Normal direction local binary pattern for fragment reconstruction

Fragment reconstruction aims to restore broken images and documents via matching spatial adjacent fragments. As the existing solutions in the literature still remain problematic, we present a novel feature descriptor, Normal Direction Local Binary Pattern (termed as ND-LBP), for document/image fragment matching. ND-LBP is based on the conventional LBP descriptor, however, it outstands LBP by introducing new features derived from shapes and contents of fragments to promote its discrimination. With normal direction operation, ND-LBP is rotation-invariant, and thus could effectively and efficiently match fragments of arbitrary orientation. According to our extensive evaluations on real world datasets, the fragment reconstruction approach with ND-LBP feature has high precision, robustness and efficiency, and outperforms existing features.