Context-Aware Confidence Estimation for Rejection in Handwritten Chinese Text Recognition

Yangyang Liu, Yi Chen, Fei Yin, Cheng-Lin Liu

Abstract: Handwritten Chinese Text Recognition (HCTR) has been advanced largely by deep learning in recent years. However, the remaining recognition errors still hinder reliability-critical applications where zero-error is desired. Rejecting low-confidence patterns can help reduce the error rate but the increased rejection rate is also harmful. In this paper, we propose a character confidence estimation method incorporating contexts for character rejection in HCTR. Based on a text line recognizer outputting character segmentation and classification results, the confidence of each segmented character is estimated by combining the scores of a re-trained character classifier, the linguistic and geometric contexts. We introduce a probabilistic formula for estimating the confidence by combining the classifier and contextual scores, and an improved approach for scoring the geometric context using unary and binary geometric features. Experimental evaluations on the CASIA-HWDB and ICDAR2013 datasets demonstrate that our method can significantly improve the rejection performance in respect of low error rate at moderate rejection rate. The re-trained classifier, the linguistic context and the geometric context are all justified effective to improve the confidence.