Research On Image Restoration Technology Based On Deep Learning

In recent years,with the development of deep learning technology,it has provided a new direction in the field of image restoration.Many image restoration algorithms based on deep learning methods have emerged.Previous studies have mainly focused on image denoising and super-resolution reconstruction in image restoration.Essentially,these two fields have the same research goal,that is,to reconstruct the missing information.Image denoising mainly aims at the problem of Gauss white noise denoising.Existing methods based on prior or learning have the problems of blurred image details or edges and high time-consuming.Image super-resolution reconstruction mainly enlarges and reconstructs a single image with fixed scale.Existing methods based on reconstruction model or learning have some problems,such as incomplete image detail reconstruction and large amount of algorithm parameters.To solve these two problems from the perspective of deep learning technology is basically the same,that is,to extract more effective image features.After a thorough investigation of this two directions,the research focus and difficulties in the two fields are studied and improved respectively.In image denoising,the existing convolution neural network methods have some problems,such as slow convergence speed,slow testing speed,unable to use the same network to solve the gray and color image denoising problems.In this paper,an image denoising algorithm based on multi-scale features is proposed,which uses convolution kernels of different sizes to extract features of different scales.What's more,the proposed method uses standardized extended convolution design method,residual learning,PReLU and other commonly used deep learning techniques.Finally,we trains an end-to-end deep denoising network.Several model ablation experiments are used to determine the network structure.The experimental results show that the denoising effect of the proposed algorithm on multiple benchmark sets is better than the existing deep learning methods.The model parameters are nearly half as much as those of DnCNN^[15]and FFDNet^[65].In addition,the denoising effect of the proposed algorithm on color images is better than the existing deep learning methods without increasing the parameters.In image super resolution reconstruction,the existing super resolution reconstruction methods based on deep learning can objectively evaluate the quality of image reconstruction,but most of these methods belong to deep and wide depth network structure,which has a large amount of computation and is not conducive to practical industrial applications.In this paper,a fast super-resolution reconstruction algorithm based on channel attenuation is proposed.The main idea is to use the idea of image feature reuse to concatenate multi-layer image features.And we design a method of decreasing the number of convolution cores as the number of layers increases.Finally,we achieves the goal of reducing the number of model parameters without reducing the effect of image reconstruction.Compared with many existing lightweight deep learning algorithms and classical super-resolution reconstruction methods on several benchmark datasets,the experimental results show that the proposed algorithm can reduce at least half of the parameters of CARN-M^[78]and FALSR-B^[32]without reducing the image reconstruction effect.