A fast image denoising method based on octblock dense block

Abstract

The invention relates to a fast image denoising method based on OctBlock dense blocks, which divides an image to be denoised into a training data set and a test data set and performs enhancement and / or noise addition respectively, and constructs a denoising network based on OctBlock dense blocks, Input the training data set for training, input the test data set for image denoising, and obtain the denoised image. The sub-module OctConv of the present invention decomposes the output feature map of the convolutional layer into features of different spatial frequencies, stores the feature maps in different groups, and shares information between adjacent positions to reduce spatial redundancy, improve model efficiency, Speed ​​up image denoising; the features extracted by each sub-module inside the OctBlock dense block will be connected to subsequent sub-modules to extract rich local features and obtain denoised images faster and better; apply residual in the network Learning, reduce network parameters, and avoid gradient disappearance and gradient explosion caused by network complexity.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a fast image denoising method based on OctBlock dense blocks. Background technique [0002] In the process of image generation and transmission, the image quality is often degraded due to the interference and influence of various noises, which will adversely affect the subsequent image processing. Therefore, in order to suppress noise and improve image quality, the image must be denoised preprocessing. [0003] Image denoising is one of the fundamental computer vision problems and has received great attention in academic research. Existing image denoising methods include traditional methods and deep learning-based methods. [0004] Traditional methods are divided into image denoising methods based on spatial domain, image denoising methods based on transform domain, image denoising methods based on image self-similarity, and image denoising methods based on sparse repr...

AI Technical Summary

Problems solved by technology

[0006] The paper "Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising" uses a deeper convolutional network for image denoising, the denoising effect is obvious, and the edge and texture features of the image are preserved; however, a deeper network structure will Leading to a large increase in parameters, which increases the difficulty of training

The paper "FFDNet: Toward a Fast and Flexible Solution for CNN based Image Denoising" downsamples the input image into multiple sub-images as the network input, and then upsamples the output sub-images to obtain the final output, which reduces the network parameters. Increasing the receptive field makes the network more efficient; however, the network only reduces network parameters by downsampling the input image as the input of the network, ignoring the problem that the feature map generated by CNN convolution also has a large amount of redundancy in the spatial dimension

Images