Semantic boosting rules for improving text recognition

摘要

The accuracy of a text recognition process can be improved using a set of semantic boosting rules, as may be contained in a sequence or other such arrangement. When text is output from a text recognition process, that text can have alternatives and confidence values for different characters or portions of the string. In order to improve the accuracy, this data can be processed using the organized rules, where rules are applied as long as any preconditions for that rule are satisfied, and each rule has the ability to modify the confidence values or modify one or more of the alternatives. When a result it produced with a minimum confidence level, or all applicable rules have been applied, the result can be provided as a refined text output of the recognition process.

主权项

1. A system, comprising:
at least one processor; and memory device including instructions that, when executed by the at least one processor, cause the system to: obtain an image including text; process the image with a text recognition algorithm to produce text string data, the text string data including at least two options for at least one portion of the text string, each of the at least two options having a respective confidence value; process the text string data using a rule decision tree, the rule decision tree including a plurality of hierarchical nodes, at least a portion of the hierarchical nodes corresponding to a respective semantic boosting rule, wherein processing the text string using the decision tree includes, for at least one node of the nodes in the decision tree: determine that a pre-condition is satisfied for the semantic boosting rule, for a first node of the decision tree, with respect to the text string; apply the semantic boosting rule for the first node to the text string in response to determining that the pre-condition is satisfied, the applying of the semantic boosting rule causing in at least one confidence value for the text string to be adjusted and a refined version of the text string to be generated; and upon receiving the refined version of the text string, provide the refined version as recognized text for the image.