Research on Image Super-resolution Restoration Algorithm Based on High Order Degradation Model
Ziting Gong, Yu Xu, Hongyu Liu · 2023
Appropriate training of the design is crucial for real-world super-resolution Low-Quality (LQ) images, which are difficult to obtain matched High-Quality ground truth images (HQ), or realistic degraded LQ observations from synthetic photography. Most of the recent work is concerned with modelling degradation using artificial or estimated degradation parameters, but these metrics cannot model complex types of real-world degradation, leading to limited quality improvement. We have demonstrated excellent performance of the image Super-Resolution (SR) model based on deep convolutional neural networks in restoring basic High resolution (HR) images from Low resolution (LR) images obtained from pre-defined downscaling methods. It is generally believed that when the assumed degradation model departs from the single image sequence (SISR) method in the real image, it performs poorly. Other factors are considered in many degradation models, such as Blur, but they are still insufficient to cover the different degradations of real-world images. In the case of high order degradation, we introduce the synchronous design of the process of obtaining the core of the de-degradation pool and modeling the noise. In order to better simulate the deterioration of the complex real world, high-resolution learning gives a realistic picture.