The region adaptive image coding is expected to be used to encode the image segments with an arbitrary shape, such as the segmented regions in the region-based image coding system and the VOPs (Video Object Planes) in the MPEG-4. The existing methods for coding these regions are largely classified into two categories. One is the region-based coding methods, GOT (General Orthogonal Transform) by Gilge$^{[3]}$ and the iterative coding by Kaup and Aach$^{[5]}$, the other is the block-based coding methods, the mean stuffing, the MBP (Macro Block Padding), and the SA-DCT (Shape Adaptive DCT). They are a trade-off in terms of the hardware complexity and the subjective quality. An goal of this thesis is to propose an algorithm based on the block-based coding scheme which is simple to implement and has high performance. This dissertation presents a new approach, a one-dimensional extension-interpolation (1D-EI)/2D-DCT, for the region adaptive image segment coding to satisfy the goal mentioned above. In addition, we propose three improvement methods of the 1D-EI/2D-DCT and apply the 1D-EI/2D-DCT to coding of the motion compensated error in the region-based coding system and to coding of the significant coefficients in the wavelet domain.
In the proposed 1D-EI/2D-DCT, the image region is spanned into an 8×8 rectangular block and its intermediate luminances are interpolated. After all luminances in the 8×8 block are obtained from pixels in the region, they are transformed by the 2D-DCT. The 1D-EI/2D-DCT is compared with conventional ones, such as the SA-DCT and the GOT, etc., under two aspects: peak signal-to-noise ratio (PSNR) and hardware complexity. Simulation results show that the performance of the proposed method is superior to that of the conventional ones.
Since the 1D-EI/2D-DCT is the 1D technique using a 1D interpolation in horizontal and vertical directions i.e., it uses the 1D method to encode two dimensional (2D) arbitrarily-shaped (AS) image segments, there is a difference in relative correlation amounts between the horizontal and the vertical directions. It results in degrading the performance of the 1D-EI/2D-DCT. We analyze this problem of the 1D-EI/2D-DCT and devise a new non-symmetric zig-zag scanning, a repetitive 1D-EI/2D-DCT, and a 2D-EI/2D-DCT to reduce this effect.
Finally, we apply the EI/2D-DCT to two image coding systems, the region-based image coding and the wavelet coding. We analyze a distribution of motion compensated (MC) error in the region based coding system and propose a new MC error coding method. The proposed MC error coding is a region adaptive method using the contour information of the regions and AS image segment coding. In the case of the wavelet coding, the EI/2D-DCT is applied to code significant coefficients in the wavelet domain. Because it is proved that the significant coefficients have some correlation each other by experiments, it is effective that they are transform-coded by AS image segment coding using the EI/2D-DCT.
