See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221048185

Building Digital Ink Recognizers Using Data Mining: Distinguishing between

Text and Shapes in Hand Drawn Diagrams

Conference Paper · June 2010

DOI: 10.1007/978-3-642-13022-9_36 · Source: DBLP

CITATIONS READS

3 55

4 authors, including:

Beryl Plimmer John Grundy

University of Auckland Monash University (Australia)
132 PUBLICATIONS   1,189 CITATIONS    517 PUBLICATIONS   6,777 CITATIONS   

SEE PROFILE SEE PROFILE

Yong Wang
University of Auckland
21 PUBLICATIONS   213 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Software Engineering Tools and Software Engineering Environments View project

Software Engineering for Big Data and Cloud Computing View project

All content following this page was uploaded by Beryl Plimmer on 04 June 2014.

The user has requested enhancement of the downloaded file.

 Building Digital Ink Recognizers using Data Mining:
Distinguishing Between Text and Shapes in Hand Drawn
Diagrams

Rachel Blagojevic1, 1, Beryl Plimmer1, 3, John Grundy2, 2, Yong Wang1, 4

1
University of Auckland, Private bag 92019, Auckland, New Zealand
2
Swinburne University of Technology, PO Box 218, Hawthorn, Victoria, Australia 3122
1 2
[email protected], [email protected],
{3beryl@cs., 4yongwang@}auckland.ac.nz,

Abstract. The low accuracy rates of text-shape dividers for digital ink diagrams
are hindering their use in real world applications. While recognition of
handwriting is well advanced and there have been many recognition approaches
proposed for hand drawn sketches, there has been less attention on the division
of text and drawing. The choice of features and algorithms is critical to the
success of the recognition, yet heuristics currently form the basis of selection.
We propose the use of data mining techniques to automate the process of
building text-shape recognizers. This systematic approach identifies the
algorithms best suited to the specific problem and generates the trained
recognizer. We have generated dividers using data mining and training with
diagrams from three domains. The evaluation of our new recognizer on realistic
diagrams from two different domains, against two other recognizers shows it to
be more successful at dividing shapes and text with 95.2% of strokes correctly
classified compared with 86.9% and 83.3% for the two others.

Keywords: Sketch tools, recognition algorithms, sketch recognition, pen-based

interfaces.

1 Introduction
Hand drawn pen and paper sketches are commonplace for capturing early phase
designs and diagrams. Pen and paper offers an unconstrained space suitable for quick
construction and allow for ambiguity. With recent advances in hardware such as
Tablet PC’s, computer based sketch tools offer a similar pen-based interaction
experience. In addition, these computer based tools can benefit from the ease of
digital storage, transmission and archiving. Recognition of sketches can add even
greater value to these tools. The ability to automatically identify elements in a sketch
allows us to support tasks such as intelligent editing, execution, conversion and
animation of the sketches.
2 Rachel Blagojevic1, 1, Beryl Plimmer1, 3, John Grundy2, 2, Yong Wang1, 4

a) Directed graph b) Organization chart c) User interface

Fig. 1. Example sketched diagrams for training set
A number of sketch tools have been developed, however they are yet to achieve
general acceptance. One of the outstanding challenges is considerably more accurate
recognition. Recognition rates from laboratory experiments are typically in the range
of 98% to 99% and above [1-3]. However, rates achieved in less controlled
conditions, where data is not limited to produce optimal performance, are usually
much lower, for example accuracy rates between 84% and 93% are reported in [3-6].
Furthermore, many of these tools are limited as they are not able to distinguish
between drawing elements (shapes) and text strokes in a sketch [1-3]. Most natural
diagrams consist of both writing and drawing as shown in figure 1.
While recognition of handwriting is well advanced and there have been many
recognition approaches proposed for hand drawn sketches, there has been less
attention on the division of text and drawing. People can comprehend writing and
drawing seamlessly, yet there is a clear semantic divide that suggests, from a
computational perspective, it is sensible to deal with them separately. Several
recognizers [7-9], commonly referred to as dividers, have been proposed for this
purpose, but recognition rates in realistic situations are still unacceptable. Limited
investigation of machine learning for text-shape division has been found to be
effective [8]. This work extends [8] by drawing on a larger set of ink features and
using a range of data mining techniques systematically selected and tuned.

