An Analytical Study of Hyperparameters and Architectural Impacts for Image Super-Resolution

Abstract

Super-resolution is particularly helpful for applications like larger computer screens, portable electronics like cellphones and cameras, and high-definition television sets. It plays a crucial role in improving visual perception by transforming low-resolution images into clearer, high-resolution ones. The integration of deep learning methodologies has demonstrated significant promise in mitigating the constraints of conventional approaches. Three traditional deep learning networks are used in this experiment to perform picture super-resolution tasks. This experiment uses the well-known Very Deep Super-Resolution (VDSR) network as a reference and evaluates the effects of different hyperparameters on the results. The experiment then presents the Convolutional Neural Network with Recurrent Neural Network (CNN+RNN) and Generative Adversarial Network (GAN). The benefits of these various network topologies are contrasted, and the differences and enhancements provided by each configuration are examined. Lastly, subjective comparison graphs are provided and objective measurements are used to assess the outcomes. The quality of the experimental results is directly impacted by the network structure's complexity.