We rely on the segmentation of the word hypotheses into characters (Sn) and on their labels (Hn) to build a segmental neural network to carry out verification at the character level in an attempt to better discriminate between characters and reduce the ambiguity between similar words. Given the output of the HRS, character alternatives are located within the word hypotheses by using the character boundaries, as shown in Fig. 2. Another module is used to extract new features from such segments and the task of the segmental neural network is to assign a posteriori probabilities to the new feature vectors representing isolated characters, given that their classes are known a priori. Further, the probabilities of the individual characters can be combined to generate confidence scores to the word hypotheses in the N-best word hypothesis list. Then, these confidence scores are combined with the recognition scores provided by HRS through a suitable rule to build a recognition and verification system. Fig. 3 presents an overview of the integration of the modules of the handwriting recognition system, the verification stage, the combination stage, and the decision stage. The following
sections describe the main components of the verification stage and how they are built and integrated into the handwriting recognition system.
We rely on the segmentation of the word hypotheses into characters (Sn) and on their labels (Hn) to build a segmental neural network to carry out verification at the character level in an attempt to better discriminate between characters and reduce the ambiguity between similar words. Given the output of the HRS, character alternatives are located within the word hypotheses by using the character boundaries, as shown in Fig. 2. Another module is used to extract new features from such segments and the task of the segmental neural network is to assign a posteriori probabilities to the new feature vectors representing isolated characters, given that their classes are known a priori. Further, the probabilities of the individual characters can be combined to generate confidence scores to the word hypotheses in the N-best word hypothesis list. Then, these confidence scores are combined with the recognition scores provided by HRS through a suitable rule to build a recognition and verification system. Fig. 3 presents an overview of the integration of the modules of the handwriting recognition system, the verification stage, the combination stage, and the decision stage. The following
sections describe the main components of the verification stage and how they are built and integrated into the handwriting recognition system.
การแปล กรุณารอสักครู่..