2 Background
Two particular applications of dividers are freehand note-taking and hand drawn
diagrams. The research on sketched diagram recognition includes dividers but has
also addressed recognition of basic shapes and spatial relationships between diagram
components. This project has drawn on the work from both applications of dividers.
The majority of recognizers rely on information provided by various measurements
of the digital ink strokes (digital ink is represented as a vector of x, y points, each
point has a time and possible pressure attribute) [1, 2, 10], as well as specific
algorithms to combine and select the appropriate features.
In the area of sketched diagram recognition many systems focus only on shapes [1-
3]. There have been some attempts at incorporating text-shape division in domain
specific recognizers [11, 12] and domain independent diagramming tools [10, 13].
These systems are predominantly rule-based, using stroke features chosen
heuristically to distinguish between text and shapes.
 Building Digital Ink Recognizers using Data Mining: Distinguishing Between Text and
Shapes in Hand Drawn Diagrams 3
Research in the area of digital ink document analysis for freehand note-taking has
explored text-shape division [14-19]. However as the content of documents is mainly
text these methods hold some bias which make them unsuitable for sketched
diagrams. In addition, as Bhat and Hammond [7] point out, some of these methods
would have difficulty with text interspersed within a diagram. There has also been
some work separating Japanese characters from shapes in documents [18, 20].
Three reports specifically on dividers are [7-9]. Bishop et al [8] use local stroke
features and spatial and temporal context within an HMM to distinguish between text
and shape strokes. They found that using local features and temporal context was
successful. They report classification rates from 86.4% to 97.0% for three classifier
model variations.
In our previous work [9] we developed a domain independent divider for shapes
and text based on statistical analysis of 46 stroke features. A decision tree was built
identifying eight features as significant for distinguishing between shapes and text.
The results on a test set showed an accuracy of 78.6% for text and 57.9% for shapes.
Part of the test set was composed of musical notes which had a significant effect on
this low classification rate. However, when evaluated against the Microsoft and
InkKit dividers, it was able to correctly classify more strokes overall for the test set.
A more recent development in this field is the use of a feature called entropy [7] to
distinguish between shapes and text. First strokes are grouped into shapes and
words/letters and then stroke points are re-sampled for smoothing. The angle between
every point and its adjacent points in the stroke group is calculated. Each angle from
the stroke group is matched to a dictionary containing a different alphabet symbol to
represent a range of angles. This results in a text string representation of each stroke
group. Using Shannon’s entropy formula (as cited by Bhat et al [7]) they sum up the
probabilities of each letter in the string to find the entropy of that group. This value of
entropy is higher for text than shapes as text is more “information dense” than shapes.
They report that 92.06% of data which it had training examples for were correctly
classified. For data the divider had not been trained on it had an accuracy of 96.42%,
however only 71.06% of data was able to be classified. We have re-implemented this
algorithm for our evaluation. As our evaluation will show, this divider has been
trained and tested on limited data and constrained conditions and does not perform at
the reported rate of 92.06% on realistic diagrams.
The choice of features and algorithms is critical to the success of the recognition,
yet heuristics currently form the basis of selection. Given that features provide such
value as input to recognition algorithms, a feature set should be chosen carefully
using statistical or data mining techniques. While others have used some data mining
techniques [8, 15] to the best of our knowledge no one has done a comprehensive
analysis of algorithms. We present below a comprehensive comparative study of
features and algorithms to select the most accurate model. In particular we are looking
at the problem of distinguishing between text and shapes as a first step to recognizing
sketched diagrams; a fundamental problem required to preserve a non-modal user
interface similar to pen and paper.
4 Rachel Blagojevic1, 1, Beryl Plimmer1, 3, John Grundy2, 2, Yong Wang1, 4

3 Our Approach
In order to use data mining techniques to build classifiers we first compiled a
comprehensive feature library which is used in conjunction with our training set of
diagrams to generate a training dataset. We investigated a wide range of data mining
algorithms before focusing on seven that were producing the most promising results.
These seven algorithms and the training dataset were used to build new dividers.

3.1 Features

Our previous feature set [9] of 46 features has been extended to a more
comprehensive library of 115 stroke features for sketch recognition. It has been
assembled from previous work in sketch recognition, includes some of our own
additions, Entropy [7], and our previous divider [9]. Our previous divider is used for
several features: pre-classification of the current stroke, pre-classification of strokes
close by (for spatial context), and pre-classification of successive strokes (for
temporal context).
Many researchers have developed features that measure similar attributes. In order
to give the reader some sense of the types of features we have categorized the feature
library into ten categories, summarized in table 1.
This feature library is available with full implementation within DataManager [21]
from www.cs.auckland.ac.nz/research/hci/downloads.

