Deep Noise Estimation And Removal For Real-world Noisy Images

Noise is introduced inevitably when capturing under low-light conditions or high ISO mode,even though the quality of digital cameras is improving.There are many cases that we need to utilize high ISO mode of the camera,such as recording details at night or capturing fast moving objects.There is spatial and channel correlations in the captured realistic noise,which is more complex than Gaussian noise.Therefore,the Gaussian noise removal methods cannot deal with this kind of noise well.Based on the above observations,we propose a novel convolutional neural network(CNN),which has promising capability in feature learning,to jointly solve the realistic noise estimation and removal task.Our contributions are summarized as follows:1.The original captured Raw image is characterized by the Bayer Pattern due to the color filter array(CFA).After the interpolation in demosaicing,the variance of noise at the interpolation location is lower than its neighbors.Therefore,the noise variance map also presents Bayer pattern.Based on this observation,we propose a noise estimation network which takes advantage of the Bayer pattern prior of the noise variance maps.Compared with the method without utilizing the Bayer prior,the proposed method presents better noise estimation performance and consumes less memory.2.Since the realistic noise has spatial and channel correlations,we propose a RSD(Residual SE Dilated)block which combines the advantages of dilated convolution and SE block to fully take advantage of the spatial and channel correlations.Since the estimated noise variance map and the original image have similar structures,the estimated noise variance map is weighted concatenated with the noisy input to further boost the denoising performance.3.Since it is much difficult to collect paired noisy and noise free images,we propose to synthesize noisy images according to the imaging pipeline.During training,the network is first pretrained on the synthesized images,and then we use a small amount of captured paired images to finetune the network,which greatly reduces the demanding of captured paired images.Experimental results demonstrate that the proposed method outperforms competing methods for both noise estimation and removal tasks.