Colorization of black-and-white cartoons

EXCOL: An EXtract-and-COmplete Layering Approach to Cartoon Animation Reusing

“…Our work belongs to the scope of cartoon animation processing. There are many techniques published in the domain of cartoon processing, e.g., cartoon creation, displaying and editing [4], [5], [8]. But only few techniques are designed for reusing cartoon animations due to the challenging nature of the involved inverse problems.…”

“…For example, assuming a single foreground layer consisting of homogenous regions bounded by clearly visible outlines, Sýkora et al [4] use outlines to segment black-and-white cartoons and then classify segmented regions by area thresholding. As we shall demonstrate later (e.g., in Figure 10), such simple area criterion might not already be robust for two-layer decomposition of modern cartoons with complex scenes, making its extension to multi-layer extraction rather challenging.…”

“…Constructing a panoramic background image essentially requires locating and registering the background layer L N k of every frame in a single global coordinate system [4]. Each frame corresponds to a warped region in the global system by a homography transformation H k : L N k →L N k .…”

“…Second, a layer completion step is further needed to fill missing parts caused by occlusion between layers. Although cartoon layering or segmentation is proved vital to effective reusing of classic cartoons, the existing works [1], [2], [4], [5] are limited to the problem of two-layer decomposition (i.e., to extract a foreground layer and a background layer) and do not deal with layer completion at all.…”

“…Noticing that cartoon animations, created with cellpainting or vector graphics tools, usually exhibit smooth and well-defined shapes, as well as large regions with simple colorization and visible outlines [4], [7], we fully explore shape cues extracted from an input cartoon video to facilitate both extraction and completion of layers (see an overview of our algorithm in Section 3). EXCOL is able to not only extract multiple semantically meaningful and temporally coherent layers in a single pass, but also faithfully recover structural features within possibly large missing regions in the extracted layers (see Figure 1 and the accompanying video).…”