Research On Real-World Camera Photographes Denoising Based On Convolutional Neural Network

Even though nowadays CNN-based denoising algorithms have achieved great performance on Gaussian denoising problem,deep learning methods still achieve very limit performance on real image denoising task.To promote the performance of CNN-based denoising method in practial use,we first design a blind denoising convolutional neural network,the main contraibution is that considering the effect of ISP process to noise,a more realistic noise model is proposed,next we design a twostage network structure to utilize the generalization ability of non-blind denoising structure.Firstly,a small network is utilized to estimate the noise level,then a UNet structure is used for non-blind denoising.For noise estimation subnetwork,we design an asymmetric loss function and utilizing a TV regularization at the same time to let the estimated noise level more suitable for non-blind denoiser.To further prove the denoising performance,real noise training data and synthesic training data are alternately used to update the network.Our network has proved to be very effective on multi real denoising datasets.To further improve the denoising performance on real image,also to promote the research process of multi frame denoising on real photographes,we collect a real denoising dataset(CIDB).Compared with previous real-noisy datasets,our dataset collect more scenes,the nosie level of images in the dataset is higher and contain more camera settings.To obtain more clean groundtruth image,also can remove the effect of moving objects in the scenes,we design a more compression and more robust post-processing algorithm,including intensity alignment,spatial alignment and content-aware iterative weighted meraging algorithm.We benchmark performance of most nowadays denoising algorithms on our dataset.Through utilizing the training dataset in our dataset,single frame and multi frames denoising networks are trained.Through experiments,these networks performance well on real phorograph denoising task.