Image super-resolution (SR) refers to the image processing technology using computer to process a low resolution (LR) image or sequence to recover a high resolution (HR) image.
Super-resolution has important applications in video monitoring, satellite imagery and medical imaging. SR can be roughly classified into two categories: reconstructing a high resolution image from multiple low resolution images and reconstructing a high resolution image from single low resolution image.
For a low-resolution image, there may be many different high-resolution images corresponding to it. Therefore, a priori information is usually added to the normalization constraint when solving high-resolution images. In the traditional approach, this prior information can be learned through the example of several pairs of low-high-resolution images that appear. The SR based on deep learning directly learns the end-to-end mapping function of the low resolution image to the high resolution image through the neural network.