Scene text character recognition based on image-to-class distance metric learning

With the increasing needs from a variety of real world applications like context retrieval and aid reading, scene text recognition is attracting more and more attention from the computer vision community. Scene text character (STC) recognition plays an important role in this task. However, recognition of STC is a challenging task due to a series of problems, like different illumination conditions, heavy occlusions and complex backgrounds. STC recognition can be further divided into two stages: feature representation and multi-class classification. Most work have been done to find better feature representation. In this paper, we focus on the stage of multi-class classification and propose a novel method names KNN based image-to-class distance metric learning (T2CDML). We first implement a global histogram of oriented gradients (GHOG) descriptor for feature representation. Then the KNN based I2CDML method is introduced to deal with the STC classification. Our KNN based I2CDML is robust to high intra-class variations. Besides, the proposed method supports incremental learning. To better evaluate our method, we conduct a series of experiments on two benchmark datasets, CHARS74K and ICDAR2003CH. Compared with other existing methods, our experimental results on these two datasets demonstrate very promising performance of the proposed method on recognizing characters in complex natural scenes.