Research On Denoising Method Of Group Sparse Residual Image Based On Deep Learning

In the process of obtaining images,noise may generate due to various hardware devices and during the transmission process,thus,the images observed by the human eye are often of poor quality.In order to understand and analyze the image,image noise may interfere with the recognition of information in clear images by people or computers.Image noise tends to cause serious mistakes while high-level visual tasks are performed on computers.Especially in some special fields,such as the aerospace field,the interference caused by image noise may cause irreparable damages.Therefore,image denoising is mainly important in the image preprocessing stage,and it is also an important research direction of many experts and scholars.In recent years,Generation Adversarial Networks has made major breakthroughs in visual tasks such as image restoration a nd image style conversion.It is fully proved that the generation of the confrontation network can preserve the texture details in the image and improve the performance of image denoising in the process of alternate update and optimization.Therefore,this paper proposes to use the idea of generating confrontation to perform image denoising.Most of the existing convention denoising methods only consider the non-local self-similarity prior method(NSS)of the noise input image,and only collect similar imag e blocks from the degraded input image,as a result,the obtained prior information of the image is not accurate enough.The quality of image denoising depends largely on the input image itself,so,in order to make the prior knowledge of the image more a ccurate,this paper uses the combination of deep learning and group sparse residual constraints to effectively estimate the group sparse coefficient of the clear image.For the problem of noise removal during image restoration,this paper works as follows:(1)Using the discriminant network to judge the denoised image obtained by the generated network,measure the difference of the data distribution between the denoised image and the clear image by using the least squares loss function,and feedback the meaningful gradient to the generation network through backpropagation.The training process is more stable to avoid the problem of gradient disappearance,so that the data distribution of the generated image is as close as possible to the data distribution of the real image,and the visual signal-to-noise ratio of the denoised image is improved while having a better visual effect.(2)For the image smoothing domain,which is easy to be blurred and unclear in the denoising process,the convolution operation is performed on the image with different convolution kernels,and the obtained image features are used in the residual block of the generated network to denoise,and then reconstruct the image features into a denoised image.The content loss function is used in the feature layer to perform a simple L2 loss calculation on the features of the denoised image and the target image,so that the edge texture details of the generated image are closer to the original image.(3)Aiming at the problem of using image prio r knowledge only by using noise image,this paper proposes an improved image denoising prior model by combining convolutional neural network and group sparse residual constraint.Fusing the two NSS of the noise input image and the preprocessed image,and u sing them for image denoising to improve the performance of group sparse image denoising.