Reducing Overconfidence In Neural Networks By Dynamic Variation of Recognizer Relevance

Contemporary recognition systems use various methods of symbol recognition and post-processing methods designed for enhancing the quality of text recognition. For some recognition problems it may be difficult to create an adequate dataset for training symbol recognizers, so several symbol recognizers are used to ensure better performance. In this paper the concept of recognizer relevance is introduced as a way of analysing the recognizer output. A method is described using this concept, allowing to use external information about the input samples in order to balance the contributions of the recognizer and the post-processing subsystem.