Using machine learning to select and optimise multiple objectives in media compression

Abstract

The growing complexity of emerging image and video compression standards means additional demands on computational time and energy resources in a variety of environments. Additionally, the steady increase in sensor resolution, display resolution, and the demand for increasingly high-quality media in consumer and professional applications also mean that there is an increasing quantity of media being compressed. This work focuses on a methodology for improving and understanding the quality of media compression algorithms using an empirical approach. Consequently, the outcomes of this research can be deployed on existing standard compression algorithms, but are also likely to be applicable to future standards without substantial redevelopment, increasing productivity and decreasing time-to-market. Using machine learning techniques, this thesis proposes a means of using past information about how images and videos are compressed in terms of content, and leveraging this information to guide and improve industry standard media compressors in order to achieve the desired outcome in a time and energy e cient way. The methodology is implemented and evaluated on JPEG, WebP and x265 codecs, allowing the system to automatically target multiple performance characteristics like le size, image quality, compression time and e ciency, based on user preferences. Compared to previous work, this system is able to achieve a prediction error three times smaller for quality and size for JPEG, and a speed up of compression of four times for WebP, targeting the same objectives. For x265 video compression, the system allows multiple objectives to be considered simultaneously, allowing speedier encoding for similar levels of quality.