Image super-resolution techniques using deep neural networks
Meilin Guo · Applied and Computational Engineering · 2023
Super-resolution (SR) based on deep convolutional neural networks is a rapidly developing field with many real-world applications. In this paper, we examine cutting-edge super-resolution neural networks in-depth using freshly released difficult datasets to test single-image SR. We present a taxonomy that divides existing techniques into six categories, including upsampling, residual, recursive, dense connection, attention-based, and loss function designs. This taxonomy is applicable to deep learning-based SR networks. The comprehensive analysis shows that in the past few years, the accuracy has increased steadily and rapidly, while the complexity of the model and the accessibility of large-scale information have also increased accordingly. It has been noted that the present techniques have greatly outperformed the past techniques that were indicated as benchmarks. On this basis, this paper will put forward some suggestions for future research.