Limitations of learning-based Super-Resolution

In recent years, studies on Super Resolution (SR) have increased because of the introduction of 4K and 8K resolution. Super-Resolution image Reconstruction (SRR) and Learning-Based Super-Resolution (LBSR) are typical SR techniques. SRR performance limitations have been mentioned in several studies. However, practical performance of LBSR has not been clarified. If LBSR is used on a video device such as a television, it is necessary to assess the degree of image quality improvement in various types of images inputs to the device. Furthermore, the conventional Bicubic and Lanczos interpolation techniques are in general use as image enlargement techniques. LBSR is expected to perform better than such techniques. Accordingly, we conduct an objective evaluation of the practical performance of LBSR to determine its limitations. Applying LBSR to multiple images, we show that there are limitations to its ability to improve resolution, and that there is no benefit in improvement of resolution as compared with images enlarged using conventional interpolation techniques. Furthermore, real-time processing is essential for SR to apply to a video. However, the latest research shows that it is still difficult to process in real time. We discuss challenges in applying LBSR to a video with reference to the latest research. We evaluate the limitations of LBSR based on these issues.