Fast Bidirectional-Similar-Function-Based Image Synthesis And Editing

Varieties of images and videos are produced with the development of mobile internet and it’s very necessary to process and edit these data. Patch-based image editing methods are very useful when applied to many image editing applications, which are essential among the image editing methods. Bidirectional-similar-function-based (BSF) image editing method is one of the patch-based methods and has been widely applied to image summary, texture synthesis and image inpainting etc. However, this method still exist deficiencies in the respect of run speed and its effects.Based on BSF, this thesis doing some research to speed up the running speed and improve the process effects. First of all, to accelerate the time-consuming nearest neighbor searching problem, it analyses and compares two approximate nearest neighbor searching methods (k-coherence and PatchMatch), in addition, it adopts GPU to run all compute-intensive kernel functions. Then to improve the effect of the results, we adopt the visual saliency model and redesign a controllable bidirectional similar function to adaptively process different patches with different importance.The main works and contributions of this thesis are as follows:1) Accelerating the nearest neighbor search which is used in the BSF based image editing. This thesis adopts the CUDA platform to develop parallel k-coherence methods and PatchMatch methods and assign the search of each patch in one CUDA thread which leads to better time performance thanCPU-based methods.2) According to the properties of the texture optimization and flow visualization, this thesis modifies BSF to get a new energy function to suit to those applications and the EM energy optimization algorithm is adopted to minimize the new energy which makes the produced texture similar with the original texture as far as possible and makes the produced new frames not only similar to the original one but also represent the attributes of the flow greatly. What’s more, those applications are accelerated by adopting parallel k-coherence which is developed in the CUDA platform.3) Conducting the research to confirm the role of BSF in the image/texture summary and then this thesis analyzes the influence of the function in the aspect of controllability and effects. Based on these, it redesign the energy function by utilizing the visual saliency model which makes the image/texture summary more controllable. At last, this new function is applied to the texture design, importance driven texture compression and image retargeting etc.