以抽樣激發配合線性預測所設計的畫面內視訊編碼法

Abstract

提出一種在一維空間內的畫面內預測方法, 針對畫面內較複雜和有重複紋理的區域,以抽樣激發配合線性預測的方式進行處理, 相較於AVC/H.264 的畫面內預測方法, 在高畫質的情況時, 此方法有較佳的壓縮效益

In this thesis, we propose a new intra-prediction method for very high quality image coding. Unlike many new image coding standards, such as the intra coder of AVC/H.264 or JPEG-XR, which apply 2-D spatial predictions to remove correlation in image data, the proposed technique converts 2-D image signals to 1-D signal using Hilbert curve scan patterns before predictive coding. A linear filter is used to estimate the predictor of the 1-D signal. The prediction errors are non-uniformly down-sampled using a closed-loop optimization process, and used as the excitation signal of the predictor model. The predictor can then be constructed by using a synthesis filter and the coded excitation signal. The error residuals between the original image signal and the reconstructed predictor is then computed and coded into image bitstreams. For residual coding, 1-D integer cosine transform is used to further compact the energy in residuals. After transform coding, arithmetic coding on the predictor description and the residuals are applied. From the experiments, the proposed intra-prediction method has much better prediction quality compares to the intra prediction method in AVC/H.264. In particular, the technique performs well for image areas with complex repeated textures. Since current CAVLC/CABAC coders in H.264 are not suitable for very high bitrate coding, some modifications of CABAC is also proposed in this thesis to improve entropy coding efficiency. The proposed intra-coding method is integrated into JM16.1, the reference implementation of AVC/H.264, as a new coding mode and the experimental results shows that the proposed techniques are very promising for future applications.