quality images in training data set X, each sample image is
associated with a N-WER i, which measures the document
image’s quality. Given the feature fi = [D
; R
]T of a document
image Ii, the goal of quality accessment is to find a
hypothesis H : fi ! i, which maps the feature fi of the
document image to a N-WER i.
Firstly, we calculate the N-WER for each document image
on both training set and testing set: