Symbol recognition in natural scenes by shape matching across multi-scale segmentations

Abstract

Symbol recognition in natural scenes plays an important role in a variety of applications such as driver assistance and environment awareness. We propose a solution including 3 phases: (1) Image segmentation, (2) component-level shape matching, and (3) structure matching. To improve the robustness, we alter the parameters to obtain image segmentation at multiple scales and perform component-level template matching across the image segmentation results obtained at all scales. By means of such exhaustive search across all possible segmentations, the chance to obtain finely matched components is increased. Some initial experimental results are obtained, which are encouraging. © 2013 Springer-Verlag.