ABSTRACT
Over the last two decades, great improvements have been made in image and video compression techniques driven by a growing demand for storage and transmission of visual information. State-of-the-art JPEG 2000 and MPEG-4 AVC/H.264 are two examples that significantly outperform their previous rivals in terms of coding efficiency. However, these mainstream signal-processing-based compression schemes share a common architecture, namely, transform followed by entropy coding, where only the statistical redundancy among pixels is considered as the adversary of coding. Through two decades of development, it has been becoming difficult to continuously improve coding performance under such architecture. Specifically, to achieve high compression performance, more and more modes are introduced to deal with regions of different properties in image and video coding. Consequently, intensive computational efforts are required to perform mode selection subject to the principle of rate-distortion optimization. At the same time, more and more memory-cost context models are utilized in entropy coding to adapt to different kinds of correlations. As a result, small improvements in coding efficiency are accomplished with great pain of increased complexity in both encoder and decoder.
