lyuzinmaxim/vur-net

PyTorch model VUR-Net for phase unwrapping

VUR-Net

PyTorch model VUR-Net for phase unwrapping

This is a PyTorch realisation of deep convolutional Unet-like network with extra convolutional layers, described in arcticle [1].

Usage

<pip install -r example-requirements.txt>

Changes

I've added following moments to the structure:

1. Replication padding mode in conv3x3 blocks, because experiments have shown that it's important at the edges of phase maps, otherwise unwrapping quality will be low

Dataset

Dataset was generated synthetically according to articles [1,2] So, dataset data was generated using two methods (in equal proportions):

1. Interpolation of squared matrixes (with uniformly distributed elements) of different sizes (2x2 to 15x15) to 256x256 and multiplying by random value, so the magnitude is between 0 and 22 rad
2. Randomly generated Gaussians on 256x256 field with random quantity of functions, means, STD, and multiplying by random value, so the magnitude is between 2 and 20 rad. From experiments with real simple phase images it's clear, that that method makes net more adapted for real-life examples

Model

Model can be shown as following (from original article [1]):

Training info

In original paper authors describe train hyperparameters as follows:

loss: pixelwise MAE

optimizer: Adam

learning rate: 2e-4 and "divided by 2 when learning stagnates"

Succeed train to zero cost (0.0162) at epoch 500 with Adam 0.0001 very fast

Todo's

1. code refactoring

References

1. Qin, Y., Wan, S., Wan, Y., Weng, J., Liu, W., & Gong, Q. (2020). Direct and accurate phase unwrapping with deep neural network. Applied optics, 59 24, 7258-7267 .
2. K. Wang, Y. Li, K. Qian, J. Di, and J. Zhao, “One-step robust deep learning phase unwrapping,” Opt. Express 27, 15100–15115 (2019).
3. Spoorthi, G. E. et al. “PhaseNet 2.0: Phase Unwrapping of Noisy Data Based on Deep Learning Approach.” IEEE Transactions on Image Processing 29 (2020): 4862-4872.

vur-net's Issues

Give some guidance on the parameter setting

Hello, I have observed you mentioned the succeed train to zero cost (0.0162) at epoch 500 with Adam 0.0001 very fast. Howerver, I cannot obtain the loss as low as you said. I used the automatic generated synthetic data provided by dataset_generation.py. My question is how many phase images are gernerated? My accuracy of unwrapping cannot attain the 50% above. I am hoping you could give me some guidance on the parameter setting.

can you supply the pre-train model?

hello , I use your code , and run the train, but i find cannt unwrap the phase. can you supply the pre-train mode?