Segmentation of heterogeneous document images : an ... - Tel
Segmentation of heterogeneous document images : an ... - Tel Segmentation of heterogeneous document images : an ... - Tel
same location of i th zone in sub-figure (a) and (b), respectively. where the weights are: SP R i = w j = M∑ j=−M w j P R i+j . exp −3|j| M+1 ∑ M −3|k| k=−M exp M+1 Weights are defined to decay exponentially with respect to the distance from the zone in the middle. Finally, the first derivative of a smoothed projection may be obtained using a symmetric difference equation as follows: tel-00912566, version 1 - 2 Dec 2013 2 2h∑ △SP R i (j) = h( h 2 + 1) k. (SP R i (j + k) − SP R i (j − k)) . k=1 where h is set to the closest odd integer value to the mean height of all CCs. The upper and lower bounds of text lines are the local maxima and local minima of △SP R i respectively. Likewise gap regions are identified as the areas between consecutive minima and maxima of △SP R i . Having the first derivatives of the smoothed projection profiles, finding the initial text and gap regions is a matter of applying a threshold to these first derivatives. Figure 5.3 shows the initial text and gap regions for the two image in figures 5.1 and 5.2. Initial separators for separating text lines can be drawn in the middle of each gap in each vertical zone. In this regard, separators are line segments that are defined in the middle of each gap region inside every vertical zone. They are used later to find text lines. 5.2 Refinement of initial text line separators Initial text line separators contain two types of errors: redundant separators resulting from misclassified text and gaps due to poor local extrema, introduced in the smoothed derivatives and separators that cut through descenders or ascenders of text characters. In order to correct these errors and to locate separators more accurately, a Hidden Markov Model (HMM) [78, 91] is formulated with parameters drawn from statistics of the initial text and gap regions. For each document, an HMM is formulated separately on-the-fly and is applied once to each vertical strip of text and gap regions. Viterbi decoding scheme [34] is applied on each zone to obtain a better succession of text and gaps. We briefly note some formal definitions of HMM. Our HMM is characterized by the following items: 88
tel-00912566, version 1 - 2 Dec 2013 (a) Binary image (b) Vertical projection profiles (c) Smoothed projection profiles (d) First derivatives of smoothed profiles Figure 5.1: First steps in line detection to obtain initial lines and gaps for a single document image. 89
- Page 47 and 48: tel-00912566, version 1 - 2 Dec 201
- Page 49 and 50: is assigning a label to a region of
- Page 51 and 52: fixed range. When the elongation ap
- Page 53 and 54: tel-00912566, version 1 - 2 Dec 201
- Page 55 and 56: The second method calculates the co
- Page 57 and 58: 3. Repeat for m = 1, 2, ..., M •
- Page 59 and 60: tel-00912566, version 1 - 2 Dec 201
- Page 61 and 62: Chapter 4 Region detection tel-0091
- Page 63 and 64: The next advantage of using CRFs is
- Page 65 and 66: weights that are assigned to edge a
- Page 67 and 68: { 1 if ys = text and y f 1 (y s , y
- Page 69 and 70: (a) Document (b) Filled text compon
- Page 71 and 72: tel-00912566, version 1 - 2 Dec 201
- Page 73 and 74: tel-00912566, version 1 - 2 Dec 201
- Page 75 and 76: f = [y c = 0] × [y tl = 0] f = [y
- Page 77 and 78: (a) Ground-truth (b) y c = 0 tel-00
- Page 79 and 80: ∂l λ = ∑ ( ∑y∈Y f k (y s ,
- Page 81 and 82: incorrect [100]. Several sufficient
- Page 83 and 84: tel-00912566, version 1 - 2 Dec 201
- Page 85 and 86: tel-00912566, version 1 - 2 Dec 201
- Page 87 and 88: tel-00912566, version 1 - 2 Dec 201
- Page 89 and 90: tel-00912566, version 1 - 2 Dec 201
- Page 91 and 92: Table 4.3: TION COUNT WEIGHTED SUCC
- Page 93 and 94: tel-00912566, version 1 - 2 Dec 201
- Page 95 and 96: tel-00912566, version 1 - 2 Dec 201
- Page 97: Chapter 5 Text line detection tel-0
- Page 101 and 102: tel-00912566, version 1 - 2 Dec 201
- Page 103 and 104: Having specified the model, a verti
- Page 105 and 106: • The fifth step is to remove ext
- Page 107 and 108: tel-00912566, version 1 - 2 Dec 201
- Page 109 and 110: text lines can be divided into two
- Page 111 and 112: the two children. The root node rep
- Page 113 and 114: leaves of the tree which contain on
- Page 115 and 116: tel-00912566, version 1 - 2 Dec 201
- Page 117 and 118: tel-00912566, version 1 - 2 Dec 201
- Page 119 and 120: tel-00912566, version 1 - 2 Dec 201
- Page 121 and 122: currently working on some of these
- Page 123 and 124: • fn (false negative) is the numb
- Page 125 and 126: 2 ∗ RA ∗ DR F − Measure = RA
- Page 127 and 128: • ”-tn”: This option uses the
- Page 129 and 130: [12] T. M. Breuel. Two geometric al
- Page 131 and 132: [39] B. Gatos, A. Antonacopoulos, a
- Page 133 and 134: [64] K. P. Murphy, Y. Weiss, and M.
