Super Resolution Using GANs and U-Nets
Using PyTorch and Deep Learning applications to increase the resolution of images.
First approach is using Generative adversarial networks (GANs). GAN approach uses a loss function, which calls another model to decide “Is this a high res image or a low res image”
This project is then improved using U-Nets and Feature Loss Applications inspired by the paper “Perceptual Losses for Real-Time Style Transfer and Super-Resolution” by Johnson et al. Instead of using MSELoss on pixel values and a CrossEntropy loss for the Critic model, we compare features displayed in the real image and the generated image.
Johnson, et al. “Perceptual Losses for Real-Time Style Transfer and Super-Resolution.” Computer Vision – ECCV 2016 Lecture Notes in Computer Science, 2016, pp. 694–711.
Daniel Diamond
data watches music travel