Image Reconstruction Method Based On Convolutional Neural Network

As an important research topic in the field of image processing,image reconstruction has received extensive attention.This paper applied the CNN to the problem of image directions of image denoising and the CFA image demosaicing.The specific research contents include:(1)This paper proposed a multi-scale feature fusion based parallel dense residual denoising network framework.First,the framework adopts the parallel network structure to combine the image feature information of different depths.Each branch of the network consists of a sequence of dense residual blocks(DRB).On this basis,the remote skipping connections between the DRB was added to overcome the gradient disappearance and gradient dispersion problems during the network training and strengthen the network's ability.In addition.on the basis of combining the shallow and deep information of the image,the multi-scale feature fusion blocks(MFFB)were added in each branch of the network to obtain multi-scale image features with different depths.Finally,the performance of the network was further improved by using residual learning.The method can better retain the effective feature information of the high frequency part for the image,and greatly improve the clarity and naturalness of the image,and has better denoising effect.(2)This paper proposed a generative adversarial neural network based on dense residual for CFA image demosaicing.The generator is designed by using deep dense residual network which includes dense residual blocks and remote skipping connections.The discriminator consists of a series of stacked convolutional units.In addition,the adversarial loss,the pixel loss and the feature loss were combined to improve the network's ability to reconstruct the image.The method can better restore the effective information of the edge and detail for the image,and has a well work on suppressing the artifact phenomenon of the high frequency part of the image compared with some traditional algorithms.(3)This paper proposed a generative adversarial neural network based on U-net and dense residual for CFA image demosaicing.The generator is designed by improved U-net model,and the discriminator used dense residual network.The method can effectively suppress the artifact phenomenon,and it can better recover the high frequency information such as the texture and edge of the image,and produce the high quality image close to the fact.