Comparison of Four Approaches of Image Compression for Wireless Communication

Abstract

With the advent of modern wireless communication standards, it becomes a common scenario that images and videos are transmitted more often in our day-to-day applications. However, images and videos consume a large bandwidth if transmitted uncompressed. In many cases after compression and transmission through channels, the quality of the images and videos often gets deteriorated. Now, some applications require a considerable amount of image quality for better understanding and interpretation, e.g., transmission of medical images. The present standards have certain limitations in that, particularly if the noise is associated with images. With this paper we compared four image compression approaches JPEG, autoencoder, VGG, and vision transformer (ViT) for standard images and standard images with introduced Gaussian noise and CIFAR10 datasets. This paper consists of two parts, the first part gives general overview of JPEG which are still in use, being a benchmark for every compression algorithm since they are the foundation of image compression algorithms and then recent trends like compression using deep learning which includes autoencoder, neural network-based compression like VGG and the transformer-based compression like ViT which are trending and are giving more promising results. The second part consists of the comparison of these four approaches, calculating their MSE, PSNR and Compression Ratio using CIFAR10 datasets, standard images and standard image with introduced Gaussian noise to get better and promising results of image compression maintaining its quality. Thus, among all four approaches, ViT and VGG give the best compression ratio for standard images and CIFAR datasets, respectively.