Describir: Figure-Ground Discrimination and Distortion-Tolerant Recognition of Color Characters in Scene Images