereas L2 regularization has an inconsistent impact on generalization. Logarithmic scaling stabilizes training but reduces noise tolerance, while linear scaling provides the opposite behaviour. Finally, we demonstrate that noise in the high-intensity regions contributes minimally to model performance, whereas it dominates in the low-intensity (idle) regions the VGG11’s decision process. These results highlight the discriminative power of CPS and identify practical strategies to increase the robustness of deep-learning models trained on synthetic optical datasets.