Saliency Based Color Image Compression and Evaluation
Event date:  July 21, 2016 10:00 AM  to 11:00 AM

Talk Abstract:

Humans are the end users to visual media. Therefore, in order to develop an effective quantitative method for visual computing, we must take into account how humans perceive visual quality. Recently, on the basis of human visual saliency information, H. Alers et al. showed that the quality of the foreground (FG) of an image is more important than that of its background (BG) in visual quality assessment. Therefore, in image compression an adaptive bitrate allocation that favors the image foreground can be expected to increase the visual quality of decoded images.

Two approaches for JPEG image compression will be discussed. Firstly, we consider a bi-level saliency based coding based on foreground/background image segmentation. However, this approach has discontinuous visual quality at borders between FG and BG blocks. Secondly, in our multi-level saliency based approach the bitrate is adapted blockwise and offers smooth visual quality. In order to evaluate our approach by subjective image quality assessment, we performed a very large crowdsourcing study using pairwise comparisons for a set of 44 image sources at five bitrates in each case comparing 12 different coding strategies with the standard JPEG coded image. Our results show a significant improvement of subjective visual quality.


Speaker’s Bio

Sung-Hwan Jung received a Ph. D. in Electronic Engineering from Kyungpook National University, Korea in 1988. From 1992 to 1994, he was a postdoctoral research staff of the Department of Electrical and Computer Engineering at the University of California at Santa Barbara (UCSB). From 1999 to 2000, he also worked for the Colorado School of Mine (CSM) in Golden, Colorado. From 2008 to 2009, he had experience on the medical information processing at the Dental School of the University of Missouri at Kansas City (UMKC). His research interests include content-based image retrieval, steganography, watermarking, internet-based remote monitoring, medical image processing, computer vision and pattern recognition, etc.



Event location:  University of Konstanz
Z 613
Universitätsstr. 10