Automatic Document Orientation Detection and Categorization through Document Vectorization

1
Automatic Document Orientation Detection and
Categorization through Document Vectorization
Shijian Lu ( dcslsj_at_nus.edu.sg ) and Chew Lim Tan
(tancl_at_comp.nus.edu.sg) School of Computing,
National University of Singapore
Document vectors must be normalized first before
they can be used for document orientation
detection and categorization
Introduction and objectives Thanks to the
advances of document capturing capabilities, more
and more document images of different scripts are
being produced by scanners, digital cameras, and
fax machines. Without human checking, physical
documents may be improperly positioned as
illustrated in Figure 1 during manual or machine
feeding process. In addition, in a multilingual
environment, visual inspection is required to
categorize document images according to the
underlying scripts. Automatic document
orientation detection and categorization
according to the underlying scripts are in demand
to reduce the human involvements during document
digitalization process.
For each script currently under study, two vector
templates corresponding to the correct and upside
down orientations are constructed as the mean of
multiple training document vectors. For scripts
Korean and Roman, Figure 3 below shows the
constructed vector templates in two opposite
orientations.
Figure 3 Character classification based on the
CART structure
Figure1 Sample documents of different scripts
with skew angle arbitrarily lying within (a) 0
90 (b) 90 180 (c) 180 270 (d) 270
360, respectively.
The orientation and script of the query document
image can accordingly be determined based on the
Bray Curtis distance given below between its
document vector and the constructed vector
templates
Methods
We carry out document orientation detection and
categorization through document vectorization,
which converts each document image into an
electronic document vector through the
exploitation of the density and distribution of
vertical component runs (VCR) as illustrated in
Figure 2 below.
where Nv is equal to 32. VRVj represents the jth
element of the query document vector. Parameter
VTij corresponds to the jth element of the ith
constructed document vector template. As a
result, the orientation and script of the query
image are determined to be the same as that of
the vector template with the smallest Bray Curtis
distance VDi.
Experimental Results
Experiments have been conducted and the results
show that orientation of document images can be
correctly determined when document images contain
enough text. However, the performance may degrade
when document images contain too little text as
shown in Table 1 below
Figure2 Definition of vertical component run.
For each labeled document component, a component
vector of size 32 can be constructed where the
first 8 elements specify the number of VCR along
text lines and the following 24 record the
position of VCR within the top, middle, and base
text zones, respectively, as illustrated in
Figure 2. We set the upper limit of VCR number at
8 as the number of VCR in each scanning round is
normally no bigger than 8 for most languages
currently under study.
Take the English character n in Figure 2 as an
example. The corresponding component vector is
10000000 10000000 00000000 00000000 because
there is just one vertical run that occur in the
upper text zone. Document vector can thus be
determined as the sum of all component vectors.