Artistic and semantic progressive image coding

Progressive image coding provides web users a faster full-image preview when only a small fraction of the file has been transmitted. JPEG is the most widely used progressive compression technique. However, the compressed quality is low at low bitrates and it entirely ignores the semantics of images. We propose to render different artistic visual effects and leverage semantic information in the images. The key insight is human visual system is more sensitive to luminance and semantic salient objects. Our learning based progressive coding approach learns to encode a smooth transition from grayscale images to color images. During decoding, our approach will display sharp grayscale images first instead of blurry color images, which allows viewers to preview images faster. In addition, our approach can allocate more bitrates to important objects according to segmentation masks. The approach can be extended to encode other artistic styles such as mosaic style and users can easily create their own decoding patterns. Our progressive image coding method generates clearer content than JPEG and previous learning based progressive compression method in both quantitative metrics and user studies.

Name thesis:degree_name

Master of Science in Computer Sciences

Level thesis:degree_level

Masters

Discipline thesis:degree_discipline

Computer Science

Grantor

The University of Texas at Austin

Year dc:date.issued

2019

Author dc:creator

• Cheng, Chia-Wen

Advisor dc:contributor.advisor

• Krähenbühl, Philipp

Committee member dc:contributor.committeemember

• Grauman, Kristen

• Image compression
• Progressive image coding
• Deep image compression
• Bit allocation

Language dc:language.iso

en

Identifier URI

http://dx.doi.org/10.26153/tsw/5692

OAI identifier oai:identifier

oai:repositories.lib.utexas.edu:2152/78636