rushipatel7 / image-denoising

WDnCNN: ENHANCEMENT OF A CNN-BASED DENOISER BASED ON SPATIAL AND SPECTRAL ANALYSIS

Dependencies

Python >= 3.6.5, Pytorch >= 0.4.1, and cuda-9.2.

Pretrained Models

Two pretrained WDnCNN models, WDnCNN_model_gray and WDnCNN_model_color, are used for evaluating the denoising performance on grayscale images and color images, respectively. Run demo.py to test the WDnCNN for both synthetic and real-world noise removal.

Network Architecture

Band Discriminative Training(BDT)

Please refer to the paper and the README.txt file.

Results

Grayscale AWGN Removal

Color AWGN Removal

Visual Results on Real-world Noisy Images

Noisy	CBM3D
FFDNet	Ours

Real-world Denoising Benchmark

We also evaluate our method on the 1,000 cropped real-world noisy images from Darmstadt Noise Dataset. You can find this benchmark at DND. For denoising the real-world noisy images in DND, we further fine tune our model on PolyU-Real-World-Noisy-Images-Dataset PRWNID. In the fine tuning, we adopt the sub-network for noisy level estimation in CBDNet, and jointly fine tune the sub-network with our WDnCNN.

You can find our results as WDnCNN+ on the DND official website. We achieve 38.87dB on sRGB images.

One PyTorch Implementation of CBDNet can be found at CBDNet_PyTorch.