This repository showcases implementations of two advanced deep learning models for image super-resolution: Fast Super-Resolution Convolutional Neural Network (FSRCNN) and Super-Resolution Generative Adversarial Network (SRGAN). Aimed at upscaling low-resolution images with high fidelity to the original, these models leverage different techniques for enhancing image resolution. The repository includes FSRCNN.py and SRGAN.ipynb as the primary implementation files.
The FSRCNN model is encapsulated in FSRCNN.py. It acts as a baseline for performance comparison, utilizing a series of convolutional layers to improve low-resolution images. Its architecture comprises feature extraction, shrinking, non-linear mapping, expanding, and deconvolution layers. Optimized for the DIV2K dataset, this model has been adjusted to suit the specific input tensor requirements of the dataset.
The SRGAN.ipynb file implements the SRGAN model. This model is based on a generative adversarial network framework and is designed to upscale low-resolution images into high-resolution ones. The generator in SRGAN focuses on upscaling the image, while the discriminator differentiates between the generated high-resolution images and real high-resolution images. This adversarial process, combined with a specialized loss function that includes content and adversarial loss, enables SRGAN to produce images that are not only high in resolution but also excel in perceptual quality.
Both models are trained and tested using the DIV2K dataset, which contains a variety of 2K resolution images. This dataset offers both low-resolution (downsampled) and high-resolution versions of images, making it an ideal resource for training and testing super-resolution models.
The models are evaluated using the PSNR metric, which measures the quality of reconstructed images. In our experiments, SRGAN achieved a significantly higher average PSNR value of 28.064 dB, compared to FSRCNN's 15.9588 dB. This indicates the superior capability of SRGAN in generating high-resolution images that are closer in quality to the original images, demonstrating its effectiveness in both upscaling and enhancing image resolution.