Noise Estimation And Removal For High ISO JEPG Images

When the lighting is poor or capturing fast motion objects,the camera needs to be adjusted to high ISO mode to record the details clearly.However,capturing images at high ISO mode will introduce much realistic noise,which is not simple Gaussian white noise,but a kind of content correlated noise.Most of the traditional image denoising methods are designed for Gaussian or Poisson noise,which are not suitable for realistic noise introduced in the complicated imaging pipeline,so,it is diffcult for traditional denoising algorithms to remove the realistic noise.Therefore,studying denoising algorithm that can handle images taken at high ISO mode has great significance.Based on the study of the distribution characteristics of realistic noise,this paper improves the traditional collaborative filtering method,and explore the advantages of collaborative filtering and deep learning denoising algorithms to study the noise estimation and removal algorithm for JEPG images captured at high ISO mode.The work and contribution of this paper are summarized as follows.1.We propose a noise estimation method based on Bayer sampling.Due to the demosaicing process in imaging,the noise variance maps of captured JPEG images are spatial-dependent and characterized by Bayer pattern.Therefore,we introduce the prior of Bayer pattern into noise estimation and design a new noise estimation algorithm.Firstly,the noise image is downsampled according to the Bayer pattern to obtain four sub-images.Then,the noise estimation is performed on the each sub-image to produce the noise variance maps of the four sub-images separately.Finally,noise variance map of the whole image is obtained by up-sampling the noise map of for sub-images to their original position according to the Bayer pattern.Experiments show that the proposed noise estimation method outperforms many noise estimation methods.2.We propose fusing the collaborative filtering and convolutional filtering to remove the realistic noise.For collaborative filtering,we introduced the nosie map and Bayer pattern prior into CBM3 D,which is designed for Gaussian noise,so that it can handle the realistic noise.For convolutional filtering,we constructed a deep convolutional neural network to remove the realistic noise.Since collaborative filtering is good at recovering repeatable structures and convolutional filtering is good at removing noise in flat regions and recovering irregular patterns,we propose to fuse the strengths of two methods via deep CNN to improve the denoising results.Experiments show that the proposed method can preserve the textures while remove remove the noise in smooth region thoroughly,and the proposed outperform state-of-the-art denoising methods in objective measurements.In addition,we construct a dataset with high ISO short exposure and low ISO long exposure images to facilitate research on this topic.