論文の公開元へResearch on Image and Video Coding Algorithms for Compressive Imaging
概要
The traditional camera based on a hundred-year-old sampling theorem developed by Whittaker–Nyquist–Kotelnikov–Shannon has resulted in a massive problem of redundant data in image and video applications, which oversamples signal twice higher than information rate. It necessitates the use of complex lossy coding algorithms to reduce redundancy. However, the most recent coding algorithms are going far beyond coding efficiency; for instance, improving coding performance by 20% would cost roughly 50% more complexity and resources, which is still a significant issue today. A new camera architecture based on block-based compressed sensing (CS) has recently gained popularity because it offers lower sampling costs and produces far less amount of raw data. Meanwhile, it is sufficient to represent the original content accurately. CS is based on the Johnson–Lindenstrauss lemma, which deals with low-distortion embedding of points from high to low dimensions via random projection, resulting in a compressed vector. It theoretically eliminates the need for coding algorithm. However, the recent studies found that raw data from the CS camera is still redundant in the form of linear combination, potentially necessitating additional coding to reduce redundancy. This thesis presents a new sensing matrix that outperforms existing sensing matrices in data acquisition performance and speed at low sampling rates while dramatically improving image quality. Furthermore, a newly developed data structure of a block-based CS camera called data cube is introduced, making coding raw CS data easier. Simplified image and video coding algorithms for compressive imaging, both vector-based and data cube-based, are introduced in software and hardware, including intra-prediction, inter-prediction with quantization, and entropy coding to improve bitrate reduction performance.
