چکیده:
A main problem in vector quantization (VQ) is codebook designation. The
traditional method used for VQ codebook generation, is the Generalized Lloyd
Algorithm (GLA). The efficiency of the GLA algorithm is hardly dependent on the
initial codebook selection. But, GLA algorithm usually gets trapped into local
minimum of distortion, resulting in a random codebook initialization. In this paper,
an effective codebook initialization algorithm based on Kernel density estimation
has been proposed. Experimental results show that the proposed algorithm not only
improves the quality of generated codebook but decreases the computation time
compared to the GLA algorithm
خلاصه ماشینی:
New codebook design algorithm for image vector quantization based on Kernel density estimation Ali Darroudi*, Ghazaleh Sarbisheie, Hadi Jafarnia, Jabber Parchami Ali Darroudi, darroudi.
In this paper, an effective codebook initialization algorithm based on Kernel density estimation has been proposed.
Experimental results show that the proposed algorithm not only improves the quality of generated codebook but decreases the computation time compared to the GLA algorithm.
Key words: Vector quantization, Codebook generation, GLA algorithm, Image compression, Kernel density estimation.
So, major steps of image compression for our proposed method are summarized as follows: (View the image of this page)Figure 2: Investigation of data density in different areas of 2-dimensional space for training vectors of set I.
Estimate distribution function of the training vectors magnitude by Parzen windowing method Segment space areas based on Euclidean length Table 1: Data density in different areas for training vectors of set I (View the image of this page) Integrate the pdf of Euclidean norms on different areas Choose initial cluster centers in each region based on data density Use the initial codebook generated in previous step, as input for GLA algorithm and obtain the final codebook Use the final codebook to obtain the index table Send codebook and index table to decoder.
(View the image of this page) Figure 3: Probability density function of Euclidean lengths for training vectors in set I 4.
Experimental results To evaluate the performance of the proposed algorithm, four standard gray level images, has been used as our training set to generate codebooks.