Computational-Imaging-RU / Bagged-DIP-Speckle

Official Code of Bagged Deep Image Prior for Recovering Images in the Presence of Speckle Noise (ICML 2024)

Bagged Deep Image Prior for Recovering Images in the Presence of Speckle Noise

Preview

There are 4 python files in this repo. The test data is under the /data/ folder.

• train_BaggedDIP_speckle.py: training Bagged-DIP based PGD algorithm.

• function_grad.py: i) explicit gradient function of the MLE loss function, and ii) implementation of Newton Schulz algorithm for efficently approximating matrix inverse.

• decoder.py: basic network structures of the Deep Image Prior/Deep Decoder we use in Bagged-DIP.

• utils.py: all the other helper functions.

Running the code

Run the Bagged-DIP PGD algorithm for recovering images:

python train_BaggedDIP_speckle.py

Specify the hyperparameters and experiment setting:

E.g., recover images from L=100, m/n=0.5 down-sampled complex-valued measurements:

python train_BaggedDIP_speckle.py --compression_rate 0.5 --patch_size 128 --lamb 1.0 --crop True --lr_NN 1e-3 --lr_GD 0.01 --outer_ite 100 --num_look 100 --weight_decay 0.0 --test_name 'Set11' --DIP_avg True --use_complex True

Relevant works on speckle noise

[1] Chen, Xi, Zhewen Hou, Christopher Metzler, Arian Maleki, and Shirin Jalali. "Bagged Deep Image Prior for Recovering Images in the Presence of Speckle Noise." arXiv preprint arXiv:2402.15635 (2024). paper

[2] Chen, Xi, Zhewen Hou, Christopher Metzler, Arian Maleki, and Shirin Jalali. "Multilook compressive sensing in the presence of speckle noise." In NeurIPS 2023 Workshop on Deep Learning and Inverse Problems. 2023. paper

[3] Zhou, Wenda, Shirin Jalali, and Arian Maleki. "Compressed sensing in the presence of speckle noise." IEEE Transactions on Information Theory 68.10 (2022): 6964-6980. paper