Our paper, "Object-Based Image Coding: A Learning-Driven Revisit", is accepted by the IEEE ICME 2020.
Abstract: The Object-Based Image Coding (OBIC) that was exten-sively studied about two decades ago, promised a vast application perspective for both ultra-low bitrate communicationand high-level semantical content understanding, but it hadrarely been used due to the inefficient compact representationof object with arbitrary shape. A fundamental issue behindis how to efficiently process the arbitrary-shaped objects at afine granularity (e.g., feature element or pixel wise). To attack this, we have proposed to apply the element-wise masking and compression by devising an object segmentation network forimage layer decomposition, and parallel convolution-basedneural image compression networks to process masked fore-ground objects and background scene separately. All compo-nents are optimized in an end-to-end learning framework tointelligently weigh their (e.g., object and background) contri-butions for visually pleasant reconstruction. We have con-ducted comprehensive experiments to evaluate the perfor-mance on PASCAL VOC dataset at a very low bitrate sce-nario (e.g.,.0.1 bits per pixel - bpp) which have demon-strated noticeable subjective quality improvement comparedwith JPEG2K, HEVC-based BPG and another learned image compression method.
Citation:
Qi Xia, Haojie Liu, and Zhan Ma, Object-Based Image Coding: A Learning-Driven Revisit, IEEE ICME, London, UK, July 2020.