Table 1. Summary of stroke feature categories.

1. Curvature (e.g. the line above 6. Pressure (measure the pressure applied to
has a greater curvature than the the screen when drawing a stroke. Pressure is
line below). dependent on the capabilities of the hardware).
2. Density (e.g. the text has larger density of 7. Size
points than the shape).
3. Direction (this is related to the 8. Spatial context (with sub categories:
slope of the stroke). curvature, density, divider results,
intersections, location and size).
4. Divider Results (these features 9. Temporal context (with sub categories:
provide the results of text/shape curvature, density, divider results, length,
divider algorithms). location/distance and time/speed).
5. Intersections (e.g. the diagram 10. Time / speed (includes total, average,
shows intersecting strokes). maximum and minimum times or speed).

3.2 Dataset

For the training set we have collected and labeled sketched diagrams from 20
participants using DataManager [21]. Each participant has drawn three diagrams; a
directed graph, organization chart and a user interface e.g. figure 1. There are a total
of 7248 strokes in the training set, with 5616 text strokes and 1632 shape strokes.
 Building Digital Ink Recognizers using Data Mining: Distinguishing Between Text and
Shapes in Hand Drawn Diagrams 5
Using this collection of diagrams we have generated a dataset of feature vectors for
each stroke using DataManager. DataManager’s dataset generator function is able to
take the diagrams collected and calculate feature vectors based on the implementation
of our feature library.

3.3 Building Classifiers

Weka (developer version 3.7) [22], an open source data mining tool, has a large
number of machine learning algorithms that can be used to perform our data analysis
and build, tune and test classifier models for dividers. We found that 60 of the
algorithms within Weka were possibly suited to the divider problem. We began our
analysis with a preliminary investigation of all these algorithms. This involved
building classifier models for each algorithm using the training data. Some clearly
performed better than others while some needed tuning of their specific parameters to
optimize their results. Upon discussion of the preliminary investigation we were able
to narrow the search down to seven algorithms that are likely to gain the best
classification accuracy for a divider1.
The chosen classifiers are: Bagging [23](with an REP tree base learner), LADTree
[24](alternating decision tree using the LogitBoost strategy), LMT [25](logistic model
tree), LogitBoost [26](additive logistic regression with Decision Stump or REP tree
base learner), MultilayerPerceptron [22] (neural network), RandomForest [27](forest
of random trees) and SMO [28](support vector machine). Using the training dataset of
feature vectors generated from the diagrams collected we built dividers by training
each classifier. While Weka provides sensible default parameters for most algorithms,
some classifiers required tuning to optimize their results.
For Bagging [23] we tuned the algorithm by varying the number of bagging
iterations that the algorithm performs. This parameter is indicative of the number of
trees that can be produced. The default value for this in Weka is 10 iterations. We ran
an experiment using 10-fold cross validation for Bagging with REPTree (a fast
decision tree learner) at 10, 100, 500, 1000 and 5000 iterations. Paired t-tests (α=0.05)
showed no significant difference in the results at each level of iterations. The highest
result is shown in table 2, this was produced at 500 and 1000 iterations.
To tune the LADTree [24] we varied the number of iterations of the algorithm to
10, 100, 500 and 1000 iterations. We were unable to increase the number of iterations
to greater than 1000 due to time and memory constraints. This algorithm takes a long
time to train therefore we chose to run the experiment with 5-fold cross validation as
opposed to 10-fold. Paired t-tests (α=0.05) showed that the LADTree with 500 and
1000 iterations were significantly more accurate than the others. There was no
significant difference between the LADTrees with 500 and 1000 iterations. The
highest result shown in table 2 was produced at 1000 iterations.
The default parameters in Weka for LMT [25] are sensible for this problem and did
not require tuning. The result of 10-fold cross validation using our training dataset on
LMT with default parameters is shown in table 2.
To begin tuning LogitBoost [26] we ran a preliminary 10-fold cross validation
experiment to see if there were any significant differences in using Decision Stump or

