Colorization Using Quaternion Algebra with Automatic Scribble Generation

In current colorization techniques, major user intervention is required in the form of tedious, time-consuming scribble drawing. Moreover, color leakage usually occurs across contours and object boundaries. In this paper, we focus on automatic scribble generation and structure-preservation mechanism, which are still open issues of colorization. Firstly, we generate scribbles automatically along points where the spatial distribution entropy achieves locally extreme value. Given the color scribbles, we compute quaternion wavelet phases to conduct colorization along equal-phase lines. These lines across scribbles and monochrome patches locate textures with similar pattern distribution. Contour 'strength' model is also established in scale space to direct color propagation among similar edge structures. Finally, we reconstruct color image patches as vector elements using polar representation in quaternion algebra, well-preserving interrelationship between color channels. The experimental results demonstrate that the proposed colorization method can achieve natural color transitions between different objects with automatically generated scribbles.