| Budget Amount *help |
¥4,020,000 (Direct Cost: ¥3,600,000、Indirect Cost: ¥420,000)
Fiscal Year 2007: ¥1,820,000 (Direct Cost: ¥1,400,000、Indirect Cost: ¥420,000)
Fiscal Year 2006: ¥2,200,000 (Direct Cost: ¥2,200,000)
|
|---|
| Research Abstract |
Vector quantization (VQ) has been widely studied for image and video compression. We are taking notice of VQ for image and video compression. VQ requires the large computational cost to construct a code book (CB). However, once we construct a CB, decoding involves only search form the CB. Therefore, the computational cost to encode and decode an image is negligible. This is very attractive point for image compression and applications. VQ consists of two parts: encoding and decoding. For encoding and decoding with VQ we first prepare a CB. To construct a CB is essential for VQ compression and the method to construct is purpose of this project. In, VQ compression rate and quality of the decided image depends on the size of code vectors (CVs) in a CB. In principle, the larger the size of CVs in a CB is, the lower is the compression rate of the decode image, while the smaller the size of CVs in a CB is, the higher is the quality of the decided image. This is a trade-off between compression
… More
rate and quality of a decoded image. To solve this problem, division with variable block size was proposed. The previous methods divide an image based on quad-tree (QT) decomposition A QT decompose image based on homogeneity of local regions of an image. However, the complex regions may have wealth information, but homogeneous region may not. In this sense, we implemented VQ with variable block size using local fractal dimensions (LFDs). We evaluated a decoded image by perceptual image quality measure (PIQM). In this project, we proposed an optimal CB design suing a genetic algorithm (GA). A GA is stochastic search method for finding optimal solution. The idea of a GA is based on the mechanism of natural selection and genetics. The basic procedure of a GA consists of selection, a crossover, and mutation. Gas have been widely used in complex optimization problems and have been shown to provide good solutions for these problems. An advantageous point of a GA is its ability to find a global optimal solution in multidimensional space, and this ability is also useful for constructing an optimal CB of VQ for image compression. We used the real-coded GA to design a CB. Its variable space is continuous, while the binary-coded GA is not. In the real-coded GA, genes are coded by real values instead of binary values. By using real-coded GA, since it is not necessary to decode from genotype to phenotype, individuals can be represented in shorter size than those represented by binary value. This ian advantageous point of the real-coded GA. We use the minimal generation gap (MGG) algorithm for the selection of individuals. In the MGG algorithm, simulated binary crossover 'SBX) is employed to generate a ne population. Since an image is a color image, we divide in into three components: red, green, and blue. The CB is constructed for individual color components. The image of each color component is divided into blocks. The blocks are learning vectors and there are 16384 vector, each of which is 16 dimensions. We generate 255 CVs from these vectors using GA. The algorithm to compute CVs is as follows. First, 255 CVs are chosen out of 16384 vectors as the initial CVs randomly. These 255 CVs are connected to generate one individual for the real-coded GA. Then, 30 individuals are generated as population. A fitness function is then computed. We propose vector quantization with variable block size for color images. Image division was carried out using local fractal dimension. Encoded image was evaluated by normalized perceptual image quality measure (NPIQM). Compression rate was also evaluated by bit per pixel Results of experiments show that compression rate is almost the same as that in the case of a GB with fixed block size. However decoded image quality is superior to that decoded by the CB with fixed block size. NPIQM is larger than 4.0, meaning perceptual level 4 (good). Less
|