1 Thanks to Eibe Frank for his advice on the selection of algorithms.

6 Rachel Blagojevic1, 1, Beryl Plimmer1, 3, John Grundy2, 2, Yong Wang1, 4

REPTree as a base classifier for LogitBoost. A paired t-test (α=0.05) showed no

significant difference between the two at 120 iterations of the algorithm. Based on
these results we decided to continue to investigate both trees as base classifiers.
To further tune LogitBoost we varied the number of iterations the algorithm
performs and also the shrinkage parameter. Shrinkage is a parameter that can be tuned
to avoid overfitting the LogitBoost model to the training dataset. When a classifier is
overfitted it reduces the likelihood of the model retaining the same level of accuracy,
achieved with training data, on a new test dataset. Small values for shrinkage reduce
overfitting. We ran experiments using 10-fold cross validation for LogitBoost with
the following options: base classifier as a Decision Stump or REPTree; number of
iterations at 10, 100, 500, 1000 or 5000; shrinkage at 1.0 (Weka default value) or 0.1.
Using all combinations of the above options resulted in 20 models for LogitBoost,
10 for each base classifier. For LogitBoost, the model that had the highest level of
accuracy was with a Decision Stump base classifier, 5000 iterations and a shrinkage
value of 0.1, the result for this model is shown in table 2. Paired t-tests (α=0.05)
showed that it was significantly better than all other Decision Stump models except
two that were not significantly different; they had a shrinkage value of 1.0 and
number of iterations set at 1000 and 5000. When compared with the REPTree models,
it was only significantly better than the REPTree model with 10 iterations at a
shrinkage value of 0.1, for all
Table 2. Best results obtained from selected
others there was no significant
classifiers
difference.
% Correctly classified
Classifier The default parameters in Weka
(10-fold cross validation) for MultilayerPerceptron [22] are
LADTree 97.49* (5-fold) sensible for this problem and did
not require tuning. The result of
LogitBoost 96.70* 10-fold cross validation using the
RandomForest 96.45 training dataset on
MultilayerPerceptron with default
SMO 96.41
parameters is shown in table 2.
Bagging 95.67 To tune RandomForest [27] we
varied the number of iterations of
MultilayerPerceptron 95.02
the algorithm to 10, 100, 500 and
LMT 94.85 1000 iterations. We were unable to
increase the number of iterations to
greater than 1000 due to memory constraints. Paired t-tests (α=0.05) showed no
significant difference between any of the models. The highest result shown in table 2
was produced at 500 iterations.
SMO [28] is a more complicated classifier to tune. There are two parameters that
can be tuned; the complexity value of SMO and the gamma value of the RBF kernel
used by SMO. To find the best model we used the GridSearch function in Weka
which allows you to optimize two parameters of an algorithm by setting a maximum,
minimum, base value and step value for how much a parameter can increase by for
each test. One of the main advantages of GridSearch is that the parameters of interest
do not have to be first level parameters, for example gamma is not a first level
parameter as it is a value used by the RBF kernel, where the RBF kernel is a
parameter of SMO. We found the optimal value for complexity was 100, with a
 Building Digital Ink Recognizers using Data Mining: Distinguishing Between Text and
Shapes in Hand Drawn Diagrams 7
gamma value of 0.1. The results of 10-fold cross validation on SMO for our training
data is shown in table 2.
The best results of 10-fold cross validation (except LADTree which was 5-fold) for
each classifier on our training set is shown in table 2. Paired t-tests (α=0.05) show that
LogitBoost and LADTree are significantly better than the other classifiers. There is no
significant difference between LogitBoost and LADTree. This is not surprising as
LADTree uses the LogitBoost strategy.

3.4 Implementation

In order to run a comparative evaluation of our two new models against other dividers
we integrated our models into DataManager’s Evaluator [6]. We also integrated our
old divider [9] and implemented the Entropy divider [7].
The Entropy divider had to be trained as no thresholds were provided by [7]. We
trained it on the same data as our new dividers using 10-fold cross validation with the
decision stump algorithm from Weka [22] to find an optimal threshold. We chose the
decision stump algorithm as this generates a decision tree with one node, essentially
producing one decision based on the Entropy feature. The 10-fold cross validation
reported that 85.76% of the training data was correctly classified; other algorithms
such as OneR, a rule based method, and a J48 tree (C4.5 decision tree) showed similar
results. Our divider developed from previous work [9] was not re-trained; it was
implemented with the same thresholds as the original decision tree.

