Abstract
The use of images and videos in this era is almost everywhere. There is a high demand to produce high-resolution images, but sometimes due to limitations of image capturing devices, we cannot produce a high-resolution image. The task of increasing image resolution, also known as the Single-Image Super-Resolution (SISR) problem, has gained popularity in the field of deep learning. The goal of this project is to present an efficient framework, Hybrid SRGAN, for SISR using Generative Adversarial Networks (GANs). One of the major challenges for increasing image resolution is to recover the image details. Specifically, we simplified the baseline Super-Resolution Generative Adversarial Network (SRGAN) model while maintaining accuracy at the same level. We removed batch normalization layers and resized low-resolution (LR) images using bicubic interpolation to avoid up-sampling in the Generator. Avoiding batch normalization improves training speed and generalization for different types of images. Resizing input images using bicubic interpolation avoids unnecessary computation for up-sampling features during training. To further improve the model performance, we applied the idea of Progressive Growing GANs (PGGAN) to train the proposed Hybrid SRGAN. To evaluate the efficiency and accuracy of our model, we used the DIV2K dataset containing a variety of images in different domains, such as the environment, people, and scenery. Compared to the baseline SRGAN model where the average Peak Signal to Noise Ratio (PSNR) was 20.44 dB (decibels), and the Structural Similarity Index Measure (SSIM) was 0.6384, the Hybrid SRGAN model achieved 24.45 dB PSNR and 0.7097 SSIM for 2000 epochs. As for efficiency, for 2000 epochs, the time required to train Hybrid SRGAN was 19.3% faster than that required to the baseline SRGAN model.