Image Super-Resolution Based on Additional Self-Loop Supervision

Relying on the prevalence of deep neural networks, only one image super-resolution (SR) research has taken a big step forward. Many network models use the difference between the real image and the generated super-resolution image as a loss function, ignoring the difference between the small-scale image and the super-resolution generated image. We adopt an end-to-end network model with two-way supervision, which can ensure that the image content is similar in both large and small scales. Only real natural images are used for supervision. As the scale of the image becomes larger, the difference becomes larger, and the effect of correcting image details becomes smaller. Adding the loss between the input image and the degraded image of the generated image can not only keep the content of the image similar on a small scale, but also ensure the image details. Compared with several state-of-the-arts, our method obtains the best results.