Deep Residual Variational Autoencoder for Image Super-Resolution

Generating a higher version from a low-resolution image is a challenging computer vision task. In recent studies, the use of generative models like Generative Adversarial Networks and autoregressive models have shown to be an effective approach. Historically, the variational autoencoders have been criticized for their subpar generative performance. However, deep variational autoencoders, like the very deep variational autoencoder, have demonstrated its ability to outperform existing models for producing high-resolution images. Unfortunately, these models require a lot of computational power to train them. Based on variational autoencoders with a custom ResNet architecture as its encoder and pixel shuffle upsampling in the decoder, a new model is presented in this study. Evaluating the proposed model with PSNR and SSIM reveal a good performance with 33.86 and 0.88 respectively on the Div2k dataset.