Texture synthesis for image compression

Still image compression methods have changed little over the last ten years. Mean- while, the quantity of content transmitted over limited bandwidth channels has increased dramatically. The currently available methods are content agnostic: that I is they use the same compression process independent of the content at any given spatial location. Region specific coding provides one possible route to increased compression performance. Texture regions in particular are usually not conceptually important to a viewer of an image, but the high frequency nature of such regions consumes many bits when encoding. Texture synthesis is the process of generating textures from a sample or parameter set, and thus if these texture regions can be encoded by spec- ifying texture synthesis at the decoder, it may be possible to save large amounts of data, without detriment to the decoded image quality. This thesis presents a number of adaptations to the Graphcut patch based texture synthesis method, to make it suitable for constrained synthesis of texture regions in natural images. This includes a colour matching process to account for luminance and chrominance changes over the texture region, and a modification to allow constrained synthesis of an arbitrarily shaped region. This architecture is then integrated into two complete image compression by synthesis systems based on JPEG and JPEG2000 respectively. In each case the image is segmented, anal- ysed and synthesis occurs at the decoder to fill in removed texture regions. In the system based on JPEG2000 a feedback loop is included which makes some assess- ment of the quality of the synthesis at the encoder in order to adapt the synthesis parameters to improve the result quality, or to skip synthesis entirely if deemed necessary. The results of these systems show some promise in that substantial savings can be made over transform coded images coded at the same Q value as the residual image. However it is observed that synthesis can be detrimental to the quality of the image in comparison to an equivalent traditionally coded image at the same bitrate. Two methods of texture orientation analysis for non-homogeneous textures are presented. One of these in particular produces a good assessment of the texture orientation. This method uses a Steerable Pyramid transform to analyse the orientations. Then, two methods of sample selection and synthesis using the analysed texture orientation are presented. These methods aim to recreate the original texture's orientation variation from a smaller texture sample and the orientation map. The best of these methods selects one or more samples containing multiple orientations and selects texture patches appropriately oriented to the current location of synthesis.EThOS - Electronic Theses Online ServiceGBUnited Kingdo