Study of image compression techniques based on deep learning.

Abstract

This thesis explores various image compression techniques, with particular focus on approaches grounded in deep learning. It presents detailed analysis and comparison between traditional compression algorithms such as JPEG, JPEG 2000, and BPG and modern deep learning methods, including factorized and hyperprior models. While conventional techniques have been widely used for their balance of image quality and bitrate, recent advances in deep learning offer more effective and efficient alternatives. Neural network-based models, in particular, have demonstrated superior capabilities in achieving higher compression rates while maintaining perceptual image quality. This study underscores the strengths of these advanced techniques, establishing deep learning as a promising and powerful direction for the future of image compression.