Optimizing Generative Adversarial Networks for Low-Resolution Image Enhancement

Current high-resolution camera and video systems require expensively complex equipment and an excessive amount of digital storage to function, effectively limiting their practicality and availability. The research seeks to address this issue by optimizing Generative Adversarial Networks (GANs) for image super-resolution using an evolutionary-based scheme for successive network modification. The capability of GANs to selectively enhance the resolution of a desired image without increasing costs is extremely significant for a substantial range of industries and practices. The network's capabilities are expressed in terms of produced image quality and network metrics, with preference given to increases in image quality. Using the DIV2K dataset as common input, the highest performing network achieved a comparative increase of 2.22 dB PSNR (29.1%) over the base model, the Super-Resolution Generative Adversarial Network (SRGAN). The most notable contribution to network enhancement with regards to image quality was caused by removing the batch normalization layers from the Generator network. Network performance with regards to subjective image quality was most affected by the inclusion of a second convolutional layer in each residual block of the modified SRGAN Generator. Possible applications of the improved system include enhancing images of license plates for traffic monitoring systems and improving still images from body camera footage to reveal crucial details. Beyond the scope of surveillance, resolution-enhancing GANs may be applied to develop media content or national defense capabilities.