應用手寫辨識於UML與樂譜作曲之系統

行政院國家科學委員會補助專題研究

行政院國家科學委員會補助專題研究

行政院國家科學委員會補助專題研究

行政院國家科學委員會補助專題研究計畫

計畫

計畫

計畫

□

□

□

□ 成 果 報 告

成 果 報 告

成 果 報 告

成 果 報 告

■

■

■

■期中進度報告

期中進度報告

期中進度報告

期中進度報告

應用手寫辨識於

應用手寫辨識於

應用手寫辨識於

應用手寫辨識於 UML 與樂譜作曲之系統

與樂譜作曲之系統

與樂譜作曲之系統

與樂譜作曲之系統

計畫類別：■ 個別型計畫 □ 整合型計畫

計畫編號

：NSC 98－2221－E－009－117－MY2

執行期間： 98 年 08 月 01 日 至 100 年 07 月 31 日

計畫主持人：陳玲慧

共同主持人：無

計畫參與人員： 博士班研究生-兼任助理人員：李惠龍

博士班研究生-兼任助理人員：陳俊旻

博士班研究生-兼任助理人員：陳盈如

博士班研究生-兼任助理人員：楊文超

成果報告類型(依經費核定清單規定繳交)：■精簡報告 □完整報告

本成果報告包括以下應繳交之附件：

□赴國外出差或研習心得報告一份

□赴大陸地區出差或研習心得報告一份

□出席國際學術會議心得報告及發表之論文各一份

□國際合作研究計畫國外研究報告書一份

處理方式

：除產學合作研究計畫、提升產業技術及人才培育研究計畫、列管計

畫及下列情形者外，得立即公開查詢

□涉及專利或其他智慧財產權，□一年■二年後可公開查詢

執行單位：國立交通大學資訊工程學系(所)

中 華 民 國 99 年 5 月 31 日

中英文摘要 一、中文摘要 在計畫中，我們建構了一個統一建模語言線上手寫辨識系統。根據我們的觀察，統一建模語言 的圖形多半為類似方形或是菱形的圖形，因此在本系統中利用決策樹的方式，來達到辨識的效果。 首先我們擷取使用者輸入圖形的幾何特徵，來進行第一階段的分類。接著從輸入圖形擷取我們需要 的特徵，和其所屬分類中各個圖形的特徵向量進行比對，即可得到最後的辨識結果。本系統之優點 在於可以接受使用者任意筆順的輸入，並且辨識的方法較之前更為簡單有效，正確結果出現在前三 名的辨識率為 91.24%。 關鍵字：統一建模語言、線上、手寫辨識 二、Abstract

We construct an online handwritten recognition system of UML diagrams. We use a decision tree to do recognition. According to our observation, the shapes of the notations of UML diagrams almost look like rectangles or diamonds. Based on this characteristic, an input notation is first classified to the correct category. Then some notation features are extracted from the input notation and used to do final recognition. The advantages of our system are that we can accept free style input and our method is simpler and more efficient than previous methods. The recognition rate of the top three choices is 91.24%.

1. Introduction 1.1 Motivation

In recent years, the development of the handheld devices and pen-based computing hardware, such as PDAs, electronic whiteboards and tablet computers, is grown rapidly, and the handwritten systems which can work in the freehand drawing environment are short of demand. There exists some handwritten recognition systems in some different applications, including math formula [1], engineering drawings [2], table detection [3] and geometric shapes [4-5]. However, the Unified Modeling Language (UML) are widely used in many different domains but there is no handwritten recognition system supporting them.

UML diagrams are widely used in the field of software engineering. Early in the software design cycle, software engineers need to sketch UML diagrams to represent the whole structure of the system. Engineers may draw these diagrams on paper, whiteboard or computer. There are many Computer Assisted Software Engineering (CASE) tools like Rational Rose or Visio to sketch UML diagrams on computer. The functionality of these CASE tools is robust but they have some drawbacks. The most serious drawback of CASE tools is that their design concepts are technique oriented. Technique oriented design provides strong capability but it is not convenient to use. Due to these reasons, we want to build a handwritten recognition system which can allow people enjoying the freedom of drawing UML diagrams by hand.

1.2 Previous Works

In 2000, Damm et al. [8] proposed the Knight Project which is a gesture based system for entering and editing UML diagrams. Gestures are some simplified shapes designed by the designer to replace the complex notations. Due to that all of the shapes are simplified, the advantage of gesture based systems is easy to recognize the input notations. However, the user needs to learn what the gestures stand for because they are designed by the designer. In Knight Project, the gestures are separated into two classes, compound gestures and eager gestures, and they use Rubine’s algorithm [9] to recognize their gestures. The drawbacks of the Knight Project are that the gesture based system is not intuitional enough. Besides, they do not illustrate the notations supported by their system and there is no experimental result to show their recognition rate.

