A Method for Face Hallucination Based on Generative Adversarial Network by Extending ESRGAN Architecture

Abstract

Deep learning has emerged as a pivotal approach across many scientific and industrial applications, primarily due to rapid advancements in computational power. Face hallucination, defined as enhancing the resolution of facial images, plays a significant role in computer vision applications such as facial recognition, facial feature analysis, and human identity analysis. Recently, deep generative models like Generative Adversarial Networks (GANs) have been extensively used for this purpose. However, further improvement in the precision and quality of the results is necessary. We present a novel GAN-based approach for face hallucination by expanding the Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) architecture. This research introduces a customized version of ESRGAN using the pre-trained version of VGG16 architecture but in a reduced size to achieve a good trade-off between the accuracy of the final output image and the complexity requirements. It has been demonstrated during the experiments that the suggested changes enhance the face hallucination accuracy, reaching a training data classification accuracy with PSNR: 30.30, SSIM: 0.8757 and LPIPS: 0. 0817. This performance exceeds state-of-the-art approaches, emphasizing the significance of the modifications undertaken.