Portland, Ore. An algorithm that simplifies the labor-intensive task of colorizing black-and-white motion pictures has been created at Hebrew University in Jerusalem. The algorithm lets simple scribbles accomplish what would otherwise demand exacting definition.
"Our colorization algorithm does not require precise image segmentation or accurate region tracking," said Dani Lischinski, an assistant professor at the university's School of Computer Science and Engineering.
Instead of precisely defining areas, as computer-generated techniques require, he said, the algorithm allows an artist simply scribble over areas with "crayons" of different colors. The algorithm then takes over, filling in the regions with the various colors and tracking them from frame to frame.
Lischinski developed the algorithm with fellow professor Yair Weiss and graduate student Anat Levin.
Computer-generated (CG) colorization techniques were themselves considered a major advance in the controversial practice of adding color to vintage black-and-white movies, automating what had been a manual process. Television's first colorized film, the 1986 broadcast of 1947's Miracle on 34th Street, used CG techniques.
But CG colorization requires highly skilled artists to designate precise color "regions." After that is done, complicated modeling and tracking algorithms can run for a few frames before failing. Then the process must be repeated.
Hebrew University's algorithm, the engineers said, is far less error-prone and labor-intensive.
"Our method is based on a simple premise that neighboring pixels with similar intensities should have similar colors," said Lischinski.
By using a quadratic cost function, Lischinski's algorithm optimizes colorizing regions with only the need to annotate the image with color scribbles. The algorithm automatically propagates the colors in both space and time to produce fully colorized images and sequences. The rudimentary annotations supply enough information to enable the algorithm to fill in subsequent frames slightly differently, thereby rendering their color correctly without explicit object tracking.
By contrast, selecting and linking disparate portions of occluded and fuzzy images using traditional CG techniques requires a separate tracking algorithm.
The researchers have applied for a patent on the algorithm but are making it available to engineers while the patent is being processed. The Matlab source code for the colorization algorithm can be downloaded at www.cs.huji.ac.il/~yweiss/Colorization/colorization.zip.