GISdevelopment.net ---> AARS ---> ACRS 1999 ---> Image Processing High Efficient compression encoding using vector Quantization for the Satellite Image Tsukasa Hosomura Professor, Department of Computer and Information Engineering Kanazawa Institute of Technology 7-1 Ohgigaoka, Nonoichi- machi, Ishikawa 921 -8501 Tel : (81)-76-294 -6710 Fax: (81)-76-294-6736 E-mail: hosomura@neptune.kanazawa-it.ac.jp JAPAN Keyword: Codebook, Landsat TM, Innerband, Interband Abstract: Block transformation encoding system using the DCT(discrete cosine transform) has been standardized by JPEG (Joint Photographic Experts Group), and there is a problem of remarkably recognizing the block strain at the low rate. The application of wavelet transform to the image coding was started using analysis (Multi-Resolution Analysis :MRA) degree of the multiple solution images by Mallet in 1988. The picture quality for the sub-band code system using wavelet transform at the low rate is high. The data volume seems to become enormous by high resolution of the image with performance enhancement of the sensor and increase in the band number for the data got by the sensor of satellite. By it, there is the necessity of compressing the data. In this study, the image was divided into blocks for multiband image of the 7 bands got by Landsat TM, and the block was made to be a unit, and vector quantization was tried every each block. And, the spectral reflectance of the object is different by the type of the object, and the spectral reflection from the object proposes the method for using the vector quantization using distinguishing the object by the spectral reflection, every pixel, since the effect of the spectral reflectance is being received. The lossy encoding technique using this vector quantization of the every block and, lossy encoding technique which used wavelet transform for vector quantization of the every pixel and predictive coding is compared. Introduction The research on the encoding of the image has the history over 30, and by extending present, the research is actively carried out more and more. This depends on the development of research environments. The representative equipment is a computer, and it is also the speedup complicated encoding. In the encoding of the still picture, because the effect which reduces the redundancy of information is high, the bandwidth compression using DCT has been standardized by JPEG, and there is a problem of remarkably recognizing the block strain at the low rate DCT block transformation encoding system. The application of the wave let conversion to the image coding was started using the analysis degree of multiple solution images by Mallet in 1988. For the block transformation encoding using the DCT, the picture quality for sub-band code system using the wave let conversion at the low rate is high-grade. The data volume seems to become enormous by high resolution of the image with performance enhancement of the sensor and increase in the band number for the data got by the sensor of satellite. By it, it is necessary to compress the data, when the burden of the transfer from satellite and satellite in the receiving station memory capacity. In this study, the image was divided for multiband image of the 7 bands got by Landsat TM in large number of blocks, and the block was made to be a unit, and vector quantization was tried every each block. And, the method for using the vector quantization using distinguishing the object by the spectral reflection is proposed. Vector Quantization The whole image region is divided into the block of n X n, and the method for expressing gray level of all pixels in this block unit at 1 vector is the vector quantization. Vector quantization is efficient compression method of image and system for noticing in pattern recognition. The algorithm is shown in the following. Vector quantization algorithm of the innerband pixel block The encoding is done in the m dimension Euclidean space, when the picture element number in the block was made to be m = n X n. That is to say, the picture signal of one block seems to be a vector of m dimension with the component of the m piece, and it forms 1 vector in the m dimensional space. With that the vector is obtained for all screens of the image of 1 band given this time, the class with approached vector value is formed. Then, it is possible that it selects representative vector of each class and approximates it by the representative vector, when the optional input vector was given. For example, the vector of the 16,384 is formed in the whole screen, when 4 X 4 arranging block for the image of the 512 X 512 pixel is made to be vector 16th order vector. The data compression is realized by conducting the clustering for this vector group, and representing at 256 vectors. Encoding algorithm of the vector quantization of innerband pixel block is shown in Fig. 2.1. Fig. 2.1 Encoding algorithm of the vector quantization of innerband pixel block Representative vector of each class is beforehand set at the codebook. To begin with, the sampled-data of m piece of the input image comes in for the input buffer and is memorized. This input vector and representative vector of each class in the codebook are compared in order, and belonging to class is decided. Then, the number of representative vector of the class is decided, and it is output as an encoding output. It decides the cluster which the input vector f k belongs to fJ by using following equation. E=| f k - f j | 2            (2.1) Square distance E between vectors is calculated, and it is decided in the result of minimizing this. In the decoding side, the representative vector is read out from the decoding signal using the codebook of the content equal to the encoding side, and the decoding block is output. The preparation of the codebook for the vector quantization of the innerband pixel block The algorithm in making the codebook is shown in the following. The vector is early uniformly placed in the past 16th order space. The input of the original picture is assigned in the class of the closest representative vector. The new representative vector is made in the mean value in each class. It is halved from the class of which the dispersion is the biggest, when representative vector which is not used produced. The work of 2 ~ 4 is repeated The vector quantization of the interband identity position pixel block. The vector quantization of interband identity position pixel block which does