Joint Demosaicing And Denoising Based On Deep Learning

At present,most digital cameras only use single-sensor camera systems to capture images for the purpose of reducing the size and cost,but single-sensor camera systems can only obtain one color component of each pixel.In order to obtain a color image acceptable to the human eye,a corresponding method must be used to recover these lost color information.These methods are called CFA interpolation,and are also referred to as CFA image demosaicing.In addition,due to the effects of electromagnetic and thermal effects,a certain amount of noise is usually introduced during the acquisition of CFA images.Therefore,demosaicing and denoising is an important part of digital cameras to obtain high-quality color images.Researching more efficient demosaicing and denoising methods can not only meet the visual requirements,but also lay a good foundation for subsequent advanced image processing tasks.The traditional joint demosaicing and denoising method mainly uses the degradation model and some prior information of CFA image to get the estimated value of color image,but these methods do not work well on the regions with more textures in the image,mainly manifested as false colors,aliasing effects,and artifacts.The joint demosaicing and denoising method based on deep learning proposed in recent years solves the problem of introducing prior information in the traditional method,which greatly improves the quality of color images.However,shallow convolution neural network on deep learning still can not fully extract details in images,which further leads to the fact that higher quality color images cannot be reconstructed.Based on the above research,this paper focuses on the study of joint demosaicing and denoising of CFA images based on deep learning.The main work and contributions are as follows:1.A joint demosaicing and denoising of CFA images based on deep convolutional network is proposed.Because the shallow convolutional neural network can not extract much detailed information in the image,this paper designs a novel deep convolutional neural network.Then,this network and a large number of training samples are used to fit the nonlinear mapping relationship between the noisy CFA image and the clear color image.Finally,this relationship is used as the guidance to joint demosaicing and denoising.Specifically,a series of convolutional layers is used to extract the spatial correlation and inter-channel correlation information contained in the CFA image,then the sub-pixel convolutional layer and a convolutional layer are used to recover the final color image.Experimental results show that this method can obtain better results of demosaicing and denoising.2.A joint demosaicing and denoising of CFA images based on generative adversarial network is proposed.Many of the existing demosaicing and denoising methods(including the one proposed in contribution 1)have lost much texture information in the reconstructed color images,and the image tends to be smooth.To solve the above problems,this paper propose a joint demosaicing and denoising method based on the generative adversarial network.This method adds another convolutional neural network based on the network model in contribution 1,and uses the network to correct the mapping relationship learned by the network in the contribution 1,so that the output of the network in the contribution 1 is close to the real color image as much as possible with probability 1.Experimental results show that this method can obtain color images with richer texture and better visual effects.