Summary: | 博士 === 國立交通大學 === 資訊科學學系 === 86 === ABSTRACTIn this dissertation, we propose several kinds of fast
searching algorithmfor various applications of Nearest Neighbour
(NN) search. The first one is a reference-point-based fast
searching method with N*(k/2 + 1) memory overhead. Here, N is
the number of codewords and k is the dimensionality of each
codeword. This method is used to accelerate the LBG codebook
generation process for Vector Quantization (VQ) design. As for
the second one, when the high/low means generated by the Block
Truncation Coding (BTC) image compression technique are to be
quantized using VQ, the computation time to search for the
nearesthigh/low mean sample can be reduced significantly by the
second approach. Thememory overhead for this approach is 2N
only. The third approach is a new approach that kicks out many
impossible candidates by a single kick-out conditionderived from
Schwarz Inequality. Due to the efficiency and simplicity of
theproposed condition, a considerable saving of the CPU time
needed to encode a data set (using a given codebook) can be
achieved. The memory overhead is as low as 1N (which is quite
competitive). The performance are demonstrated using the example
of encoding images by vector quantization (without BTC) when a
code-book is given. Finally, we introduce a fast motion
estimation method based onthe Hierarchical Use of Minkowski''s
Inequality. Experimental results and complexity analysis are
both given to show that the method outperforms many well-known
methods such as PDE, SEA and TSS.
|