vector quantization using features of multiband satellite image is proposed, because vector quantization is done for multiband satellite image. It is unique for the spectral reflectance of the object, and it is different by the type of the object. Then, in making the every pixel to be a unit using this, the value of each band is made to be a vector. Though for the image of 512 X 512 pixel of 7 bands, 262,144 vectors are formed as 7th order vector, this vector is represented at 256 vectors. The mean value of each every item was utilized using training data used in the land cover classification as a vector this time early. By making the codebook from the vector in initial stage, the high quality image is more obtained. The algorithm in making the codebook for vector quantization of encoding algorithm of interband identity position pixel is shown in the following. The vector is early placed on the basis of training data used in the past seventh space in land cover and classification, etc.. The input of the original picture is assigned in the class of the closest representative vector. The new representative vector is made in the mean value in each class. It is halved from the class of which the dispersion is the biggest, when representative vector which is not used produced. The work of 2 ~ 4 is repeated. Experiment Vector quantization of the innerband pixel block and vector quantization of the interband identity position pixel block are done, and the result is shown. And, the result of the compression encoding using wavelet transform is shown as comparison object with that vector quantization. Used satellite image data The image with the object in this experiment utilized the data which consisted of 7 bands of Landsat/TM. This images are Kanazawa City in August, 1984 and the nearby region, and they consist of each 1 pixel 8bit, 512 X 512 pixel, and the size of the image is 1835,008(byte). And, the equal image in August, 1985 and November, 1991 of the range it confirms the effectiveness of the vector early is used for the experiment. Experimental Method The compressibility used by this experiment is defined in the following equation. Compressibility = {(Amount of Compressed Image Data) /(Amount of Original Image Data)} x 100 (%)           (3.1) PSNR which shows signal-to-noise ratio is widely used as evaluated value of picture quality, and it is defined in the following equation. It is the least square error in which the above equation gives MSE in (3.2). In equation (3.3), M X N is the image size. x1(j,i) andare pixel value of original image and restoration image of band 1. Similarly, the sub script 2 ~ 7 shows each band. It is shown that the picture quality is better, as the value is higher for PSNR. The experiment carried out the compression using vector quantization of the innerband pixel block and vector quantization of the interband identity position pixel block. The huffman code was applied for the data which showed the correspondence with the number of the codebook. As a comparison object, in predictive coding and reversible encoding by huffman code and irreversible encoding, the encoding by wavelet transform is used. In the encoding by wavelet transform, the thresholding of the subband did the frequency of the subband division on 2 times and 3 times, stepsize in the quantization on case from 1 to 8 and 0% ( the thresholding is not made ), 0.5%, 1% and 2% of the largest electric power value. Experimental Result In the vector quantization of the interband pixel block, the data to which the codebook shows the correspondence between 4096 byte and number of the codebook becomes for the 114,688 byte, and it becomes a total for the 118,784 byte. Therefore, it becomes 6.47% compressibility. It becomes the 94,908 byte, when the huffman code is done for the data which shows the correspondence with the number of the codebook, and it becomes in the whole with the 990,004 byte, and it becomes 5.40% compressibility. PSNR is 32.208027dB. In the vector quantization of the interband identity position pixel block, the file to which the codebook shows the correspondence between 1,792 byte and number of the codebook becomes the 263,936 byte, when it becomes the 262,144 byte totaled. Therefore, it becomes 14.38% compressibility. It becomes the 231,707 byte, when the huffman code is done for the data which shows the correspondence with the number of the codebook, and it becomes in the whole with the 233,499 byte, and it becomes 12.72% compressibility. PSNR is 36.998659dB. In the reversible encoding, the original picture size became, and 1,835,008 byte and compression size became 1,003,129 byte, 54.666% compressibility. Summary On the case of the vector quantization, the picture quality is clearly better than the case in which the wave let conversion was used, and the compressibility has improved in the condition that the compression for information of vector number and codebook is not carried out. It seems to obtain the better efficiency, if that it makes large number of images as an element and makes the codebook in the every type of the sensor is possible in respect of the codebook. In the vector quantization of the innerband pixel block, block strain which can be clearly confirmed in visual observation appears. And, it is proven that the high frequency component has been lost. It was able to be confirmed that the image in which it is considerably similar even in visual observation to original picture was obtained in the vector quantization of the interband identity position pixel block. On the vector quantization, the preparation of the codebook seems to become an important problem. It is connected with earning considerable compressibility, if the common codebook is made. However, the codebook which made large number of images in the element is necessary. This time, it is necessary to choose the image for the codebook preparation in order to contain various regions and various seasons in order to consider and, the existence of seasons peculiar object, etc. so that the type of the object which has been projected in the image may increase. Concerning it, it seems to be effective for improving the picture quality to make small size codebook, when the common codebook was made. Reference Makoto Miyahara : "systematic image coding", IPC Co., 1991. Takeshi Agui, Masayuki Nakajima : "image processing", Morikita publication, 1991. Inst. of Television Engineers of Japan : "the image information compression", ohm Co., 1991.