Neural network object recognition error detection according to predicted token count comparisons

Abstract

Method of identifying errors in a transcription of a document image 102, comprising: generating a first document transcription 104, 106 from the document image using a neural network 110; generating a second document transcription from the document image using a second different technique than the neural network 122; and identifying one or more errors in the first document transcription based on a comparison of respective numbers of unique tokens in the first document transcription and the numbers of unique tokens in second document transcription 124. A semantic class label in the first document transcription may be identified and used to identify the second different document transcription technique. The unique tokens may be text tokens, table tokens, or formula tokens. Description tokens may be removed from the first document transcription before the comparison of tokens with the second document transcription. The description tokens may correspond to markup content that is generated by the neural network. Erroneous content tokens may be removed from the first document transcription before the comparison of tokens with the second document transcription. The neural network may be implemented as part of a batch transcription system (310,Fig.3). [Figure 1]