Browsing by Author "Mykkänen, Arttu"

Lossy image compression algorithms are used everywhere, ranging from general photography to natively running applications and the web. The most important question regarding a lossy image compression algorithm is arguably its rate-distortion performance. However, recent research has emphasized the fact that subjective experience of image quality can come at the expense of added or more pronounced distortions for low bit rates. The main focus of this thesis is to form a comprehensive view into standard lossy image compression, and to compare both classical and neural lossy image compression methods in order to inform their selection and use. Moreover, we focus on image quality in addition to traditional rate-distortion considerations. Various lossy image compression algorithms are discussed in some detail for a comprehensive view. Furthermore, we compare them using the most widely used objective IQA measures, the PSNR, the SSIM, the MS-SSIM, as well as a crowdsourced subjective image quality comparison questionnaire. The results indicate that neural methods easily dominate with respect to both objective and subjective scores. However, rank ordering reveals some discrepancies, indicating that high objective measurement scores do not always align with subjective experience. The thesis gives detailed explanations as well as objective and subjective scores for eight different classical or neural lossy image compression algorithms: JPEG, JPEG2000, WebP, BPG, VVC intra coding, HiFiC, the Coarse-to-fine hyperprior model, and iWave++.