Affiliations: [a] College of Computer Science, South-Central University for Nationalities, Wuhan, Hubei, China | [b] College of Computer Science, Wuhan University of Technology, Wuhan, Hubei, China
Correspondence:
[*]
Corresponding author: Bo Li, College of Computer Science, South-Central University for Nationalities, Wuhan, Hubei, China. E-mail:libo_hust@126.com
Abstract: Document images tend to warp when scanned or photocopied
in document processing, resulting in non-linear distortion which can cause a
decrease in image quality, and bring errors to image analysis and
recognition. This paper treats the warped text lines as curves of non-linear
function, and proposes an adaptive non-linear correction model to rectify
the distorted document images. The model is based on a novel character and
word two-level structure and flexible parameters calculation method which
are robust to various kinds of distortions of text lines. After
pre-processing, character-based connected domains are detected and merged.
Some methods are used to locate the heads of text lines, trace them and
extract each text line accurately. Character connected domains are then
sorted via insertion sort algorithm, and important parameters are adaptively
calculated by adopting statistics method related to document region
partition. Finally, tilt angle is calculated using angle calculation model,
which is based on word-level unit, and curled text lines are corrected
adaptively based on character-level unit. Experimental results show that the
method proposed is effective and robust in correcting the curl of text lines
in various degrees and directions.
Keywords: Document image correction, non-linear distortion, text lines, two-level structure, document region partition
