Image Super-Resolution Reconstruction Technology Based on Deep Learning

抄録

The traditional super-resolution method has limited ability of feature extraction and feature expression, which cannot meet the requirements of high quality image in practical application. This paper mainly applies the relevant theories of deep learning to image super-resolution reconstruction technology. By comparing three classical network models used for image super-resolution (SR), finally a generative adversarial network (GAN) is selected to implement image super-resolution, which is called SRGAN. SRGAN consists of a generator and a discriminator that uses both perceived loss and counter loss to enhance the realism of the output image in detail. Compared with other algorithms, although the improvement of PSNR and SSIM values of the SGRAN network obtained by the final training is not obvious, the output high-resolution images are the best in the subjective feelings of human eyes, and the reconstruction effect in the image details is far higher than that of other networks.