搜索

x
中国物理学会期刊

基于交叉结构光的物理驱动小波神经网络超分辨成像方法

Physics-Informed Wavelet Neural Network for Super-Resolution Imaging Based on Crossed Structured Illumination

PDF
导出引用
  • 结构光照明显微技术(SIM)因其较低光毒性和明场成像特性被广泛应用于超分辨成像。然而,传统SIM通常需在每个方向上采集三幅相移图像,并依据准确相移量实现重建;同时,对于具有一定厚度的样品,离焦效应易引入重建伪影并降低对比度。为此,本文提出一种物理驱动的小波神经网络(PIWNN)用于实现基于交叉结构光照明的超分辨重建。交叉结构光照明允许两个方向的高频信息同时被编码进莫尔条纹中,提高数据获取的速度同时有效抑制不同照明条件下零级频谱对重建结果的影响,PIWNN将SIM成像物理模型与小波网络相结合,通过成像物理模型约束网络重建过程,增强样品高频信息,并利用小波变换代替池化进行上下采样,以减少传统池化操作引起的信息损失同时扩大感受野。所提出方法能够在无先验系统参数和预训练数据的条件下,使用5幅随机相移的原始图像即可完成交叉结构SIM重建,成像速度与重建鲁棒性被显著提高。仿真与实验结果表明,该方法能够有效实现分辨率的提升,并验证了在低照度和噪声条件下仍能保持较高的结构相似性和信噪比。

     

    Structured Illumination Microscopy (SIM) has been widely used in super-resolution imaging owing to its low phototoxicity and wide-field imaging capability. However, traditional SIM requires three phase-shifted images for each illumination direction and relies on accurate phase-shift estimation. Meanwhile, for samples with a certain thickness, reconstruction artifacts are easily introduced and image contrast is reduced due to out-of-focus effects. To address these issues, this paper proposes a physics-informed wavelet neural network (PIWNN) for super-resolution reconstruction based on crossed structured illumination. Crossed structured illumination enables high-frequency information from two directions to be simultaneously encoded into Moiré fringes, thereby improving data acquisition speed while effectively suppressing the influence of the zero-order spectrum under different illumination conditions on the reconstruction results. PIWNN combines the SIM physical model with wavelet network. The physical model is used to constrain reconstruction process and enhance high-frequency information of sample, while wavelet transform is employed to replace pooling operation for upsampling and downsampling, thereby information loss can be reduced and the receptive field can be enlarged. The proposed method can complete crossed-illumination SIM reconstruction using only five raw images with random phase shifts, without prior system parameters and pretraining datasets, imaging speed and reconstruction robustness can be improved significantly. Simulation and experimental results demonstrate that the proposed method effectively improves resolution and maintains high structural similarity and signal-to-noise ratio under low-illumination and noisy conditions.

     

    目录

    /

    返回文章
    返回