Phase Unwrapping with Deep Image Prior (PUDIP)

README still under construction.

Don't hesitate to contact me (Thanh-an Pham) for more information.

Relevant publication:

F. Yang, T. -A. Pham, N. Brandenberg, M. P. Lütolf, J. Ma and M. Unser, "Robust Phase Unwrapping via Deep Image Prior for Quantitative Phase Imaging," in IEEE Transactions on Image Processing, vol. 30, pp. 7025-7037, 2021, doi: 10.1109/TIP.2021.3099956.

@ARTICLE{yang2021robust, author={Yang, Fangshu and Pham, Thanh-An and Brandenberg, Nathalie and Lütolf, Matthias P. and Ma, Jianwei and Unser, Michael}, journal={IEEE Transactions on Image Processing}, title={Robust Phase Unwrapping via Deep Image Prior for Quantitative Phase Imaging}, year={2021}, volume={30}, number={}, pages={7025-7037}, doi={10.1109/TIP.2021.3099956}}

Requirements

• pytorch
• matplotlib

Tried on Macbook M1

Basic Info

PUDIP can be called in Python as a function. Just calling PUDIP without argument will unwrap an example wrapped image.

from PUDIP import PUDIP
UnwrapPhase = PUDIP(wrapped_data = None, target = None, parserin = [])

PUDIP can be called as a Python function from command line, e.g.,

python PUDIP 'example.npy'

Inputs:

• wrapped_data is either the filename of the wrapped phase (.npy) or a numpy.ndarray
• target (ground-truth) is either the filename of the target (.npy) or a numpy.ndarray
• parserin is either a json filename or a list of dictionaries that overrules the "overlapping" default parameters. Please have a look at the pars_default.json (and below) for all the possible keys.

Return:

• UnwrapPhase: reconstructed phase (size is H*W)

Keys of parser:

Parameters for data

• RealData : true or false (change the way the scalar bias is removed)
• FileName : Name of data (for saved files)

Parameters for Network

• LR : Learning rate (default:0.01)
• NoiseType : 'Normal' or 'Std': adds noise.normal_()*reg_noise_std or noise.std()*reg_noise_std (default: 'Normal') to input at each iteration
• reg_noise_std: Weight for additional noise added to input (default:0.01)
• input_num : Number of input image, 1 for normal and multiples for sequential case (default:1)
• input_depth : Input channel (default:128)
• output_depth : Output channel (default:1)
• num_iter : Total number of iterations
• reg_loss : If true will compute the weighted loss: sum(wn.*Unwraploss)
• update_ite : The weights wn will be updated at every update_ite iteration (valid when the reg_loss is true)
• boundWeights : the (1/weights) min and max value for the loss weights (see paper)
• GDeps : The constant (GDeps^2) added to loss to make the BP stable (e.g. 1e-9)
• Plot : If true will plot the intermediate results
• show_every : Iteration for intermediate plotting
• gpu : If true will run the code on GPU
• gpuID : GPU ID
• main_dir : Main path (e.g.,'./')

Parameters for network architecture

• convfilt = [128,128,128,128,128]
• kernelsize = 3
• actifuncrelu = True
• LR_decrease = False # if Ture will apply the torch.optim.lr_scheduler.StepLR
• OptiScStepSize = 1000 # the parameter is valid when the LR_decrease is True
• OptiScGamma = 0.5 # the parameter is valid when the LR_decrease is True
• act_fun = 'PReLU' # 'ReLU','LeakyReLU(0.1, inplace=True)','Swish','ELU', 'PReLU'or'none'
• INPUT = 'noise' # 'noise' or 'meshgrid'
• OPTIMIZER = 'adam' # 'LBFGS'
• OPT_OVER = 'net' # 'input', 'net', 'net,input'
• NET_TYPE = 'skip' # skip, ResNet, UNet,Decoder
• upsample_mode = 'bilinear' # 'nearest', 'bilinear'; for UNet, deconv is valid
• pad = 'zero' # 'reflection', 'none','zero'

MAIN PARAMETERS (default)

• SaveRes = True # Save figures and result
• ItUpOut = 100 # Compute Loss, SNR, (save if SaveRes) every ItUpOut iterations
• figSize = 6 # Figure size
• BatchSize = 1 # Batch size (only one in DIP)
• bgwin = [3,52,3,52] #Used if RealData is True. Area used to compute the scalar bias (should be a background, no sample-induced phase delay)