New Features for the Recognition of German-Kurrent-Handwriting
with HMM-based Offline Systems
Klaus Prätel
Department of Computer Vision and Remote Sensing, TU Berlin, Marchstr. 23, 10587 Berlin, Germany
Keywords: New Geometrical Features, German Cursive.
Abstract: In 2007, the project Herbar-Digital was launched at the University of Applied Sciences and Arts Hannover
(Steinke, K.-H., Dzido, R., Gehrke, M., Prätel, K., 2008). The aim of this project is to realize a global
Herbarium, to compare findings quickly and catalog. There are many herbaria, i.e. collections of herbarium
specimens worldwide. Herbarium specimens are paper pages where botanical elements are glued on. This
herbarium specimens are provided with some important clues, such as the name of the submitter, barcode,
color table, flag for first record, description of the findings, this often handwritten. All information on the
herbarium specimens should be evaluated digitally. Since a number of discoveries in the 19th Century took
place, Alexander von Humboldt is to be mentioned, here is the challenge to identify specific manuscripts from
this period. This paper describes the topic recognition of old German handwriting (cursive).
1 INTRODUCTION
Approach to handwriting recognition:
Writer independent offline handwriting recognition.
Based on architecture of Hidden Markov Models.
Module for extracting text lines from image data.
Modules for pre-processing such as Slant, Slope, and
Scale.
Sliding-Window Serialization.
Automatic limitation of the model states as a function
of training material (30 as the upper limit).
Parameter-Estimation using Standard-Baum-Welch-
Training.
Decoding (recognition) on the Viterbi algorithm.
Semi-continuous HMMs (state reduction) for Writing
Model.
Statistical n-gram models as language models.
Linear and Bakis topology.
Figure 1: Schematic representation of a typical architecture
of a handwriting recognition system (Fink, G. A., 2008).
Preprocessing:
- Binarization
Separation of the foreground from the
background.
- Skeletonization
Thinning by line following, reduction to one-
pixel thickness.
- Determination of reference lines
Upper limit line, midline, baseline, lower
profile
366
Prätel, K.
New Features for the Recognition of German-Kurrent-Handwriting with HMM-based Offline Systems.
DOI: 10.5220/0006185403660371
In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 366-371
ISBN: 978-989-758-222-6
Copyright
c
 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

- Correction of font orientation
Estimate of the font baseline
- Correction of font size
- Correction of the slant
- Feature extraction
Comparison of German cursive hand with the
simplified output font:
As you can see the writing samples, a cursive
hand, here just the large letters, is characterized by
significantly increased crossing processes (possibly
multiple entry between upper and lower contour) than
in the Simplified output font.
German cursive (uppercase)
A B C D E F G H IJ K L M N O P Q R S T U
V W X Y Z
Simplified output font (uppercase)
A B C D E F G H I J K L
M N O P Q R S T U V W X
Y Z
German cursive (lowercase)
a b c d e f g h i j k l m n o p q r s t
u v w x y z
Simplified output font (lowercase)
a b c d e f g h i j k
l m n o p q r s t u v
w x y z
Feature extraction
Now features are determined to make the following
step, a classification can be used for the pre-
segments. On one hand, the features may be
discriminative, being very different from each other
when they come from different classes. On the other
hand, they should differ little when they come from
the same class. In order to achieve this, to realize this,
the peculiarities of the German cursive must be
considered. Knowledge: the feature extraction must
be extended. The following are the major principles
of the feature extraction are described first of all to
create the feature vector:
In order to allow sufficiently accurate detections,
the preprocessing must map the Characteristic
Scripture extensively, the resulting amount of data
must be restricted in order to allow decoding, i.e.
recognition, in an acceptable time (online and
offline).
To describe handwriting sufficiently accurate in
vectors it here needs editing of 11 features, which are
described below. First, nine features are explained
briefly, then describes the additional features are
needed to better recognize cursive.
In this case, geometric features are examined the
subject sliding window will be described briefly.
Sliding-Window: A systematic subdivision of the
lettering after the pitch- and orientation-correction.
These corrections are only useful when a
handwriting-recognition is given. In a realization of a
writer-recognition, this would be counter-productive,
because writer-specific features would be eliminated.
It is pushed from left to right, a window of the text
line. The height of the window is based on the font
height, the window width normally is 4 pixels that
overlap two pixels. These parameters are adjustable.
It is shown that an increase of 14 pixels width and 13
pixels overlap window to significantly better results.
Figure 2: Previous features (Wienecke, M., Fink, G. A., and
Sagerer, G., 2005).
(a) Mean distance of the upper outline of the type
to the baseline
(b) Mean distance of the lower outline of the type
to the baseline
(c) Average distance between the y-coordinates of
the focal points columns
(d) Orientation of the upper contour
(e) Orientation of the lower contour
New Features for the Recognition of German-Kurrent-Handwriting with HMM-based Offline Systems
367

