Assisted Analysis of Classical Cartoon Animations

and contributions In this thesis a novel approach to analysis and processing of classical cartoon animations is presented. It focuses on a popular animation technique where each frame is created as a planar composition of two visually distinct layers: static textural background and dynamic homogenous foreground. The aim of the proposed framework is (1) to precisely detach these two layers to reach the state before the final composition and (2) to estimate structural correspondences between animation frames and homogenous regions in the foreground layer. Such analysis allows to partially understand the 2.5D structure of classical cartoon animations that is crucial for numerous applications including semi-automatic colorization, example-based synthesis, and video compression. The main contributions of the thesis: 1. Unsupervised sub-pixel accurate outline detection and image segmentation algorithms suitable for cartoon images where outlines represent region boundaries. 2. Hierarchical estimation of frame-to-frame and region-to-region correspondences based on local structural similarity and neighborhood relations. 3. Example-based approach to semi-automatic colorization of black-and-white cartoon animations dramatically reducing the amount of manual interventions. 4. Example-based synthesis of cartoon animations using an intuitive scribble-based interface allowing to quickly sketch new cartoon characters and poses in the style of masters. 5. Efficient video compression technique for classical cartoon animations based on a hybrid encoding scheme providing state-of-the-art visual quality for low encoding bit-rates.