1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
|
These are suggestions/remarks sent by Rodrigo Readi.
---
You write, Clara OCR may be seen as a productivity tool for typists. I think
here may play a comfortable graphic interface a big role, and the simple in
outlook, the better in my opinion. It was not the unnecessary scrolling
welcome what convinced me of how efficient is clara. I also hold the
"auto pop-up" item in "Options" menu, "fonts" and "hide scrolbars" items
in "view" menu for unnecessary, perhaps is a configuration file a better
place for changing these defaults. A disturbing thing is the blink of the
image of the scanned page after entering the transliteration of a pattern,
specially disturbing when ones eyes are concentrated on the text arround
the cursor, I was after a time tired because of it. I think it would be
enough if the typed transcription appears written somewhere. The name
"indo-arabic numbers" is not only too long: it sounds ideological. The
decimal system is indian, but the number characters used in the west are
arabic, usually one says "arabic numeral", "cipher" or "cypher" (an arabic
word). The system of the greek alphabet, even the name of the letters and
their order in the alphabet, are phoenician, the characters of the alphabet
are greek, one does not say "punico-greek alphabet".
There is a lot of places for specifying alphabets. The tab "pattern (types)",
the menu "alphabets" (for what is it?), the buttons on the left and the item
"set pattern type" in the edit menu (are they for the same purpose?). I was
not able to specify a pattern type in the "pattern (types)" tab, and my changes
in the parameters "ascent", "descent", DD were ignored, reset to "-1" after
reentering clara. I think this tab is the place for specifying if the font
is "latin type" (with three levels, base, ascent, descent), "arabic type"
(with much more levels and cursive). Perhaps "the "number of levels",
their heigh, the length of words, the fonts name is better data to specify
for each font, and not "latin", "greek", etc. For what is the account? What
is "sk", "bm", "pd", "QTY" (Hints on abrebiations would not be bad)? Note
that one needs much more characters than the ones appearing in the accounts:
ligatures, punctuation, etc. And perhaps one decides to use clara for a
completely unknown alphabet, for which there is no account :)
The tab "page fatbits" is very good for measuring the book fonts. From
there one can get numbers to put in the tab "pattern (types)", like
descent, ascent, DD, XH, etc. It would be nice to have the possibility to
put rulers, as much as one want, for example at the base line, descent line.
Also to have a way for measuring distances and angles by clicking. What you
call zoom is in reality not a zoom: it doesnt go from a general view to a
detail view in the scanned text/fatbits.
In automatic entering of the transliteration, the patterns are set on the
upper link corner. How to distinguish "P" from "p" here? When entering
transliteration of the german Fraktur I had a lot of other problems here.
The context is at this place necessary. I think also this transliteration
could be given in the tab "page": after giving a transliteration, the cursor
could directly jump to the next pattern for which a transliteration is
demanded, instead to the next at the right. It would also be nice to have
the possibility to enter transliterations in the tab "patterns (list)"
directly, but a little of context, perhaps the "web clip" would be, as
noted, necessary. Also the possibility to change data (like transliteration,
and type), as well to delete a pattern directly in "patterns (list)" would
be a nice thing.
My atempts to change parameters in the tab "tune (skeleton)" were also ignored
and reset to the old values after reentering clara. I couldnt reclasify symbols
after adding new patterns, although I selected "Re-scan all patterns" in
"Edit" menu. It would be nice to have the possibility to force a classification
without using a pattern and without using the symbol as pattern: for example to
tell clara that a symbol is "d", although it appears to be a perfect "a". In
general one cannot distinguish very well if clara is dealing with a pattern or
with a symbol. The following question has in my oppinion now answer:
"click on better alternative"
r "b" (pattern)
4 "4" (symbol)
both are "b"
both are "4"
symbol is "4", pattern "b"
What is the best??? There should be a way to directly say what is the
pattern and what is the symbol, without "multiple choice". Sometimes there
does not appear any pattern in this questions. -- The information in the
last window in "page" tab is not understandable: what does mean
"best match 3160, shape status 1 (best match 3268)"? After clicking on
3160 one gets "pattern 3268 (id 3377), from act 3494", and after clicking
3494 one gets "symbol 25". Too much unexplained numbers and IDs!
