Poisson denoising using Poisson2Sparse

This code shows how to restore a single image corrupted by Poisson noise using Poisson2Sparse, without requiring external training or knowledge of the noise level.

This method is based on the paper “Poisson2Sparse” Ta et al.[1] and restores an image by learning a sparse non-linear dictionary parametrized by a neural network using a combination of Neighbor2Neighbor Huang et al.[2], of the negative log Poisson likelihood, of the \(\ell^1\) pixel distance and of a sparsity-inducing \(\ell^1\) regularization function on the weights.

import deepinv as dinv
import torch

Load a Poisson corrupted image

This example uses an image from the microscopy dataset FMD Zhang et al.[3].

# Seed the RNGs for reproducibility
torch.manual_seed(0)
torch.cuda.manual_seed(0)

device = dinv.utils.get_freer_gpu() if torch.cuda.is_available() else "cpu"

physics = dinv.physics.Denoising(dinv.physics.PoissonNoise(gain=0.01, normalize=True))

x = dinv.utils.demo.load_example(
    "FMD_TwoPhoton_MICE_R_gt_12_avg50.png", img_size=(256, 256)
).to(device)
x = x[:, 0:1, :64, :64]
x = x.clamp(0, 1)
y = physics(x)

Selected GPU 0 with 5152.625 MiB free memory

Define the Poisson2Sparse model

backbone = dinv.models.ConvLista(
    in_channels=1,
    out_channels=1,
    kernel_size=3,
    num_filters=512,
    num_iter=10,
    stride=2,
    threshold=0.01,
)

model = dinv.models.Poisson2Sparse(
    backbone=backbone,
    lr=1e-4,
    num_iter=200,
    weight_n2n=2.0,
    weight_l1_regularization=1e-5,
    verbose=True,
).to(device)

Run the model

Note that we do not pass in the physics model as Poisson2Sparse assumes a Poisson noise model internally and does not depend on the noise level.

x_hat = model(y)

# Compute and display PSNR values
learning_free_psnr = dinv.metric.PSNR()(y, x).item()
model_psnr = dinv.metric.PSNR()(x_hat, x).item()
print(f"Measurement PSNR: {learning_free_psnr:.1f} dB")
print(f"Poisson2Sparse PSNR: {model_psnr:.1f} dB")

# Plot results
dinv.utils.plot(
    [y, x_hat, x],
    titles=["Measurement", "Poisson2Sparse", "Ground truth"],
    subtitles=[f"{learning_free_psnr:.1f} dB", f"{model_psnr:.1f} dB", ""],
)

  0%|          | 0/200 [00:00<?, ?it/s]
  4%|▎         | 7/200 [00:00<00:02, 68.79it/s]
  8%|▊         | 16/200 [00:00<00:02, 78.54it/s]
 12%|█▏        | 24/200 [00:00<00:02, 59.65it/s]
 16%|█▌        | 31/200 [00:00<00:03, 47.42it/s]
 18%|█▊        | 37/200 [00:00<00:03, 42.89it/s]
 21%|██        | 42/200 [00:00<00:03, 40.89it/s]
 24%|██▎       | 47/200 [00:01<00:03, 38.95it/s]
 26%|██▌       | 52/200 [00:01<00:03, 38.01it/s]
 28%|██▊       | 56/200 [00:01<00:03, 37.27it/s]
 30%|███       | 60/200 [00:01<00:03, 36.85it/s]
 32%|███▏      | 64/200 [00:01<00:03, 36.61it/s]
 34%|███▍      | 68/200 [00:01<00:03, 36.28it/s]
 36%|███▌      | 72/200 [00:01<00:03, 35.72it/s]
 38%|███▊      | 76/200 [00:01<00:03, 35.71it/s]
 40%|████      | 80/200 [00:01<00:03, 35.98it/s]
 42%|████▏     | 84/200 [00:02<00:03, 35.90it/s]
 44%|████▍     | 88/200 [00:02<00:03, 35.71it/s]
 46%|████▌     | 92/200 [00:02<00:03, 35.65it/s]
 48%|████▊     | 96/200 [00:02<00:02, 35.27it/s]
 50%|█████     | 100/200 [00:02<00:02, 35.96it/s]
 52%|█████▏    | 104/200 [00:02<00:02, 35.98it/s]
 54%|█████▍    | 108/200 [00:02<00:02, 35.42it/s]
 56%|█████▌    | 112/200 [00:02<00:02, 35.49it/s]
 58%|█████▊    | 116/200 [00:02<00:02, 35.15it/s]
 60%|██████    | 120/200 [00:03<00:02, 35.34it/s]
 62%|██████▏   | 124/200 [00:03<00:02, 35.51it/s]
 64%|██████▍   | 128/200 [00:03<00:01, 36.08it/s]
 66%|██████▌   | 132/200 [00:03<00:01, 35.72it/s]
 68%|██████▊   | 136/200 [00:03<00:01, 35.72it/s]
 70%|███████   | 140/200 [00:03<00:01, 35.64it/s]
 72%|███████▏  | 144/200 [00:03<00:01, 35.54it/s]
 74%|███████▍  | 148/200 [00:03<00:01, 35.15it/s]
 76%|███████▌  | 152/200 [00:03<00:01, 35.24it/s]
 78%|███████▊  | 156/200 [00:04<00:01, 35.12it/s]
 80%|████████  | 160/200 [00:04<00:01, 35.52it/s]
 82%|████████▏ | 164/200 [00:04<00:01, 35.01it/s]
 84%|████████▍ | 168/200 [00:04<00:00, 35.27it/s]
 86%|████████▌ | 172/200 [00:04<00:00, 35.61it/s]
 88%|████████▊ | 176/200 [00:04<00:00, 35.60it/s]
 90%|█████████ | 180/200 [00:04<00:00, 35.40it/s]
 92%|█████████▏| 184/200 [00:04<00:00, 35.59it/s]
 94%|█████████▍| 188/200 [00:05<00:00, 35.92it/s]
 96%|█████████▌| 192/200 [00:05<00:00, 35.49it/s]
 98%|█████████▊| 196/200 [00:05<00:00, 35.51it/s]
100%|██████████| 200/200 [00:05<00:00, 35.44it/s]
100%|██████████| 200/200 [00:05<00:00, 37.43it/s]
Measurement PSNR: 27.5 dB
Poisson2Sparse PSNR: 31.1 dB

References:

[1]

Calvin-Khang Ta, Abhishek Aich, Akash Gupta, and Amit K Roy-Chowdhury. Poisson2sparse: self-supervised poisson denoising from a single image. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 557–567. Springer, 2022.

[2]

Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14781–14790. 2021.

[3]

Yide Zhang, Yinhao Zhu, Evan Nichols, Qingfei Wang, Siyuan Zhang, Cody Smith, and Scott Howard. A poisson-gaussian denoising dataset with real fluorescence microscopy images. In CVPR. 2019.