4 Evaluation
In order to test the accuracy of these dividers on data that they are not trained on we
used a new set of diagrams from different domains to the training set. The test set was
composed of ER and process diagrams (see figure 2) collected from 33 participants
who drew one diagram from each domain. The participants were asked to construct
the diagrams from text descriptions so that they are realistic in individual drawing.
There are a total of 7062 strokes in our test set which is similar in size to our training
set. There are 4817 text strokes and 2245 shape strokes. Table 3 shows the results for
each divider on the test set of diagrams. LADTree is the most accurate of the four
tested with 95.2% correctly classified closely followed by LogitBoost at 95.0%. The
Entropy divider is the least accurate at a rate of 83.3%. It is clear that entropy has a
large bias towards text as only 50.5% of the shapes in the test set are correctly

a)ER diagram b) Process diagram

Fig. 2. Example sketched diagrams for test set
8 Rachel Blagojevic1, 1, Beryl Plimmer1, 3, John Grundy2, 2, Yong Wang1, 4

Table 3. % Correctly classified for each divider. classified. Our previous divider is
Divider % Correct % Text % Shapes slightly more accurate than Entropy;
however its bias towards text is not
LADTree 95.2 98.3 88.5
as extreme. In fact the results show
LogitBoost 95.0 98.1 88.4 that all dividers classify text much
Old Divider 86.9 93.1 73.5 more accurately than shapes.

Entropy 83.3 98.7 50.5

5 Discussion
The high accuracy of the results we have obtained by using data mining techniques to
build dividers demonstrates the effectiveness of this approach. We believe that other
recognition problems would also benefit from a similar study of data mining
techniques. However there is still room for improvement in these divider algorithms.
In terms of tuning, for all the algorithms where we varied the number of iterations
we found that a high number of iterations usually resulted in significantly better
results. We could tune these further by increasing the number of iterations for some
algorithms however we are constrained by time and memory. Although, these
constraints are for training, once the classifier is trained the memory requirements are
minimal and actual classification time on instances is very fast in all cases.
We can also study the common types of failures that occur with recognition, in
particular for shapes as they are the main source of misclassification. Data mining
these misclassified strokes could identify features that may help correctly distinguish
them. Studying error cases may also lead to the identification of new features that
account for these misclassified shapes.
Feature selection strategies may also contribute to recognizer improvement. This
involves using feature selection algorithms to isolate the features that perform well.
When training an algorithm insignificant features can have a negative effect on the
success of classification algorithms [22] therefore careful feature selection is a very
important step to developing recognition techniques. We were surprised that 100+
features were employed by our top two dividers and speculate that some features are
redundant or detrimental. Redundant features will only slow execution time whereas
our concerns are with features that have a negative effect. Further exploration of
feature selection strategies could identify features that should be excluded.
Combining different classifiers into a voting system is also worthy of investigation.
Classifiers predictions can be weighted according to their performance and combined
to produce one overall classification for an instance [22]. We are yet to investigate
whether the different algorithms have a large number of common failures. If they all
fail on the same cases then voting is not useful. For future work we plan to investigate
the main cause of failures that occur for the original seven algorithms and identify
what proportions are common between them.
We chose to train and test on diagrams of different domains to create a general
diagram divider. Each diagram domain has its own syntax, semantics and mix of
drawing shapes. Given the difference between the training 10-fold validation values
 Building Digital Ink Recognizers using Data Mining: Distinguishing Between Text and
Shapes in Hand Drawn Diagrams 9
and the test results (~ 2.3%), it may be worthwhile to data mine and train a divider for
each diagram domain.

6 Conclusion
We have built seven new dividers using data mining techniques to distinguish
between text and shapes in hand drawn diagrams. The two best dividers, LADTree
and LogitBoost, are able to correctly classify 95.2% and 95.0% respectively of a test
set that they have received no training for. A comparative evaluation of these dividers
against two others shows that the new dividers clearly outperform the others. The
success of our new dividers demonstrates the effectiveness of using data mining
techniques for sketch recognition development.

7 Acknowledgements
Thanks to Associate Professor Eibe Frank for expert advice on using data mining
techniques. This research is partly funded by Microsoft Research Asia and Royal
Society of New Zealand, Marsden Fund.

