The Research Of Fast Image Editing Methods Based On Exemplar Learning

The image editing methods based on exemplar learning take reference images of a certain style as the learning object, and use the learned visual features to change the color and texture attributes of the target image, which makes the output image have the similar visual effects of the reference images. These methods have a wide range of applications for the fields of artistic image stylization, color transfer, image content editing and super-resolution detail synthesis.This thesis starts with the editing of image texture and color, analyzes the texture features and learns the semantic characteristics of the image regions to mine the abundant information of reference images, takes the understanding, extraction and reuse of the image features as the main research line to transfer the extracted feature to the target image, with the help of the GPU based acceleration technology to realize fast and high quality image editing, achieving the editing tasks of image content editing, artistic image stylization and color transfer.The main contribution of this thesis is as follows:(1)The texture adaptive content editing algorithm is proposed. Based on image texture features, the algorithm can choose suitable nearest neighborhood searching strategies according to different texture areas quickly and flexibly. It adopts GPU based global accurate neighborhood searching and k-coherence based coherence searching to achieve neighborhood matching process, which can learn the mapping relations of the sample images fast, and uses synthesis magnification to generate high resolution images with high quality quickly. Experimental results show that the algorithm is able to adjust the parameters conveniently, greatly speed up the neighborhood matching process, and improve the editing quality to a certain extent.(2) To further improve the quality of content editing task, a fast content editing method based on global optimization is proposed. The method deals the content editing task as a discrete global optimization problem by defining an energy function which measures the image similarity, and uses a modified GPU based global optimization algorithm to compute the result iteratively. This method uses an initialization process based on chamfer measurement and region growing to generate the initial sample, served as the input of the global optimization, which can not only speed up the convergence speed, but also improve the synthesis quality; The defined measurements of different texture regions capture the structural texture characteristics, which makes the synthesis results in accordance with human visual perception. The algorithm is implemented using CUDA architecture, which greatly speeds up the synthesis process. The experimental results show that the algorithm is able to learn the texture and color features of the sample images, at the same time maintain the smooth transition between the image regions, achieving high quality content editing.(3) The existing artistic image stylization methods have poor usability and flexibility. Aiming to solve these problems, a fast artistic image stylization method supporting multi-exemplar is proposed. The algorithm can not only learn the style characteristics of multiple samples to realize hierarchical rendering, but also provide a simple and feasible way of artistic image design to meet the users variability requirements of artistic style learning. By the guide of image direction field, the proposed flow based automatic style transfer algorithm modifies anisometric texture synthesis to learn the texture and brush style of the sample images quickly, and also improves the synthesis quality; The GPU based algorithm can transfer different brush style to different regions in parallel, which decouples the pixel/patch dependence existing in the previous style transfer algorithms to accelerate the style learning process; On the basis of automatic synthesis, this thesis also proposes the user interactive artistic image design method, which lets the user to further optimize the automatic generated result, makes the style learning process more flexible, and obtains more abundant style transfer effect.(4) The previous color transfer methods only consider regional low-level color statistics, and ignore the analysis of image content, this thesis proposes a semantic based color transfer method, which not only reduces the artifacts caused by inaccurate automatic mapping, but also needs a little user interactions. First, it modifies the original Normalized Cut segmentation method, which can segment more complete and homogeneous image regions as the input of the semantic learning process; Then, a MPEG-7descriptor based regional semantic labeling method is used to establish the correlation between the high-level semantic features and the low-level image features, which improves the labeling accuracy; At last, according to the learned regional semantic words of the reference image and the target image, it establishes region mapping to obtain more accurate and intelligent color transfer effects.