Rejection strategies for offline handwritten text line recognition

Confidence Measures for Error Correction in Interactive Transcription Handwritten Text

“…On the contrary, if at a first glance no significant portion of the text line seems to be correctly recognised, then he/she might ignore system output and transcribe the whole text line manually. On the other hand and by contrast to previous works [9], here confidence measures are based on posterior word probabilities estimated from word graphs since, at least in the case of speech recognition, experimental evidence clearly shows that they outperform alternative confidence measures, and even posterior word probabilities estimated from N -best lists [7,8].…”

“…Although the use of confidence measures for offline handwritten text line recognition is not new (see [9] and the references therein), here we go a step further and confidence measures are proposed to guide the human supervisor in locating possible system errors and deciding how to proceed. For instance, if a small number of transcription errors can be tolerated for the sake of efficiency, then he/she might validate the system output after only checking those (few) words, if any, for which the system is not highly confident.…”

Evaluating Retraining Rules for Semi-Supervised Learning in Neural Network Based Cursive Word Recognition

“…Finally, the recognition's confidence value of the word's output is set to p(y, n) if that value exists, otherwise to p(n). For more details on these confidence values, we refer to [11].…”

Rejection Strategies with Multiple Classifiers for Handwritten Character Recognition

“…While high-accuracy character recognition has been achieved, in some applications even very few recognition errors are extremely costly. By implementing rejection strategies in a handwriting recognition system, we are able to improve the reliability of the rejected characters, and increase the accuracy of the remaining characters [1].…”

“…For such an approach rejection strategies can be formulated as simple thresholding operations. A large number of confidence measures have been proposed in the literature [1,3,9,13,14]. There are two main categories for producing confidence measures.…”