In 2001, Lank et al. [10] proposed an online recognition algorithm for UML diagrams. The algorithm is composed of the domain dependent kernel and the domain independent kernel. The domain independent kernel deals with the preprocessing steps, including capturing the input strokes, stroke grouping and so on, and the domain dependent kernel is the part of recognition. In the recognition algorithm, they use size, number of strokes, the input order of strokes and the stroke’s bounding box size to recognize the input notations. Their algorithm does not allow user drawing the notations in various order. Besides, there is no experimental result to show their recognition rate.

In 2003, Chen et al. [11] proposed another gesture based recognition system for UML diagrams called SUMLOW. The recognition kernel of SUMLOW combines several multi-stroke shape recognition algorithms to recognize their gestures. The characteristic of SUMLOW is that they allow user modifying, copying, replacing, and deleting input notations via pen-based input technique. Their system has high recognition rate, but there are only six experienced UML designers to participate in their experiment.

In 2006, Costagiola et al. [12] proposed an online recognition method for hand-drawn diagrams based on grammar formalism, namely Sketch Grammars. The method uses a parse tree and the Sketch Grammar to recognize input notations. To enhance the recognition rate, the authors propose a language recognizer which can help the original recognizer to select the best interpretation. This method can be adapted to any notation besides UML diagrams and has high recognition rate. However, a troublesome problem for this method is how the grammars train for new notations.

2. UML NOTATION DATABASE

UML diagrams have thirteen different types and more than forty different notations. However, some of these notations are used rarely and their shapes are more complex. In this project, we choose 23 notations based on UML concepts and the frequency of usage to recognize. These 23 notations are shown in Figure 2.1.

In the project, we invite 20 persons to draw the 23 notations ten times for each and collect the ink data they draw. We randomly choose half of the ink data for training, and the rest for testing.

Figure 2.1 Supported notations of the system. 3. PROPOSED METHOD

The proposed method is based on a decision tree and shown in Figure 3.1. The whole process consists of four major phases：geometric feature extraction, category classifier, notation feature extraction (NFE), and final classifier. In the geometric feature extraction phase, some geometric features, such as convex hull, bounding rectangle, PA ratio and Area ratio, are extracted from the input notation. In the category classifier phase, the features extracted in the previous phase are used to classify the input notation to the belonging category. In the notation feature extraction phase, the input notation is divided into primitives and then we extract features like direction, location and distance from these primitives. In the final classifier phase, based on the extracted features, a similarity measure is provided. Based on the similarity measure, the result notation that is most similar to the input notation is determined.

Figure 3.1 The proposed method. 3.1 Geometric Feature Extraction

According to our observation, the notations supported in the system can be divided into five categories, i.e. circle, line, rectangle, diamond, and others, based on their geometric properties. This phase extracts geometric features from input notation to classify it to the correct category. The geometric features we used include convex hull, bounding rectangle, PA ratio and Area ratio. Each of these features is described below.

Geometric Feature Category Classifier

Result

Circle

Line Rectangle Diamond Other

s Final Classifier Result Final Classifier Result Final Classifier Result Final Classifier Result Final Classifier NFE

NFE NFE NFE NFE

Pen Input

3.1.1 Convex Hull

The convex hull for a set of points X is the minimal convex set containing X. Figure 3 gives two examples to illustrate convex hull. We use the Graham scan algorithm [14] to find the convex hull of the input notation. Figure 3.2 (b) shows the convex hull of an input notation “Actor”. The blue line denotes the convex hull. After finding the convex hull, we compute its perimeter and the area. These values will be used in the following section.

Figure 3.2 Two examples to illustrate convex hull (a) A convex hull of a set of points. (b) The convex hull of an input notation “Actor”.

3.1.2 Bounding Rectangle

The bounding rectangle is the minimum rectangle containing the input notation. We scan all points of input notation to find the minimum values of x and y coordinates, and the maximum values of x and y coordinates. After finding these coordinates, we use them to establish the bounding rectangle of the input notation. Figure 3.3 shows an example of the bounding rectangle of an input notation “Actor”. The bounding rectangle is shown by red lines. After finding the bounding rectangle, we compute its perimeter and area. These values will be used in the following section.

Figure 3.3 An example of the bounding rectangle of an input notation “Actor”. 3.1.3 PA Ratio

PA ratio proposed by Kimura [6] is defined as：

CH 2 CH/Area Perimeter ratio = PA _{, (1)}

where PerimeterCH denotes the perimeter of the convex hull of the input notation, and AreaCH denotes the area of the convex hull of the input notation. Note that the perimeter and area partly define the shape of an object. This ratio will be a constant for some kinds of shape. For instance, PA ratio = 16 for any square rectangle and PA ratio = 4π for any circle. Size independent is the main advantage of PA ratio. In the project, PA ratio is used to classify circle and line.

3.1.4 Area Ratio

Area ratio is also proposed by Kimura [6]. The ratio is defined as：

BR CH/Area

Area ratio=

Area _{, (2)}

where AreaBR is the area of the bounding rectangle of an input notation.

Area ratio also has the property of size independent. In the project, we use this ratio to distinguish the rectangle and the diamond shape.

