FC-AIDE: Fully Convolutional Adaptive Image Denoiser

code for testing
- 'sigma_estimation.py'
  : estimate a noise sigma of the noisy image
- 'test_fc_aide_sup.py'
  : denoise a noisy image using the supervised model trained on the specific noise level
- 'test_fc_aide_ft.py'
  : denoise a noisy image using the supervised model + fine-tuning
- 'test_fc_aide_blind_sup.py'
  : denoise a noisy image using the supervised model trained on various noise level
- 'test_fc_aide_blind_ft.py'
  : denoise a noisy image using the supervised model + fine-tuning
- 'test_fc_aide_blind_estimated_sigma_ft.py'
  : estimate a sigma of the noise level and then denoise using the supervised model + fine-tuning
weights
- 'sigmaX.hdf5' : a weight trained on a specific noise level([15, 25, 30, 50, 75])
- 'blind.hdf5' : the weight trained on various noise levels([0,50])
test images
- 'Set13'
- 'BSD68'
- 'Medical60'

Results

Set13

The average PSNR(dB) on Set13.

Sigma	BM3D	RED	MemNet	DnCNN-S	DnCNN-B	FC-AIDE_{_S}	FC-AIDE_{_S+FT}	FC-AIDE_{_B}	FC-AIDE_{_B+FT}
15	31.98	-	-	32.21	31.58	32.08	32.59	31.72	32.52
25	29.44	-	-	29.63	29.22	29.57	30.14	29.34	30.03
30	28.56	28.91	28.83	28.64	28.41	28.73	29.28	28.50	29.19
50	26.05	26.28	26.39	26.09	26.07	26.24	26.87	26.05	26.77
75	24.16	-	-	24.03	18.33	24.24	24.97	21.07	24.89

BSD68

The average PSNR(dB) on BSD68.

Sigma	BM3D	RED	MemNet	DnCNN-S	DnCNN-B	FC-AIDE_{_S}	FC-AIDE_{_S+FT}	FC-AIDE_{_B}	FC-AIDE_{_B+FT}
15	31.07	-	-	31.72	31.60	31.63	31.75	31.47	31.72
25	28.56	-	-	29.22	29.15	29.18	29.31	29.04	29.26
30	27.74	28.45	28.42	28.35	28.34	28.35	28.49	28.24	28.44
50	25.60	26.29	26.34	26.21	26.20	26.24	26.38	26.12	26.33
75	24.19	-	-	24.62	18.68	24.74	24.87	21.42	24.76

Mismatch / blind case

Requirements and Dependencies

Python 2.7 / Python 3.6
CUDA v8.0 / CuDNN v5.1
Tensorflow v1.2.1
Keras v2.0.8

citation

@ARTICLE{2018arXiv180707569C,  
    author = {{Cha}, S. and {Moon}, T.}a,  
    title = "{Fully Convolutional Pixel Adaptive Image Denoiser}",  
    journal = {ArXiv e-prints},  
    archivePrefix = "arXiv",  
    eprint = {1807.07569},  
    primaryClass = "cs.CV",  
    keywords = {Computer Science - Computer Vision and Pattern Recognition, Computer Science - Machine Learning, Statistics - Machine Learning},  
    year = 2018,  
    month = jul,  
    adsurl = {http://adsabs.harvard.edu/abs/2018arXiv180707569C},  
    adsnote = {Provided by the SAO/NASA Astrophysics Data System}  
}

FC-AIDE: Fully Convolutional Adaptive Image Denoiser

Contents

Results

Set13

BSD68

Mismatch / blind case

Requirements and Dependencies

citation