Research On Vital Technologies Of Stroke-Based Colorization

Stroke-based colorization has been applied in many areas. Because many factors influence colorization quality and couple together, it is difficult for any existing algorithm to describe the reason for falsehood perfectly, and accordingly, these algorithms performs not good enough in particular scene. Hence, it is of theoretical and practical significance to develop technologies based on analysis to colorization mechanism.This thesis focuses on researches on two key procedures in colorization, in particular, on relieving the difficulty in finding desirable stroke by decoupling relative factor, and on developing better diffusion method to reduce sensitivity of visual result to stroke distribution in order to improve visual effect over weak boundary and scattered areas. The main contents are shown as following:A general frame for optimization-colorization is put forward, and it is proved that optimization-colorization is an exact process of chrominance blending, where individual chrominance blending weight is equivalent to the pixel’s first-reach probability to relevant stroke. In addition, its equivalent circuit is established, which convert colorization process into determining voltage distribution of all node in this circuit, and thus provides different view, electrical theory, to address colorization. The above implicit blending mechanism not only make it available to modify stroke’s color and position respectively, but also inspires developing negative label which, in contrast to traditional label, plays the role of disappear specific label, makes good the deficiencies of traditional one as well as extends flexibility in improvement for poor result. In addition, the equivalence between blending weight and first-reach probability prepares analytical tool to investigate the very reason for falsehood.It is determined analytically that common algorithms are sensitive to distribution of stroke and tend to perform poorly over weak boundary areas, and three improvement methods are proposed. The first one employs the idea of evenness, but optimizes its implementation, which leads to better robustness and high speed. The second one employs blending mechanism discussed in the last chapter, can control width of chrominance transit area via bending-weight-factor, and provides better quality result. The last one introduces a new label to indicate weak boundary and provide relevant treatment.It is proved that common colorization methods are inevitable to perform badly in scattered object. Existing function is assigned artificially as one of some kernel functions, hard to capture the affinity accurately between pixels in complicated texture area. Aiming at this problem, a feature of natural color images is employed in modeling weight function. Firstly, a local linear chrominance-luminance relation was extracted from the distribution of pixels in color images, and then served as prior assumption, generalized to the whole image and finally provides a novel chrominance weight function model. To recolorization, a two-layers propagation algorithm incorporated with regions cues is put forward. Firstly, Regions by over-segmentation are combined with pixels to construct a two layer weight graph. In contrast to previous works which only employ connection between neighboring pixels, both full-connections in region layer and inter-connections between the pixels and their corresponding regions are introduced into this graph additionally. Then all these connections are integrated in a chrominance iteration formulation, not only facilitates propagation of local grouping cues across larger image areas but partly enforces color edit consistency inside each region. Because updating is performed in the case of membership of all pixels to every label, thereby the method can be extended to segmentation via comparison to resulting membership.Existing colorization algorithms can hardly provide pleasant visual effect in highlight or shadow areas. Motivated by the general imaging principle of natural color images, a realism improvement approach based on post-process correction in chrominance is proposed. First, an imaging model is built where influence of both light and objective on resulting image is decoupled. Then, a further chrominance-luminance relation model for any pixel imaged in white light is established and employed to identify the very reason for falsehood resulted from existing colorization methods and inspires a realism improvement strategy based on chrominance correction. Since only a grey parameter is to be set, two convenient operation modes are available-automatic coarse adjustment based on default values, and fine tuning by user’s intervention. In addition, this decupling model makes it easy to extend colorization to case of color light. This proposed method could provide a general platform to improve realism of colorization result.In order to speed colorization, a novel colorization strategy characterized as superpixel instead of basic pixel during colorization is put forward. Based on analysis to shortage of existing method, the reasonable criterion is formulated according to discussion of characteristics of colorization and facilitate selecting appropriate superpixel which lead to colorization in high speed with edges preserved. And then a guided filter is used to smooth out the result based on superpixel. Combining with improvement in colorization discussed in earlier chapter, we build a fast and flexible colorization system as this algorithm’s another application. User can modifies scribbles in real-time according to visual result thanks to short runtime (below one second) until get pleasant visual result.