數位視訊壓縮之研究

Abstract

In this thesis, we studied the three methods of digital video encoder with either discrete cosine transform (DCT) or discrete wavelet transform (DWT), 2D method (DCT & DWT), 3D method (DCT & DWT) and hybrid method. The aim of the video compression is to achieve higher compression ratio with little elimination of visual quality and lower computational complexity. By exploiting the redundancies in the video sequence and the human visual response we could reduce storage of information in the video sequence and then apply the information theory (Huffman coding) which is a lossless coding method using by text compression tool such as Winzip or Winrar. 2D model with 2D-DCT is adopted by the JPEG for image compression. In application of video sequence, it is called motion JPEG. For preventing the block effects of the 2D-DCT method, DWT replaces DCT. The DWT is completed by two orthogonal or biorthogonal filter banks and then the DWT coefficients are transmitted by the SPIHT Algorithm. For exploiting the temporal redundancies, the 3D model of digital video compression is the direct method. By applying the 3D transform either with the 3D-DCT or 3D-DWT, we can transform the temporal redundancies to the frequency domain and then quantize the coefficients according to the human visual response. In the previous study, the energy of 3D-DCT coefficients is highly compacted in the hyperbolic plane, thus our mission is to quantize the coefficients outside the plane to zero for achieving higher compression ratio without lower visual quality. The quantization value could be obtained form the proposed equation. The zig-zag scan order of the 2D-DCT is not suitable for the 3D-DCT, because it can’t achieve higher compression by using variable length coding. The optimal scan order should be implemented according to the probability of non-zero coefficients. Hybrid method is also a method of eliminating the temporal redundancies of video sequence by motion estimation. It reduces the temporal redundancies by subtracting the predicted frame from the current frame, thus it is more suitable for the high motion video sequence but 3D method isn’t.