Gated Fusion Network Based On Image Denoising And Super-resolution

In the field of image restoration,image denoising and image super-resolution reconstruction are popular directions relatively,and these have important research value in criminal investigation,military,satellite navigation,map information collection,and other application scenarios.Image denoising is the process of removing the mixed noise points in the image and restoring the missing content of the pixels through the correlation of the pixels.Image denoising methods are divided into three types based on filters,based on sparse dictionary and based on deep learning.At present,the traditional image denoising algorithm is difficult to obtain a good denoising effect when facing images under the influence of more complex noise.Image super-resolution reconstruction is the process of reconstructing low-resolution images to obtain high-resolution images by analyzing image signals.Image super-resolution reconstruction methods are divided into three types based on interpolation,sparse and deep learning.The traditional image super-resolution reconstruction algorithm simply super-resolution reconstruction of clear images,when the image has defects,it will affect the final super-resolution reconstruction effect,and most of them are performed in high-resolution space,which greatly increases the computational complexity of the overall algorithm.Faced with this situation,research on super-resolution reconstruction algorithms under complex noise environment still has a lot of room for development,so this thesis proposes a image denoising and super-resolution reconstruction algorithm based on gated fusion network.The main contributions of the gated fusion network algorithm proposed are the following aspects:(1)In order to better achieve the dual tasks of denoising and super-resolution reconstruction,a dual-branch gated fusion network is proposed to de-noise and image super-resolution reconstruction performs feature extraction at the same time,and then adaptively fuse the features of the two.Through continuous training,enable the fusion features of each round to be input into the training of the next round,which improves the final image recovery ability;(2)In view of the current complex noise impact and the high proportion of noise that makes it difficult to restore the image,the asymmetric residual encoding-decoding structure is used to expand the feature acceptance field,and all the pooling layer applications are cancelled.Drop Out is used toprevent training overfitting,making the structure of noise feature extraction is more simple and effective.Finally,two additional convolutional layers are used to reconstruct the image,which effectively improves the denoising ability and robustness;(3)The multi-scale residual block and multi-level feature fusion structure are used,in which the multi-scale residual block is used for feature extraction,and the redundant features are filtered out through the multi-level feature fusion structure and the combination of the front and back features is realized,achieve better image super-resolution reconstruction effect.After comparison experiments with the current image denoising algorithm and image super-resolution reconstruction algorithm,and data analysis on the evaluation indicators,it is proved that the algorithm proposed in this thesis can still recover a clear high-resolution images in the traditional noise environment and more complex noise environment.