Design and Architecture of End-to-End GANs for Image Coding

EasyChair Preprint 14315

Abstract

Image coding and compression are pivotal in managing the vast amounts of visual data generated in today's digital world. Traditional image coding methods, while effective, often fall short in terms of compression efficiency and image quality preservation. Recent advancements in Generative Adversarial Networks (GANs) offer a promising alternative by leveraging deep learning techniques to enhance image coding processes.

This paper explores the design and architecture of end-to-end GANs specifically tailored for image coding applications. We provide an in-depth analysis of the GAN framework, focusing on its three core components: the encoder, decoder, and discriminator networks. The encoder compresses the input image into a compact latent representation, while the decoder reconstructs the image from this latent space. The discriminator plays a critical role in ensuring the reconstructed image maintains high perceptual quality by distinguishing between real and generated images.

Key challenges in this approach include balancing compression efficiency with reconstruction fidelity, as well as managing the computational complexity associated with training and inference. We address these challenges through innovative network designs and training strategies, including the use of advanced loss functions and optimization techniques.

Keyphrases: Generative Adversarial Networks, compression efficiency, image coding, latent space, perceptual quality

@booklet{EasyChair:14315,
  author    = {Samon Daniel and Godwin Olaoye},
  title     = {Design and Architecture of End-to-End GANs for Image Coding},
  howpublished = {EasyChair Preprint 14315},
  year      = {EasyChair, 2024}}