(f) Orientation of the course of columns priorities
(g) Average number of vertical header
background transitions
(h) Mean Number of font pixels per image
column
(i) Average number of font pixels between the
upper and lower outline of the type
In the used system 10 feature-extractions are realized
actually. The 10th feature is merely a small
expansion, therefore it is not mentioned here.
As is evident from the representation of the
feature extraction described, the position (s) and the
angle of the lines (the) possibly multiple passage is
not taken into account.
2 NEW FEATURE STRATEGIES
Therefore, two new features have to be introduced,
which can occur more than once.
1. Passage distance from the base line (k)
2. Orientation of the passage line for the parallel to
the base line (j)
Figure 3: Principle of a new feature.
Figure 4: The example with a puncture in the realization.
The feature 1 (k) and the feature 2 (j) have been
implemented and the results are evaluated.
Within the framework of the tests it was stated
that people must deviate from the conventional
procedure of the preprocessing. Usually a correction
of the writing level (Slope), the letter inclination
(Slant) as well as the standardization of the font size
(Scale) is carried out in the case of the word
recognition. Slope as well as Scale are used wider,
Slant not.
Reason: The intersection characteristics are falsified
by the erection of the letters in such a way that the
recognition result is worsened.
Other considerations:
Figure 5: Multiple middle passages are possible, see
German cursive capital letter K.
The realization of the extended intersection
number takes place in a later publication.
Further feature:
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
368

Furthermore, it is considered whether the
relationship between number of pixels of the top and
bottom edge of the font is set in each pixel gap in
relation to better accommodate to average values of
the whole window, the baseline differences in the
analysis. The further consideration and the
implementation will be published in another
publication.
Description of the recognition system
As a first step, a vector quantization is performed. The
object is to reduce the data amount. For this reason,
vectors from the input space are mapped to typical
representatives, so similar data vectors are grouped
together. The goal is to determine via cluster analysis
accumulation ranges of unknown data distributions
and to describe. Are formed mean vectors that are
then representative of a statistical quantization.
Set Y = y1, y2, ... yN prototype vectors yi is also
referred to as a codebook.
A typical quantization algorithm is the K-means
algorithm today.
A brief description:
1. Initialization: Random selection of k cluster
centers or the first N vectors of the sample.
2. Assignment: Each object is assigned to him
closest cluster center, i.e. determination of
optimal reproduction vector in the current
codebook.
3. Recalculation: It will recalculate the cluster
centers for each cluster, i.e. determination of a
new code book.
4. Repeat: If the assignment changes, go to step 2,
otherwise finished.
Scripture modeling
The static signature modeling is implemented via
semi-continuous hidden Markov models. Under
modeling is understood to mean the training of the
training material. There models are so trained that
correspond to the training material.
Under decoding is understood to mean the detection,
so the determination of the probability that the test
pattern corresponding to a model. When modeling
with HMM (Hidden Markov Model), we consider
two modeling components:
The Writing Model: Hidden Markov Model
On word or letter level:
Language model (lexicon)
Markov Chain Model (n-gram model
(Brakensiek, A., Rottland, J., and Rigoll, G.,
2002))
The combination of both models provides a
powerful system for the representation of
handwriting. The parameters of the models can be
predicted automatically.
Decoding (recognition):
So-called Decoding the combined model, meaning
the optimal path through the combined state space. It
is achieved optimal segmentation and classification in
a continuous system. The Markov model concept
describes the analysis of sequential data:
The HMM as a statistical model is now "state of the
art".
The recognition model:
Sequence of symbols (such as words)
w: Implementation in sequence of feature vectors X.
Objective of the recognition process:
Aim of the recognition process: Find the sequence
which maximizes the posterior-probability P(w|X) of
the symbol sequence of the given data.
^argmax




|


argmax









|







 arg max








|


Application Bayestheorem:
posterior probability P (w | x) is implemented in the
form in which the two component model of a Markov
model are obvious:
P (w) = language model probability of symbol
sequence w
n-gram model (Language Model).
P (w | X) = probability of observing the symbol
sequences as features X, according to the writing
model, namely the Hidden Markov Model (HMM).
Modeling:
HIDDEN MARKOV MODEL
General:
Using a hidden Markov model (Vinciarelli, A.,
Bengio, S., and Bunke, H., 2004) is trying to
determine the probability with which a given feature
Bayestheorem
New Features for the Recognition of German-Kurrent-Handwriting with HMM-based Offline Systems
369