I think that making tools for "Computing Aided Revision" of scanned and
OCRed text is also a challenge. Someone asked a scientist at the University
of Tuebingen to correct a latin text that lied in ASCII, the scientist
answered that he preffers to type the whole text than to read and find the
errors. He typed the text, then it was compared with the original one, and the
places where this two texts differed, were corrected. This is the origin
of a technik used by those that let text be typed in China or Rumenia:
the text is typed twice and the two versions compared. If you have two
good classification criteria and dont want to mix them, you can make
clara give, after batch OCR, a "clickable" report with the places where
both clasifiers differ. Another idea is the following: after clicking
a classified symbol in the image of the scanned text, appears a black
elipse arround the symbol, and grey elipses arround the ones classified
with the same pattern, at this point one should able to change the
transliteration of the pattern and of some symbols without altering the
pattern. We can call this symbols as checked. When one clicks another symbol,
the checked ones become unmarked and new ones are marked. Well, the idea would
be to select a "revision mode" in a menu, so that the checked symbols
become marked as "checked", at best presented with a very light grey tone,
to show that one has not anymore to care of these symbols (but not to
confuse with the unclassified ones that also appear with a light gray tone).
The game would be to revise all symbols, but no one more tham once.
When OCRing a "?" symbol, what should show the "status line" when the
symbol was succesfully classified and the transliteration (=?) found?
Unfortunately the symbol "?" is used for the untransliterated patterns!
Please, think on the possibility of other accents, like the german Umlaut
(two points over a, o or u, in TeX \"{a}, \"{o}, \"{u}). Another issue
are the "text-glosses": They are words or sentences in a very small
fontsize between the lines of texts, for making commentaries or translations
of the main text. When I use the right-arrow-key to navigate on
the image of the scanned text, the cursor jumps from the main text to
the glosses, and then comes again down to the main text (and forget what
was in between). Something similar happens when I have an "i" and a letter
with accent: it moves through the accents and comes down again. What should
happen in the output? (I was not able to get a readable output, it ommits
a lot).
I would be interested in having the possibility to decide how the
output looks like, for later processing the output with a script.
I use for example tcl and CoSt for processing SGML files, and a
SGML output in a way that I specify would be interesting, and not
TeX or HTML.
The Idea of integrating clara with GUILE is excellent, specially if it
allows to easily program new OCR strategies, perhaps text-dependent
strategies. I was thinking the whole time on "classification strategies",
perhaps I write you later some ideas. I would be glad if I hear from you
how to change the tuning parameters, as said, it doesnt work. I would like
to experiment with it.
---
(1) Instead of checking that the skeleton of the pattern is in the symbol
and that the skeleton of the symbol is in the pattern, one could check
that the skeleton of the pattern is in the symbol and that both, pattern
an symbol are (almost) equally big. The idea is: if the set A is included
in B and both, A and B, have the same number of elements, then A and B
are equal. This would allow to explicitely define skeletons, without
functions that calculate them. The functions are necessary because you
calculate the skeletons of the symbols to be automatically classified.
Of course, training could mean to compute a skeleton with an algorithm
on the basis of symbols told the computer to be equal, but also modifying
the computed skeleton, or just directly entering the skeleton.
(2) Well, one can not only check if a pattern is included in the symbol,
but also that a set of forbidden bits are 0.
(3) One can make (2) smooth. We can characterize a pattern (mean as a muster
symbol) with a "reward measure", a "punishment measure" and two
numbers "a" and "b". Each of these measures consists of a number
between 0 and 1 for each point in the bitmap. The reward measure has
higher values at points that we wish to be black, the punishment at
those we wish white. We can consider the symbol and both measures as
a vector. The reward for the symbol is the scalar product of the
symbols vector with the reward measure. The punishment the scalar
product with the punishment measure. The task is to find the measures,
a and b for a pattern, so that a symbol corresponds to the pattern if
and only if the punishment is below a and the reward over b. Of course,
if we are working with grayscale, we must "normalize" the symbol (divide
by the norm/ a norm). The measures could be calculated by taking the
arithmetical mean of the bits (and their complements for the punishment)
of some symbols the user tell the computer to belong to a pattern.
(4) As I came to (3), I noted that this is a statistical thema. I dont
have much idea on statistics, I never liked it. I asked someone about,
he told me that this kind of classification strategies are studied
under the thema "multivariates methods". There are strategies for
finding the groups and for deciding what bolongs to a group.
(5) Note that I am far away from your original strategy. You select patterns
from the scanned text, I propose to use scanned symbols to create
"artificial" patterns, they are nowhere in the scanned text, furthermore,
they are not bitmaps, but measures, numbers that can also be modified
by the user. Of course, for the symbol "a" we can still have many patterns,
it is the user who says what is similar to what. I think it would be
interesting to have programming tools for easily experimenting with OCR
strategies, among them tools for managing databases of symbols clasiffied
by the user.
---
Another thought is to try to characterize patterns by many kinds
of measures, not only height and width, but also centre of mass
(also useful for identifying points of two symbols) and anything
we can imagine.
---
|
