This Wednesday, Edward presented on CNNs for Image Super-Resolution. The abstract is presented below.
CUMIN’s paper reading group happens weekly on Wednesdays 2pm in the North Room, Department of Engineering. No prior reading of the paper is required, just show up to find out more. All are welcome!
Check out our paper reading page for information on other regular paper reading groups that are going on around Cambridge.
Convolutional neural networks (CNN) are becoming mainstream in computer vision. In particular, CNNs are widely used for high-level vision tasks, like image classification. This talk focuses on examples of using CNNs for image super-resolution (SR), which is a low-level vision task, and on practical aspects of training CNNs. U-Net architecture was published in 2015 and has been deployed in practice for image SR. RCAN (Residual Channel Attention Network) was published in 2018 and gives much superior results compared to U-Net at smaller network size. GANs have also been widely studied for image SR. Cycle GAN allows to learn image translation from low resolution space to SR space without paired data, which makes it interesting for practical applications where training data is scarce.