- Page 135 and 136: [91] M. Stamp. A revealing introduc
- Page 137 and 138: Index tel-00912566, version 1 - 2 D
same location <strong>of</strong> i th zone in sub-figure (a) <strong>an</strong>d (b), respectively.<br />
where the weights are:<br />
SP R i =<br />
w j =<br />
M∑<br />
j=−M<br />
w j P R i+j .<br />
exp −3|j|<br />
M+1<br />
∑ M −3|k|<br />
k=−M<br />
exp<br />
M+1<br />
Weights are defined to decay exponentially with respect to the dist<strong>an</strong>ce from<br />
the zone in the middle.<br />
Finally, the first derivative <strong>of</strong> a smoothed projection may be obtained using<br />
a symmetric difference equation as follows:<br />
tel-00912566, version 1 - 2 Dec 2013<br />
2<br />
2h∑<br />
△SP R i (j) =<br />
h( h 2 + 1) k. (SP R i (j + k) − SP R i (j − k)) .<br />
k=1<br />
where h is set to the closest odd integer value to the me<strong>an</strong> height <strong>of</strong> all<br />
CCs. The upper <strong>an</strong>d lower bounds <strong>of</strong> text lines are the local maxima <strong>an</strong>d local<br />
minima <strong>of</strong> △SP R i respectively. Likewise gap regions are identified as the areas<br />
between consecutive minima <strong>an</strong>d maxima <strong>of</strong> △SP R i .<br />
Having the first derivatives <strong>of</strong> the smoothed projection pr<strong>of</strong>iles, finding the<br />
initial text <strong>an</strong>d gap regions is a matter <strong>of</strong> applying a threshold to these first<br />
derivatives. Figure 5.3 shows the initial text <strong>an</strong>d gap regions for the two image<br />
in figures 5.1 <strong>an</strong>d 5.2. Initial separators for separating text lines c<strong>an</strong> be drawn<br />
in the middle <strong>of</strong> each gap in each vertical zone. In this regard, separators are<br />
line segments that are defined in the middle <strong>of</strong> each gap region inside every<br />
vertical zone. They are used later to find text lines.<br />
5.2 Refinement <strong>of</strong> initial text line separators<br />
Initial text line separators contain two types <strong>of</strong> errors: redund<strong>an</strong>t separators resulting<br />
from misclassified text <strong>an</strong>d gaps due to poor local extrema, introduced in<br />
the smoothed derivatives <strong>an</strong>d separators that cut through descenders or ascenders<br />
<strong>of</strong> text characters. In order to correct these errors <strong>an</strong>d to locate separators<br />
more accurately, a Hidden Markov Model (HMM) [78, 91] is formulated with<br />
parameters drawn from statistics <strong>of</strong> the initial text <strong>an</strong>d gap regions. For each<br />
<strong>document</strong>, <strong>an</strong> HMM is formulated separately on-the-fly <strong>an</strong>d is applied once to<br />
each vertical strip <strong>of</strong> text <strong>an</strong>d gap regions. Viterbi decoding scheme [34] is applied<br />
on each zone to obtain a better succession <strong>of</strong> text <strong>an</strong>d gaps.<br />
We briefly note some formal definitions <strong>of</strong> HMM. Our HMM is characterized<br />
by the following items:<br />
88