Image restoration method based on convolutional neural network with symmetric cross layer connection

Abstract

The invention discloses an image restoration method based on a convolutional neural network with symmetric cross layer connection. The method comprises the following steps: training data are preprocessed; a fuzzy training image is generated automatically; a neural network is built; the neural network is trained; and restoration effects are verified and parameter adjustment is carried out. On the basis of a damaged image, a clearer original image can be restored, and the effectiveness of methods such as traditional image denoising, image super resolution and image completion is enhanced. Besides, through adding symmetric cross layer connection to the convolutional neural network, a deeper neural network can be optimized more easily, the model generalization ability is improved, the detailsof an underlying image are kept at the same time, and a better restored image can be obtained. High-efficiency and distinct image restoration is realized, and the practical value is high.

Description

technical field

[0001] The invention relates to the field of image restoration, in particular to an image restoration method based on a convolutional neural network (Convolutional Neural Network, CNN) with symmetric cross-layer connections. Background technique

[0002] With the continuous rapid development of information technology, various fields are generating various types of image data at an astonishing speed every day. In the process of acquiring and disseminating a large amount of image data, images are often compressed, reduced in resolution, or disturbed by non-artificial noise. How to restore the original image from the damaged image as realistically as possible has become a very important task. Important issues. With the increasing popularity of more and more mobile devices with camera functions, such as cameras, mobile phones, tablets, etc., and the rise of social networks, people have more and more ways to obtain images, which further promotes the rapid growth ...

Examples

Embodiment 1

[0095] This embodiment describes image denoising, including the following parts:

[0096] 1. First, on the BSD natural image data set, a total of 500,000 image blocks of 50*50 are randomly intercepted, and Gaussian noise with a mean value of 0 and a standard deviation of σ is added to the image (σ takes 10, 30, 50, and 70 respectively Four experiments were performed). Normalize the image after adding noise and the image without adding noise. Thus, a data set with a size of 500,000 is obtained, in which 450,000 image blocks are used as the training set and 50,000 as the verification set.

[0097] 2. Construct a 30-layer convolutional neural network with cross-layer connections, use the ADAM algorithm to train on the constructed data set, and set the learning rate to 1e-4 uniformly, and use the verification set to verify the parameter pre-training effect in each training round. Finally, a converged network is obtained.

[0098] 3. Test on the original image of the BSD200 test...

Embodiment 2

[0103] This embodiment describes image super-resolution, including the following parts:

[0104] 1. First, on the BSD natural image data set, a total of 500,000 image blocks of 50*50 are randomly intercepted. For each image block, the linear interpolation algorithm is used to shrink by s times (s is 2, 3 and 4 for three experiments), On the basis of the reduced image, a low-resolution image with the same size as the original image but with low definition is obtained by s-fold enlargement. Combine the low-resolution image with the original image to obtain a data set with a size of 500,000, of which 450,000 image blocks are used as the training set and 50,000 as the verification set.

[0105] 2. Construct a 20-layer convolutional neural network with cross-layer connections, use the ADAM algorithm to train on the constructed data set, and set the learning rate to 1e-4 uniformly, and use the verification set to verify the parameter pre-training effect in each training round. Fina...

technical field

[0001] The invention relates to the field of image restoration, in particular to an image restoration method based on a convolutional neural network (Convolutional Neural Network, CNN) with symmetric cross-layer connections. Background technique

[0002] With the continuous rapid development of information technology, various fields are generating various types of image data at an astonishing speed every day. In the process of acquiring and disseminating a large amount of image data, images are often compressed, reduced in resolution, or disturbed by non-artificial noise. How to restore the original image from the damaged image as realistically as possible has become a very important task. Important issues. With the increasing popularity of more and more mobile devices with camera functions, such as cameras, mobile phones, tablets, etc., and the rise of social networks, people have more and more ways to obtain images, which further promotes the rapid growth ...

Examples

Embodiment 1

[0095] This embodiment describes image denoising, including the following parts:

[0096] 1. First, on the BSD natural image data set, a total of 500,000 image blocks of 50*50 are randomly intercepted, and Gaussian noise with a mean value of 0 and a standard deviation of σ is added to the image (σ takes 10, 30, 50, and 70 respectively Four experiments were performed). Normalize the image after adding noise and the image without adding noise. Thus, a data set with a size of 500,000 is obtained, in which 450,000 image blocks are used as the training set and 50,000 as the verification set.

[0097] 2. Construct a 30-layer convolutional neural network with cross-layer connections, use the ADAM algorithm to train on the constructed data set, and set the learning rate to 1e-4 uniformly, and use the verification set to verify the parameter pre-training effect in each training round. Finally, a converged network is obtained.

[0098] 3. Test on the original image of the BSD200 test...

Embodiment 2

[0103] This embodiment describes image super-resolution, including the following parts:

[0104] 1. First, on the BSD natural image data set, a total of 500,000 image blocks of 50*50 are randomly intercepted. For each image block, the linear interpolation algorithm is used to shrink by s times (s is 2, 3 and 4 for three experiments), On the basis of the reduced image, a low-resolution image with the same size as the original image but with low definition is obtained by s-fold enlargement. Combine the low-resolution image with the original image to obtain a data set with a size of 500,000, of which 450,000 image blocks are used as the training set and 50,000 as the verification set.

[0105] 2. Construct a 20-layer convolutional neural network with cross-layer connections, use the ADAM algorithm to train on the constructed data set, and set the learning rate to 1e-4 uniformly, and use the verification set to verify the parameter pre-training effect in each training round. Fina...

Images