search

Zhenxing-Dong / Fourier-inspired-neural-module

[Optics Letters] Fourier-inspired neural module for real-time and high-fidelity computer-generated holography
19 stars 4 forks source link

Normalization before rfft and irfft #2

Open beginnerinphysics opened 1 month ago

beginnerinphysics commented 1 month ago

Hello, Zhenxing, I am a beginner in this Holography, and I find in rfft2 and irfft2 there's a norm=ortho, image due to the default about these function is no normalization, I want to ask, is norm = orthonormal an experiment result, which always make the performance better? Thanks for the helping

Zhenxing-Dong commented 1 month ago

Hi! In fact, we didn't try different normalizations for FFT in the paper. In my opinion, the role of the Fourier transform is to learn the diffractive information in the hologram generation process. Therefore, I believe that the Fourier-inspired neural module can enhance performance regardless of the normalization. Of course, you can try different normalization methods to observe their impact on the final results.