Digital image is enriching people ’s daily life,which is a main source of gathering information.Color information is important for digital image.Since color image contains both luminance information and chrominance information,it is more popularized by us.Image colorization refers to the process of adding color to a grayscale image or binary image to improve its visual quality.Image colorization is a technique coming from practice,and application-oriented is its most distinct feature.Beside the colorization of ealier films,both medical image processing and remote sense image processing are its typical applications.However,image colorization is a typical undetermined “ill-posed” question.It is usually required to incorporate prior knowledge or be human-intervened effectively.How to avoid the error propogation and reduce the number of human intervention are still worthy of further investigation while guranteen the colorization effects.In this thesis,several key techniques are inves tigated which are closely related image colorization.Specifically,the main works and contributions of this work are summarized as follows.First,a Normalized Cut-based interactive image colorization scheme is proposed.Its basic idea behind the proposed approach is to combine the Normalized Cut-based image segmentation with the color-annotation-based color colorization.Firstly,the key parameters for Normalized Cut are initialized in terms of the texture complexity of input grayscale image.It incorporates with Texton histogram-based merge-and-split technique to obtain the segmentation results with appropriate number of regions.Secondly,each region after image segmentation is specified with an initial color by color annotation,and then it is transferred to obtain the image colorization results.Experimental results on several images show that the proposed approach achieves desirable and natural colorization results.Second,an efficient image colorization approach is proposed by exploiting the unique value of pixels in both reference image and target image.The unique pixel values are counted for both the reference image and the target image.Then,they are normalized with the absolute difference by considering the whole differences between them.Matching is conducted on the basic of the normalized images.Experimental results on several images show that the proposed approach achieves desirable image colorization results compared with the Levin ’s approach,while it is much more efficient.In summary,we have made some useful investigation about grayscale image colorization.However,there are still some rooms for further improvements such as how to make image segmentation more suitable for the color annotataion involved in interactive image colorization,and the color transfering mechanism by considering image texture. |