Layered Coding Of Color Image Based On Subband Decomposition

The applications of digital image and video communication have led to thedevelopment of digital image coding. The first international digital imagecompression standard for continuous-tone still images, known as JPEG (JointPhotographic Experts Group) was extended broadly. However, the advance indigital technology has opened up possibilities for new image and video serviceswhich require innovation to existing communication networks. In order totransfer images with large capacity such as HDTV, CATV and color fax, thebroadband communication is necessary. As the transmission and switching modefor Broadband Integrated Service Digital Network (B-ISDN), the AsynchronousTransfer Mode (ATM) technology has been deployed.With flexible and high-speed switching structure, ATM can realizeintegrated transmission from narrowband voice and data service to broadbandvideo communication. However, cells may be lost when congestion occurs in thenetwork. Cell loss has become the main disadvantage of ATM network. So far,there is no way to avoid cell loss completely. But, some compensation algorithmshave been proposed to reduce degradation of image quality caused by cell loss,among which priority layered coding methods have drawn significant attention.In these methods, information contained in images is partitioned into differentlayers in accordance with human visual perception, which is then packed intodifferent cells for transmission. When congestion occurs, cells with low prioritymay be discarded to maintain continuous communication and acceptable quality.Layered coding is to partition image data into several layers according to itssignificance, which are coded further respectively with different methods. Itsbasic idea is to seek out the important information for reconstructing the imagefrom the whole information. The important information is called basic layer,while other information is called enhanced layers. A remarkable characteristic ofthe priority layered coding is its ability of providing error resilience. In theenvironment where error exists, the important information of the image can betransferred on the reliable channel or transferred with more protection, so as toavoid degradation of reconstructed image quality. However, the implementationof layered coding will inevitably increase complexity and some extra codes, anddecrease coding efficiency.Layered coding of image is performed according to the significance ofinformation. The information included in the basic layer is the most importantand reflects the main content of an image. Therefore, it is endued with the highpriority. The enhanced layers which reflect the details of an image are lessimportant and are endued with the low priority. Thus, different layers of an imageare endued with different priority, the cell loss only occurs in the low priority cell,the transmission of the high priority cell is guaranteed. So the degradation ofreconstructed image quality can be reduced when network congestion occurrs.The image quality is perceived by human eyes, so the characteristic ofhuman vision can be considered when designing the scheme of image coding.Researches indicate that human eyes have the characteristic of band-pass in thefrequency doman. It is well known that human eyes are not as sensitive to highfrequency components as low frequency components. So we can apply vectorquantization to the high frequency bands which can enhance the compress ratioand make little effect on reconstructed image quality. Even if we omit some highfrequency bands, there is not much difference for perceiving by human eyes.Thus, we can partition images into high priority layers and low priority layers,which compensate the influence of cell loss in ATM by making use of thecharacteristic of human vision.In this paper, image layers are obtained by subband decomposition withconsideration of human visual properties. It can be applied in ATM network.Subband decomposition of image has an inherently layered structure. Its basicidea is to design a decomposition and synthesization filters (we use QuadratureMirror Filters in this paper), partition image into different bands in frequencyfield, and adopt different coding methods for different bands. In the paper, thecoding method of baseband is based on Discrete Cosine Transform (DCT), thecoding results of the baseband could be transmitted in high priority cells, while tothe high-frequency bands, we choose some of them and use vector quanzitation tocode them. They are transmitted in low priority cells. In the decoder, we canreconstruct the image using different bands, which have been decoded.The basic concept of subband coding was introduced by R.E.Crochiere in1976 as the algorithm of speech coding. Hereafter, it was applied broadly inspeech coding. The basic idea of subband coding is to divide the channel ofinformation into different bands. For every band, we can select the codingmethod which is suitable for the statistical characteristic of the band. Twoimportant virtues of subband coding are: firstly, error of the subband coding canonly occur in one band, and it does not extend to other bands, which avoid errorpropagation to the whole channel;secondly, we can distribute different codingbit-rate to each band, which is very important for variable bit-rate coding anddynamic assignment of bandwidth.In the beginning of this paper, the background of selecting this topic isintroduced, and the significance and possibility of image coding are alsopresented. Some basic knowledge of image coding is presented in Chapter 2.Traditional image coding methods include predictive coding, transform coding,vector quantization and entropy coding etc. The image studied in this paper iscolor image, so knowledge about the color space is presented, the spaces of RGBand YUV are introduced, and the conversion between them is also mentioned.The emphasis of this paper is put in Chapter 3 and 4. The algorithm of subbanddecomposition is explored in Chapter 3, which importance lies in designing thefilters including the basic theory, designing method, choice of steps andmanagement of mixing in the frequency. Subband coding in speech coding can beconsidered as 1-dimension, the still gray image is 2-dimension, and the3-dimension can be the dynamic gray image or the still color image. In this paper,we studied the latter. In Chapter 4, the coding scheme of each band after subbanddecomposition is presented. The baseband which is very similar to the originalimage can be used to reconstruct the original image coarsely, therefore we selecttransform coding as its coding method. It includes block partition, DCT andentropy coding. Firstly, we divide the original image into blocks (usually withsize 8×8), transform the blocks to 8×8 coefficient matrix by using DCT, andquantify the DCT coefficient matrix using a 8×8 quantization matrix. Lastly, thecoefficients after quantization are encoded by entropy encoder. Thehigh-frequency bands are encoded with low priority which can be dropped ifnecessary, therefore the coding of high-frequency bands does not need too fine. Inthis paper, the vector quantization is applied to encode the high-frequency bands.In order to evaluate the algorithm, experiments done, which include experimentsof subband decomposition and coding of different band. Experimental results andperformance analysis are given in Chapter 5. Conclusions and perspectives aremade in Chapter 6.