Image denoising method based on residual convolutional self-coding network

Abstract

The invention discloses an image denoising method based on a residual convolutional self-coding network, and aims to overcome the problems of limited feature extraction capability of a traditional shallow linear structure and weak generalization capability of an existing image denoising model based on deep learning and the like. A residual convolutional self-encoding block composed of a residual block, a batch normalization layer and a self-encoder is used as a basic denoising network structure, and the multifunctional denoising residual convolutional self-encoding neural network is provided.The image denoising method disclosed by the invention not only has blind denoising capability, but also can remove noise different from a training set in type while keeping higher denoising quality and denoising precision.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to an image denoising method based on a residual convolutional autoencoder network in image denoising. Background technique [0002] Traditional image denoising models can be divided into four categories based on spatial domain, transform domain, sparse representation and natural statistics. Among them, the representative methods are: the median filtering method based on the spatial domain, which ignores the characteristics of each pixel, and the image after denoising will appear more serious blur; BLS-GSM based on the transform domain, this method The method can retain part of the original information of the image while denoising; based on the sparse representation of NLSC, this method takes a long time to calculate denoising; based on natural statistics of BM3D, this method can only filter out a specific noise. [0003] In order to overcome the limitations of traditional de...

AI Technical Summary

Problems solved by technology

DCAENN based on self-encoding network can effectively remove the noise of chest X-ray images, but it can only remove known noise and has no generalization ability; SSDAs based on stacked sparse denoising self-encoding network, through multiple SSDAs phase The combination of sparse coding and the deep neural network pre-trained by the denoising self-encoder is used for image denoising, but SSDAs rely more on supervised training and can only remove the noise that appears in the training set; based on the residual convolutional neural network DnCNN, using residual learning and batch normalization to speed up the training process, the DnCNN model can handle Gaussian denoising with unknown noise levels, but cannot remove noise with unknown types; CGAs based on conditional generative adversarial networks combine trained networks and The combination of sharpness detection network to guide the training process, CGAs reduces the training difficulty of generative confrontation network, but denoising is easy to lose feature data at the same time

Images