Cartoon features selection using Diffusion Score

Similarity estimation is critical for the computer-assisted cartoon animation system to improve the efficiency of cartoon generations. The main issue in similarity estimation is choosing efficient features to describe cartoon images. Previous methods adopt pairwise distance to evaluate the similarity. However, this measurement is sensitive to noise. This paper proposes a novel feature selection method named Diffusion Score which captures the geometrical properties of the data structure by preserving the diffusion distance. Specifically, the Markov process is carried out to find meaningful geometric descriptions of the whole cartoon dataset. The diffusion distance sums over all paths' lengths which connect two data points. Since diffusion distance integrates “volume” of paths connecting data points, it is tolerant to noises. The time scale of Markov process can incorporate the cluster structure of data at different levels of granularity. It makes the number of the nearest neighbor K in graph construction to be an insensitive parameter. Therefore, Laplacian Score is sensitive in feature selection. Diffusion Score can effectively improve the stability by minimizing large absolute errors and large relative errors of the features. The experimental results can demonstrate the efficient performance of Diffusion Score in feature selection.

► We encode different features to build a physically meaningful subspace. ► We adopt discriminative information to obtain the optimal low-dimensional subspace for each view. ► We utilize unlabeled data to enhance the subspace learning. ► We use the alternating optimization to explore complementary characteristics of different features.