Training Techniques for GANs in Low Bitrate Image Coding

EasyChair Preprint 14313

Abstract

Generative Adversarial Networks (GANs) have emerged as a powerful tool in enhancing image compression, especially at low bitrates where traditional methods often struggle to maintain visual quality. This paper explores advanced training techniques for GANs in the context of low-bitrate image coding. We begin by discussing the fundamentals of image compression and the unique advantages GANs offer in this domain. The focus then shifts to the specific architecture and training strategies that optimize GAN performance, including the design of the generator and discriminator networks, the formulation of loss functions, and the implementation of regularization techniques to ensure stability and prevent mode collapse.

Key training techniques such as progressive training, curriculum learning, and transfer learning are examined for their effectiveness in enhancing the quality of reconstructed images. Additionally, we address the challenges inherent in GAN training, such as instability and computational complexity, and propose solutions to mitigate these issues. Evaluation metrics like PSNR, SSIM, and perceptual quality scores are used to assess the performance of the GAN-based approach compared to traditional compression methods.

Through detailed case studies and comparative analyses, this paper highlights the significant improvements in visual quality and rate-distortion performance achieved by using GANs for low-bitrate image coding. Finally, we discuss future directions for research, including potential advancements in GAN architectures and the integration of GANs with other AI techniques to further enhance image compression capabilities. This study underscores the transformative potential of GANs in digital image coding, offering a path forward for more efficient and visually appealing image compression solutions.

Keyphrases: Generative Adversarial Networks, Image Compression, Low Bitrate Image Coding, adversarial training, deep learning

@booklet{EasyChair:14313,
  author    = {Samon Daniel and Godwin Olaoye},
  title     = {Training Techniques for GANs in Low Bitrate Image Coding},
  howpublished = {EasyChair Preprint 14313},
  year      = {EasyChair, 2024}}