A confidence-based method for keyword spotting in online Chinese handwritten documents

In keyword spotting from handwritten documents, the word similarity is usually computed by combining character similarities. Converting similarity to probabilistic confidence is beneficial for context fusion and threshold selection. In this paper, we propose to directly estimate the posterior probability of candidate characters based on the N-best paths from the segmentation-recognitioin candidate lattice. The N-best path scores are converted to confidence measure (CM) using soft-max, and the posterior probability of candidate characters is the summation of confidence measures of paths that pass the candidate character. The parameter for CM is optimized using the binary cross-entropy criterion. Experimental results on database CASIA-OLHWDB demonstrate the effectiveness of the proposed method.