sequence (simply put: one-word commands) belongs
to a particular word. For each word in the lexicon, an
HMM is (i.e. for each word to be recognized) created.
Once an input value (word) is present (observation),
this is quasi compared with each HMM. The best
result represents the recognized word. Verification
with HMM is a statistical method. Determining the
degree of match between the test sample and the
reference as a probability.
2-piece stochastic process:
First Part: State machine: Describes probabilities of
the transitions of the states.
Second Part: modeling the output of the font pattern
when entering the respective state output probability
distributions.
Allocation of the state to the magazine segment.
Feature of the chain is determined by the
characteristics of the segments.
Training: Baum-Welch algorithm (optimization of
HMM)
Improvement of a given model in response to certain
sample data (training pattern) in such a way that the
optimized model generates the training set with equal
or greater probability.
Decoding: Viterbi algorithm
Calculation of the optimal path through the state
sequence (Viterbi path). So the maximum probability
of generating the observation sequence.
Is used to "discharging" of the hidden state
sequence, that generates a maximum likelihood, a
valid sequence of outputs, is given by the model.
3 RESULTS
Description of the database:
To make recognition systems comparable, this system
must come on a standardized database to the
application. In the case of the Old German
handwriting ( Kurrent ) this is not possible, since such
a database does not exist. It had to be redesigned a
database. This database consists of approximately
6000 samples. This is made up of collections of 6
writers, here also some of Alexander von Humboldt.
Characteristic lines and characteristic features of the
writing thickness allow conclusions on the writer.
First, the test was carried out with 110 test images.
The recognition rate was 0.4234. In order to present
an apparent delta, more 440 test images were
deducted from the training data.
Then the model 11 and the model features 9
characteristics was examined. The window setting
was w8X4 .
The term dimension in this context means: features or
characteristics.
Table 1: Without correction (Slope, Slant, Scale).
In the model with two additional features (12
dimensions), the difference is clear. At 550 test
images, the recognition rate was increased by
approximately 0.1!
Table 2: Without correction (Slant).
Testimages 10-Dimensions 12-Dimensions
110 0.5834 0.6038
550 0.5195 0.6284
Table 3: With correction (Slope, Slant, Scale).
Testimages 10-Dimensions 12-Dimensions
110 0.5636 0.554545
550 0.549091 0.516364
1100 0.534545 0.495455
4 CONCLUSIONS
Other Windows settings are investigated.
(e.g. w14X13).
In order to improve the recognition result, the number
of samples must be increased noticeably. In this
publication should be shown merely, that through an
extension of the feature extractions, here specific for
Old German handwriting (Kurrent), an improvement
of the recognition is reached.
REFERENCES
Fink, G. A., 2008. Markov Models for Pattern Recognition,
From Theory to Applications. Springer, Heidelberg.
Wienecke, M., Fink, G. A., and Sagerer, G., 2005. Toward
Automatic Video-based Whiteboard Reading. Int.
Journal on Document Analysis and Recognition, vol.
Testimages 10-Dimensions 12-Dimensions
110 0.4234 0.4545
550 0.327 0.425
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
370

7(2–3):188–200.
Steinke, K.-H.; Dzido, R.; Gehrke, M.; Prätel, K., 2008.
Featurerecognition for herbarium specimens (Herbar-
Digital), Proceedings of TDWG, Perth.
Vinciarelli, A., Bengio, S., and Bunke, H., 2004. Offline
Recognition of Unconstrained Handwritten Texts using
HMMs and Statistical Language Models. IEEE Trans.
on Pattern Analysis and Machine Intelligence, vol.
26(6):709–720.
Brakensiek, A., Rottland, J., and Rigoll, G., 2002.
Handwritten Address Recognition with Open
Vocabulary Using Character N-Grams. In Proc. Int.
Workshop on Frontiers in Handwriting Recognition,
pages 357–362. Niagara on the Lake, Canada.
Fink, G. A., 1999. Developing HMM-based Recognizers
with ESMERALDA.
In V. Matouˇsek, P. Mautner, J.
Ocel´ıkov´a, and P. Sojka, eds.,
Text, Speech and Dialogue ,
vol. 1692 of
Lecture Notes in Artificial Intelligence , pages
229–234. Springer, Berlin Heidelberg.
New Features for the Recognition of German-Kurrent-Handwriting with HMM-based Offline Systems
371