Image compression based on machine learning technics

Abstract

Image compression is required in a variety of situations. Unfortunately, artifacts emerge whenever a lossy compression method is utilized. Image artifacts, which are created by compression, tend to remove high-frequency detail while also adding noise or minor image structures in some circumstances. Several technologies have been employed to refine picture compression methods in order to lessen their impact on the human or software user's ability to appreciate images. We attempted to investigate ways of combining machine learning methods with one of the deep learning approaches, specifically convolution networks for picture compression, in this master's thesis. Indeed, for this assignment, we employed a technique based on Run Length Encoding and a quantification vector combined with a Convolution Autoencoder, which was efficient and based on principles. Our method produces more detailed images than generic methods and other machine learning methods combined with deep learning.