Synthesis Technology And Efficient Image

Digital image compositing is an important topic in the fields of image process and computer vision, and it also is considered a core issues in image editing. Image compositing has widely been applied in movie special effects, publishing and printing, graphic design and advertising production et al. Although the research on image composting has been developed for many years, it is still difficult to obtain a seamless composite (in the view of perception) for images have large different with few user interactions. In other words, inconsistencies between source image and target image make image compositing difficult, and the inconsistencies commonly exist in luminance, color, texture, noise pattern and style. Therefore, eliminating these inconsistencies plays a critical role to obtain a visual pleased composite. The thesis aims at alleviate the inconsistent artifacts such as "bleeding effect", color distortion and misaligned textures in gradient-based image compositing, and the research focused on how to use image coarsening, gradient integration, color correction and patch matching to improve image composting. The main contributions of this dissertation are summarized as follows:(1) Gradient-based image compositing usually produced a blur artifact called "bleeding artifact" near the boundary, and the artifact was difficult to prevent only by boundary optimization. To alleviate the "bleeding artifact", a compositing method based on image coarsening was proposed in this dissertation. The bilateral image coarsening (BIC) space bound similar pixels together in smooth regions and left pixels across edges independent, which was useful in reducing the numerical error of gradient integration. To switch the image compositing to BIC space, three different projection methods were proposed in the thesis. Direct projection projected a composite into BIC space by solving a minimization problem. Gradient projection projected boundary and gradient of a source patch instead of the result composite. Modified gradient projection computed a correction function with deleted gradient of boundary in BIC space. Experimental results demonstrated that the proposed method reduced the "bleeding artifact" significantly, and the computational cost was also reduced as the linear system solved in BIC space was far smaller than the original one. (2) Image compositing often suffered color distortion when the hue of target image was different from that of source image. To correct the color distortion, the quantitative representation of color distortion must be given first. In this dissertation, a color belief estimation method based on the general geodesic distance transform was proposed. The regions with color distortion were indicated by user strokes, and then the pixels covered by the strokes were accumulated to estimate the distribution of foreground or background color by Gaussian mixture model. The estimated distributions were incorporated into geodesic distance to measure the affinity between pixel and foreground, and a Hermite interpolation was used on the general geodesic distance to obtain a smooth color belief. Experimental results showed the estimated color belief was robust to different composites, and the accuracy of results could be increased by simply adding more user strokes.(3) To correct color distortion in image compositing, a closed form objective function constituted of a color fidelity term and a color propagation term was proposed. The color fidelity term was designed to preserve the original color for pixels had higher color belief, and the propagation term corrected the color distortion by propagating the color of high belief region to the lower one. To improve the color propagation further, an affinity function based on the enhanced long edges was proposed to prevent the color crossing over the weak edges. Experimental results demonstrated the composite obtained by the proposed method suffered little color distortion while kept a seamless compositing boundary and smooth transition between foreground and new background.(4) According to using dominant geometry transformations of patch matching, an improved patch-based image compositing was proposed. It could be observed that the geometric transformations (including shift, rotation and scale) used to find nearest patches between different images were sparsely distributed, thus, the dominant transformations could be found to represent prominent patterns of transformations of patches. Limiting the search space to a few dominant transformations and local neighborhoods improved image compositing reduced the computational cost. The experiments demonstrated the proposed method alleviated the blur in blending regions of compositing and aligned small misaligned texture better (the blur and the misaligned texture both were caused by inaccurate patch matching). The running time of patch matching also reduced significantly in the proposed method, especially for the images with high resolution.