Pyramid Real Image Denoising Network

This is the code for the paper "Pyramid Real Image Denoising Network". ( VCIP 2019 oral )

Paper Link : Pyramid Real Image Denoising Network

Training dataset : SIDD Medium Dataset

Validation dataset : SIDD Validation data

Testing dataset : SIDD Benchmark, DND, NC12

The trained model for raw_rgb : Google Driver,  Baidu Yunpan

While deep Convolutional Neural Networks (CNNs) have shown extraordinary capability of modelling specific noiseand denoising, they still perform poorly on real-world noisyimages. The main reason is that the real-world noise is moresophisticated and diverse. To tackle the issue of blind denoising,in this paper, we propose a novel pyramid real image denoisingnetwork (PRIDNet), which contains three stages. First, the noiseestimation stage uses channel attention mechanism to recalibratethe channel importance of input noise. Second, at the multi-scale denoising stage, pyramid pooling is utilized to extractmulti-scale features. Third, the stage of feature fusion adopts akernel selecting operation to adaptively fuse multi-scale features.Experiments on two datasets of real noisy photographs demon-strate that our approach can achieve competitive performancein comparison with state-of-the-art denoisers in terms of bothquantitative measure and visual perception quality.

Training

train_SIDD_Pyramid.py is used for training.

variable "dir_name": to store your training dataset.

variable "checkpoint_dir": to save your trained model.

variable "result_dir": to save the images produced in the training process.

Testing

test_SIDD_Pyramid.py is used for testing.

variable "val_dir": to store your testing dataset.

variable "checkpoint_dir": to store your pre-trained model.

variable "result_dir": the denoised result after testing.