Share Email Print

Proceedings Paper

Image super-resolution reconstructing based on generative adversarial network
Author(s): Jing Nan; Lei Bo
Format Member Price Non-Member Price
PDF $17.00 $21.00

Paper Abstract

Image super-resolution refers to the technique of reconstructing a high-resolution image by processing one or more complementary low-resolution images. It is widely used in medical imaging, video surveillance, remote sensing imaging and other fields. The learning-based super-resolution algorithm obtains a mapping relationship between the highresolution image and the low-resolution image by learning, and then guides the generation of the high-resolution image according to the obtained mapping relationship. The generative adversarial network (GAN) is composed of a generative network model and a discriminator network model, and the two play each other until the Nash equilibrium is reached, and the texture information and the high-frequency details of the downsampled image can be restored based on the super-resolution method of generative adversarial network. However, super-resolution algorithms based on generative adversarial network can only be used for one kind of magnification time, and the versatility is insufficient. Despite convolutional neural networks has achieved breakthroughs in accuracy and speed of traditional single-frame superresolution reconstruction, and can achieve a higher peak signal-to-noise ratio (PSNR). Most of them use Mean Square Error (MSE) as the minimum optimization objective function, so although a higher peak signal-to-noise ratio can be achieved, when the image downsampling factors is higher, the reconstructed image will be too smooth, lack highfrequency details and perceptually unsatisfy in the sense that they fail to match the fidelity expected at the higher resolution. When dealing with complex data of real scenes, the model's representation ability is not high; and the generative adversarial network training is very unstable, seriously affecting the model training process. This paper is based on generative adversarial network, improving the network structure and optimizing the training method to improve the quality of generating images. The following improvements have been made to the generator model: the multi-level structure is used to enlarge the image step by step, so that the model can simultaneously generate multiple images with a larger scale, and also ensure that the image obtained at a larger magnification has higher quality; ResNet model is improved by recursive learning and residual learning, and the batch normalization structure in the model is removed. On the basis of ensuring the image quality, the efficiency of the model is effectively improved. The recursive and residual learning methods can effectively improve the feature expression ability of the model, and thus significantly improve the quality of the generated image. The Expand-Squeeze method is proposed to generate images. The basic idea is to expand the dimension of the last layer of the convolution layer of the model. In this way, more context information is obtained, and then the image is generated by using the 1x1 convolution kernel. The Expand-Squeeze method can effectively reduce the checkerboard effect and improve the quality of the generated image to some extent. This paper improves the discriminator network loss function. Measure the similarity between generated image and real image by introducing Wasserstein distance. The loss function proposed consists of two parts: the loss function of resistance and the loss function of content. The experimental results verify that the improved generation of the generative adversarial network can effectively improve the quality of the generated image and effectively improve the stability of the model training.

Paper Details

Date Published: 18 December 2019
PDF: 8 pages
Proc. SPIE 11342, AOPC 2019: AI in Optics and Photonics, 113420F (18 December 2019); doi: 10.1117/12.2547435
Show Author Affiliations
Jing Nan, Huazhong Institute of Electro-Optics (China)
Lei Bo, Huazhong Institute of Electro-Optics (China)

Published in SPIE Proceedings Vol. 11342:
AOPC 2019: AI in Optics and Photonics
John Greivenkamp; Jun Tanida; Yadong Jiang; HaiMei Gong; Jin Lu; Dong Liu, Editor(s)

© SPIE. Terms of Use
Back to Top
Sign in to read the full article
Create a free SPIE account to get access to
premium articles and original research
Forgot your username?