Choosing Machine Learning Methods for Image Compression

Abstract

The article deals with the comparison and selection of neural network algorithms within an image compression framework based on machine learning. The image compression framework uses convolutional neural networks, autoencoders, and generative adversarial neural networks as core image processing procedures. We compare convolutional neural networks and generative adversarial neural networks when filling trivial image regions when using the compression framework. The trivial image regions do not contain critical data, so we do not save the trivial image regions to an archive file. Instead, we fill the trivial image regions with high quality fake data when decompressing. We perform computational experiments comparing the efficiency of various neural network algorithms in filling trivial regions in natural images within the image compression framework. We evaluate the quality of fake data within image trivial regions both visually and numerically. We choose the generative neural network that has an advantage and is promising for use in the image compression framework based on the machine learning.