Comparison of Learned Image Compression Methods and JPEG

Media such as audio, images, and videos, which occupy significant storage space in digital environments, are often compressed to save space, particularly outside professional settings. The compression process aims to achieve space savings while maintaining an acceptable level of quality. Many standard image compression algorithms are based on the principle of eliminating data that the human eye cannot perceive and selectively removing certain data after performing spatial transformation on the obtained image. In recent years, compression algorithms utilizing autoencoders or generative adversarial network (GAN) architectures have emerged. These algorithms fundamentally aim to reduce the dimensionality of data and create a representation of it. The dimension reduction stage used is considered equivalent to the compression process. In this study, the performance measurement of three different models using autoencoders and GANs is conducted, and the results are compared with the performance of the JPEG algorithm in terms of speeds and ratio. Upon examining the results, it is observed that learned image compression methods are catching up with JPEG in terms of quality. The methods achieve better results in terms of compression ratio compared to JPEG, but they operate much slower than JPEG in terms of processing time and have some artificial artifacts.