For the transmission of digital video signal at low bit-rate the video data compression technique with high compression ratio is needed. The hybrid coding using BMA(Block Matching Algorithm) and DCT(Discrete Cosine Transform) is a typical technique for the purpose. Two fundamental problems arise at low bit-rates. In BMA the amount of motion vector data is excessive as the object moves with large displacement. The excessive amount of motion vector data directly lowers the quality of reconstructed image. The MCP(Motion Compensated Prediction) error images have the characteristic of line structure. DCT has low energy compaction efficiency because of the characteristic. It generates block artefacts around moving contour and inefficiently consumes the transmission data bits.
In this dissertation two methods are proposed for the improvement of the above problems. First the backward predictive BMA is proposed for removing the transmission of motion vectors. In this algorithm the motion vector of a block to be coded is estimated using the spatially adjacent blocks which are previously reconstructed. The motion vector need not be transmitted because decoder can estimate the same motion vector. Despite of the advantage, from the analysis and measurement of the proposed method, it is not applied well in the blocks containing motion boundary of deformation of object. To overcome this defect, the backward/forward adaptive BMA is proposed and applied to low bit-rate video coding algorithm. Experimental results show that the amount of motion vector data can be reduced up to about 50% of the amount of conventional BMA. In the circumstances of camera panning and zooming, it also operates well without serious increase of motion vector data.
For the effective compression of MCP error image, the ridge based model is proposed. It describes the line structured parts with three parameters: magnitude, width and smoothness. From these parameters, the energy compaction efficiency and quantization noise effect of DCT are measured. Its efficiency becomes low as the values of parameters describing width and smoothness become small. Also the quantization error spreads in the parts which have no MCP error. Based on these facts, two compression techniques of prediction error image are proposed. One is based on the removal of ridge with small parameters and the other transmission of ridge parameters. The former reduces the amount of DCT coefficient data without loss of subjective quality. The latter transmits the prediction error image without block artefacts. Two schemes are applied to the conventional hybrid coding system based on BMA and DCT. Experimental results show quite reduced block artefacts around the moving boundary and better subjective quality in overall image.
The proposed two schemes are combined for measurement of overall contribution of this dissertation to low bit-rate video coding. At 64 kbit/s the proposed video coding scheme shows about 10 - 30% reduction of transmission bit-rate with similar subjective quality.
