Thesis

New fast Kohonen algorithm: a codebook generation algorithm enhanced by neural network

In this paper, a novel codebook generation algorithm, called New Fast
 Kohonen Algorithm (NFKA), is proposed to relieve the users from the dependence on
 the previous experience that is required in the existing fast Kohonen algorithm.
 The design of NFKA algorithm is mainly based on a self-organizing Kohonen
 neural network structure. Taking advantage of this structure, NFKA can greatly save
 both the codebook construction time and the computation resources. The concept of
 codeword utility, the usefulness of a codeword, is introduced to the Kohonen neural
 network structure to trace the changes of the representativeness of each codeword.
 The calculations of weight adjustment and partition selection for merging and
 splitting are based on the utility values so that the quality of the codebook is further
 improved. We also defined and included an important measurement, "accumulated
 error", in this algorithm to decide on the positions of the new code words formed by
 partition splitting.
 Several data sets are used in the experiments to demonstrate the superiority of
 the proposed NKFA algorithm. Comparisons of reconstructed image quality are made
 between NKF A and other code book generation algorithms including the industryaccepted
 generalized Lloyd Algorithm. The comparisons are both objective and
 subjective.

In this paper, a novel codebook generation algorithm, called New Fast Kohonen Algorithm (NFKA), is proposed to relieve the users from the dependence on the previous experience that is required in the existing fast Kohonen algorithm. The design of NFKA algorithm is mainly based on a self-organizing Kohonen neural network structure. Taking advantage of this structure, NFKA can greatly save both the codebook construction time and the computation resources. The concept of codeword utility, the usefulness of a codeword, is introduced to the Kohonen neural network structure to trace the changes of the representativeness of each codeword. The calculations of weight adjustment and partition selection for merging and splitting are based on the utility values so that the quality of the codebook is further improved. We also defined and included an important measurement, "accumulated error", in this algorithm to decide on the positions of the new code words formed by partition splitting. Several data sets are used in the experiments to demonstrate the superiority of the proposed NKFA algorithm. Comparisons of reconstructed image quality are made between NKF A and other code book generation algorithms including the industryaccepted generalized Lloyd Algorithm. The comparisons are both objective and subjective.

Relationships

Items