The Research And Implementation On Algorithms For Image And Video Colorization

With the development of computer technology, the applications of images and videos are becoming more widely. But many technological images, such as satellite images, medical images and some relics of ancient buildings images are still presented in the form of grayscale images. However, the human eye has a poor ability to distinguish grayscale, while it can distinguish millions of different colors. For this reason, color is an important property of an image or a video. In recent years, the subject of image and video colorization has become an important research direction in the digital image processing field.The image and video colorization refers to the process of adding color information to the monochromatic image or video through the assist of computer. It is a map from one-dimensional grayscale image data to the three-dimensional chromatic image data. There are many domestic and foreign researchers have studied this subject and proposed a lot of colorization algorithms.The existing colorization algorithms can be mainly divided into two categories. One is a class of non-interactive, the other is a class requires human-computer interaction. Pseudo-colorization and color transfer algorithm based on the reference image are two important methods of non-interactive algorithms. Between the interactive algorithms, colorization based on segmentation and colorization using optimization are representative.This paper proposes an innovative method based on the reference image, after the analysis and comparison between the existing methods of their own advantages and limitations. It is different from the previous methods that depend on the comparisons between the grayscale image and the reference image based on the luminance value and neighborhood statistics of each pixel, our method takes pattern continuity and spatial consistency into consideration. Our method needs a partially segmented reference color image as the input. For each pixel in the reference image, we construct a pattern feature vector using the mean and the standard deviation of the Gabor wavelet transform coefficients. Based on the training set, we transform the vectors into a low dimension space, and obtain a classifier. For the pixels in the grayscale image, we classify the pixels and learn the color information from the corresponding region. Finally, we preserve the more reliable pixels as the scribbles and employ the optimization-based method to get the final result.According to the experimental results, we can find that using our algorithm can generate better results. Our method has the following advantages compared to the existing methods. First, using wavelet analysis can better preserve the pattern continuity, to get a global optimal solution and generate better visual effect. Secondly, the process of colorization is almost automatic, requires little user effort, which saves a great amount of labor. Finally, our method generates the scribbles spontaneously, which is more important for video colorization. Besides, the method we propose is not only applicable for natural images and videos, but also suitable for cartoon images and videos, which have rich texture information. And this paper will specifically focus on the colorization results of cartoon images and videos.