Efficient three dimensional rotation estimation for camera-based OCR

Abstract

Camera-Based Optical Character Recognition (CBOCR) has attracted interests of many researchers in both computer vision and document analysis research fields. A significant challenge in CBOCR is how we handle characters of those appearances are affected by three-dimensional (3D) rotation due to locational relationship between a printing plane and camera. Proper handling of these 3D rotated characters is expected to improve the performance of both detection and recognition of camera-captured characters. In this paper, we propose an efficient implementation of 3D rotation estimation for camera-captured characters. The proposed implementation requires small memory load and short computational time. We employ Linear Discriminant Function (LDF) instead of Modified Quadratic Discriminant Function (MQDF) for further memory reduction. The results of experimental evaluation using a large-scale alphanumeric character dataset showed that small number of dimensionality of original feature vector is sufficient for keeping accuracy of 3D rotation estimation and total amount of memory required for 3D rotation estimation is reduced from 141.0 MB to 6.6 MB.