Research of image compression algorithms using neural networks

Abstract

The article presents the results of the study of image compression algorithms based on neural networks. Classical compression methods, such as JPEG, PNG, GIF, TIFF, are analyzed, and the advantages of neural network methods, in particular the use of an autoencoder, a variational autoencoder, and generative adversarial networks, are highlighted. It is concluded that the main advantages of neural network methods are the preservation of a high level of textures and details at low bitrates, as well as the ability to work with high-quality images, although this requires significant computing resources. A comparative analysis of classical compression algorithms, such as JPEG, with new approaches based on neural networks is carried out using the example of an autoencoder, and the prospects of neural networks in solving the problem of data compression are assessed. The main emphasis is placed on the analysis of the quality of image restoration and the level of compression using different neural network settings. A mathematical model is presented that describes the principle of operation of an autoencoder and shows how a neural network encodes and restores images using latent space. To achieve the best reconstruction quality, a hybrid loss function was used, which consists of three components: perceptual loss based on VGG16, SSIM loss, and MSE loss. A modular software system was developed using the Python programming language to conduct experiments. The software includes a graphical interface, a compression module for performing image encoding and decoding operations using an autoencoder model, and a quality assessment module for calculating the main quality metrics (PSNR and SSIM). It was found that traditional image compression methods demonstrate high efficiency, but are more prone to generating artifacts, especially at high compression levels, than neural network methods. As a result of the research, it was found that the autoencoder model can encode and decode images with minimal loss of quality, on a par with JPEG, but is inferior to classical algorithms in speed (1.6 seconds per image versus 0.02 for JPEG) and compression ratio (the model provides a reduction in file size by 11–18%). It is concluded that without reducing the need for computational resources, neural network compression methods will not be able to replace classical methods