8 References
1. Rubine, D.H. Specifying gestures by example. in Proceedings of Siggraph '91.
1991: ACM.
2. Paulson, B. and T. Hammond. PaleoSketch: Accurate Primitive Sketch Recognition
and Beautification. in Intelligent User Interfaces (IUI '08). 2008. New York,
USA: ACM Press.
3. Wobbrock, J.O., A.D. Wilson, and Y. Li, Gestures without libraries, toolkits or
training: a $1 recognizer for user interface prototypes, in User interface software
and technology. 2007, ACM: Newport, Rhode Island, USA.
4. Plimmer, B., Using Shared Displays to Support Group Designs; A Study of the Use
of Informal User Interface Designs when Learning to Program, in Computer
Science. 2004, University of Waikato.
5. Young, M., InkKit: The Back End of the Generic Design Transformation Tool, in
Computer Science. 2005, University of Auckland: Auckland.
6. Schmieder, P., B. Plimmer, and R. Blagojevic. Automatic Evaluation of Sketch
Recognition. in Sketch Based Interfaces and Modelling. 2009. New Orleans, USA.
7. Bhat, A. and T. Hammond. Using Entropy to Distinguish Shape Versus Text in
Hand-Drawn Diagrams. in International Joint Conference on Artificial
Intelligence (IJCAI '09). 2009. Pasadena, California, USA.
8. Bishop, C.M., M. Svensen, and G.E. Hinton, Distinguishing Text from Graphics in
On-Line Handwritten Ink, in Proceedings of the Ninth International Workshop on
Frontiers in Handwriting Recognition. 2004, IEEE Computer Society.
9. Patel, R., B. Plimmer, et al. Ink Features for Diagram Recognition. in 4th
Eurographics Workshop on Sketch-Based Interfaces and Modeling 2007.
Riverside, California: Eurographics.
 10 Rachel Blagojevic1, 1, Beryl Plimmer1, 3, John Grundy2, 2, Yong Wang1, 4

10. Plimmer, B. and I. Freeman. A Toolkit Approach to Sketched Diagram

Recognition. in HCI. 2007. Lancaster, UK: eWiC.
11. Lank, E., J.S. Thorley, and S.J.-S. Chen, An interactive system for recognizing
hand drawn UML diagrams, in Proceedings of the Centre for Advanced Studies
on Collaborative research. 2000, IBM Press: Mississauga, Ontario, Canada.
12. Hammond, T. and R. Davis. Tahuti: A Geometrical Sketch Recognition System for
UML Class Diagrams. in 2002 AAAI Spring Symposium on Sketch
Understanding. 2002.
13. Zeleznik, R.C., A. Bragdon, et al., Lineogrammer: creating diagrams by drawing,
in Proceedings of User interface software and technology. 2008, ACM:
Monterey, CA, USA.
14. Shilman, M. and P. Viola. Spatial recognition and grouping of text and graphics.
in EUROGRAPHICS Workshop on Sketch-Based Interfaces and Modeling. 2004.
15. Shilman, M., Z. Wei;, et al. Discerning structure from freeform handwritten notes.
in Document Analysis and Recognition. 2003.
16. Jain, A.K., A.M. Namboodiri, and J. Subrahmonia, Structure in On-line
Documents, in Proceedings of the Sixth International Conference on Document
Analysis and Recognition. 2001, IEEE Computer Society.
17. Ao, X., J. Li, et al., Structuralizing digital ink for efficient selection, in
Proceedings of the 11th international conference on Intelligent user interfaces.
2006, ACM: Sydney, Australia.
18. Machii, K., H. Fukushima, and M. Nakagawa. On-line text/drawings
segmentation of handwritten patterns. in Document Analysis and Recognition.
1993. Tsukuba Science City, Japan.
19. Microsoft Corporation, Ink Analysis Overview. 2008 cited 2008; Available
from: http://msdn.microsoft.com/en-us/library/ms704040(VS.85).aspx.
20. Mochida, K. and M. Nakagawa. Separating drawings, formula and text from free
handwriting. in International Graphonomics Society (IGS2003). 2003.
Scottsdale, Arizona.
21. Blagojevic, R., B. Plimmer, et al. A Data Collection Tool for Sketched Diagrams.
in Sketch Based Interfaces and Modeling. 2008. Annecy, France: Eurographics.
22. Witten, I.H. and E. Frank, Data Mining: Practical machine learning tools and
techniques. 2nd Edition ed. 2005, San Francisco: Morgan Kaufmann.
23. Breiman, L., Bagging predictors. Machine Learning, 1996. 24(2): p. 123-140.
24. Holmes, G., B. Pfahringer, et al., Multiclass alternating decision trees. ECML,
2001: p. 161-172.
25. Landwehr, N., M. Hall, and E. Frank, Logistic Model Trees. Machine Learning,
2005. 95(1-2): p. 161-205.
26. Friedman, J., T. Hastie, and R. Tibshirani, Additive Logistic Regression: a
Statistical View of Boosting. 1998, Stanford University.
27. Breiman, L., Random Forests. Machine Learning, 2001. 45(1): p. 5-32.
28. Platt, J. Machines using Sequential Minimal Optimization. in Advances in Kernel
Methods - Support Vector Learning. 1998.

View publication stats