Exemplar Driven Character Recognition In The Wild

Character recognition in natural scenes continues to represent a formidable challenge in computer vision. Beyond variation in font, there exist difficulties in occlusion, background clutter, binarisation, and arbitrary skew. Recent advances have leveraged state of the art methods from generic object recognition to address some of these challenges. In this paper, we extend the focus to Indic script languages (e.g., Kannada) that contain large character sets (order of 1000 classes unlike 62 in English) with very low inter character variation. We identify this scenario as a fine grained visual categorization task, and present a simple exemplar based multi layered classification approach to the problem. The proposed approach beats the existing state of the art on the chars74k and ICDAR datasets by over 10% for English, and around 24% for Kannada.