Unsupervised physical neural network empowers stacked imaging denoising algorithm

In view of the reconstruction problem of stack imaging technology in noisy environment, Lin Nan's team from Shanghai Institute of Optics and Mechanics, Chinese Academy of Sciences, proposed an innovative method ProPtyNet based on unsupervised physical neural network, which is expected to be applied to chip CD measurement and defect detection. The article was published in Optics and lasers in engineering under the title of "Noise robust photography using unsupervised natural network".
Its core innovations and achievements are as follows:

1. Core issues and challenges
Stacked imaging uses diffraction patterns to invert the complex amplitude distribution of objects, but faces challenges in the following scenarios:
Complex noise environment: Poisson noise (signal-to-noise ratio 24.64dB) under low light conditions and mixed noise caused by random high-energy particles can lead to the failure of traditional algorithms
Dynamic bandwidth imaging: Non monochromatic light sources make non target wavelengths a new frequency domain noise source
Hardware limitations: Traditional iterative algorithms are slow and difficult to achieve fast imaging.

2. Method innovation
Propose a dual driven framework of "physical model+deep learning"
Data processing architecture: Zero padding preprocessing (e.g. 512 → 612 pixels) and four component output (amplitude/phase x object/probe) are used to improve axial resolution
Network topology optimization: Customize U-net architecture (with only 2.5 million parameters) to achieve four-dimensional parameter joint optimization, and extend the dynamic range of the "Conv2d tanh" phase layer to 2 π
Anti noise loss function: pioneering a dual loss mechanism (β=0.85-0.95) to balance overexposed areas (γ=1 → 0.1 gradient) and probe structured constraints, resulting in a 5-fold decrease in loss function

3. Experimental verification
Verify performance through 600 sets of tests:
Noise robustness: Under mixed noise (SNR 30dB), the SSIM value reaches 0.92 ± 0.03, which is about 14 times higher than ePIE
Speed advantage: The single convergence time is 729 seconds, which is 47.8% and 31.9% faster than AD and ePIE, respectively
Wide spectral adaptability: effectively separates ± 5nm noise components in the 405nm band, achieving a resolution of 57 line pairs/mm

4. Application prospects
Extreme Ultraviolet EUV Imaging: The method has been adapted to the Fresnel propagation model and can be extended to a wavelength of 13.5nm
Low dose dynamic monitoring: Effective suppression of readout noise under 300 μ s short exposure conditions (NPS=0.12)
Multi parameter joint calibration: achieving joint calibration along the axis (zPIE) and oblique incidence (aPIE) by adding output channels [35-36]
This study provides a breakthrough in both computer principles and experimental models for coherent imaging of new light sources (X-ray/EUV) and extreme operating conditions (low temperature/irradiation).

Figure 1. ProPtyNet algorithm details. (a) ProPtyNet network flowchart. (b) Details of U-net network architecture. (c) Data preprocessing and post-processing methods.

Figure 2. Simulation reconstruction results. (a) The true amplitude and phase of objects and probes. (b) The reconstruction results of ProPtyNet, ePIE, rPIE, and AD algorithms under Poisson noise, Gaussian noise, and mixed noise.

Figure 3. Experimental results. (a) The reconstruction results of sample amplitude information under different noise environments. (b) Comparison of amplitude cross-sections in the underlined section. (c) Comparison of loss functions in the iterative process.

Figure 4. Reconstruction results under broad-spectrum illumination (a) Experimental illumination light spectrum. (b) Simulation and experimental results.

Source: opticsky