Embedded image compression with constant bit-rate

Abstract

Prediction is an important process to explore the redundancy in lossless coding. There are many researches on the optimization of prediction scheme, globally and locally. Due to its simple structure and good performance, the predictor of JPEG-LS is used as spatial prediction scheme. In addition to the spatial predictor, we design new prediction algorithm called the inter-color prediction to explore the redundancy between color components. The high similarities between neighborhood pixels in the YDbDr color coordinate system is used to estimate the color of current pixels. From experiments for various images, it is confirmed that inter-color prediction is effective to reduce the energy of residue. The performance of entropy coding as well as the prediction is critical to compression efficiency. To improve the performance of entropy coding, many researches have been focused on how to define the contexts from neighbors. In this thesis, we design a new algorithm to increase the probability inequality of symbols of each bit-plane by multiple predictors during the bit-plane coding. We also improve the algorithm to be compatible with previous context based entropy coding. The proposed compression algorithm produces embedded-structured bitstream. It is easy to control the bitstream size on it. Through this way, we can achieve constant bitrate compression. For practical implementation, we develop a parallel prediction scheme based on color components interleaving and image partitioning. For a processing unit which consists of color components of multiple pixels, each component can be predicted from neighborhoods in parallel. The prediction scheme has side effect of random sampling which is helpful to improve the worst case performance because it lessens the local characteristics in the processing unit.