3.2 Category Classifier

After extracting geometric features, we use these features to classify the input notation to the correct category. The 23 supported notations are separated to five categories including circle, line, rectangle, diamond, and others. The classification of each notation is shown in Figure 3.4. Four different filters, namely circle filter, line filter, rectangle filter and diamond filter, are provided to distinguish the five categories in the category classifier. The flowchart of the category classifier is shown in Figure 3.5.

Figure 3.4 The classification of each notation. 3.2.1 Circle Filter

In the category classifier, we use the circle filter to check for circles first. In the project, we use PA ratio for circle filter. PA ratio of a perfect circle of any size is a scalar 4π. Due to that the input may not be a perfect circle, we need to train a threshold range around 4π to classify the input notation. To train the threshold, we compute the PA ratio of the notations belonging to the circle category in the training database first. Then we find a maximum and a minimum as the upper bound and the lower bound of threshold range.

Figure 3.5 The flowchart of the category classifier. 3.2.2 Line Filter

If the input notation does not belong to the circle category, it will be checked by the line filter. Here, we use PA ratio for line filter. Due to the PerimeterCH of a line is close to twice of the length of input notation and the AreaCH of a line is closed to the product of the length of input notation and ∆h which is the maximum distance between input stroke and its convex hull, the PA ratio of a line can be approximated by h l h l l ratio PA ∆ = ∆ ×

≈ (2 )2 4 _{. Since ∆h << l the PA ratio should be large. Here, we take} 120 as a threshold value obtained by training. Figure 3.6 shows two examples to explain why the PA ratio is greater than a threshold. In Figure 3.6, the black line is user’s input and the red line is the convex hull. To avoid the error of dividing zero, we set the PA ratio equal to 200 when the area of the convex hull of a line is equal to zero.

Figure 3.6 Two examples to show the PA ratios of lines. Circle Filter Line Filter Rectangle Filter Diamond Filter Rectangle Category Line Category Circle Category Yes Yes Yes Diamond Category Yes No No No No Is Diamond? Is Rectangle? Is Line? Is Circle? Others Category Input

3.2.3 Rectangle Filter

Rectangle filter will be used when the notation does not belong to the circle or line category. In the project, we use Area ratio for rectangle filter. According to the fact that the AreaCH of a rectangle is almost equal to the AreaBR of the rectangle, the Area ratio of a rectangle is close to 1. Figure 3.7 shows two examples to explain the fact mentioned above. In Figure 3.7, the black line is user’s input, the red line is the convex hull and the green line is the bounding rectangle. To get a threshold range, we also train the rectangle notations in the training database.

Figure 3.7 Two examples to show the Area ratios of rectangles. 3.2.4 Diamond Filter

If input notation is not considered as a circle, a line or a rectangle, it will be checked by the diamond filter. In the project, we use Area ratio for diamond filter. We assume that the notations belonging to the diamond category are all upright patterns. The AreaBR of a diamond is nearly two times of the AreaCH of a diamond based on our assumption. In other words, the Area ratio of a diamond is nearly 0.5. Figure 3.8 shows two examples to explain why the Area ratio of a diamond is nearly 0.5. In Figure 3.8, the black line is user’s input, the blue line is the convex hull and the red line is the bounding rectangle. We use a threshold range which is trained using the diamond notations in the training database to check whether the input notation belongs to the diamond category or not.

Figure 3.8 Two examples to show the Area ratios of diamonds. 3.2.5 Other Notations

If the input notation does not belong to any category mentioned above, it will be classified to the others category. In our experiments, after category classification the others category contains Actor and several rectangle notations which are ill-written.

3.3 Notation Feature Extraction

After the input notation is classified to a category, some notation features will be extracted for the final classification. Before extracting notation features, we will first segment the notation into several primitives, which will be described in the following subsection. The notation features extracted include the number of primitives, the direction of each primitive, the location of each primitive, the length of each primitive, and the hollowness of the notation. In the following subsections, we will describe how to extract features.

3.3.1 Primitive

A primitive is defined to be the minimum unit of a notation, which may be a line or an arc. The advantage of segmenting a notation to primitives is that it is much easier for the shape matching procedure to find the matching notation. All the notation features are extracted in primitive level except hollowness.

To divide a notation to many primitives, we use 4-way chain code and the curvature of each point. The 4-way chain code is shown in Figure 3.9. First we compute the chain code for each point. Then we compute the curvature of each point by

) ) 1 ( ) 1 ( ( 2 2 _{+ − − +} + − − + − − = x i x i Crpi cos-1 , (3)

where x(i), y(i) denotes the x, y coordinates of point pi and Crpi is the curvature of point pi. After computing the curvature, we evaluate the curvature difference between two neighboring points to find the dominate points, which have curvature difference greater than a threshold. Finally, the notation is divided into several segments using the dominate points as cut points, each segment is considered as a primitive of the notation. When the notation is segmented to many primitives, we take the number of primitives, N, as the first feature. Note that we have two kinds of primitives: line and curve, which are decided by the sequence of chain codes of the primitive. To decide what kind of a primitive is, we evaluate the chain code difference between each two neighboring points in the chain code sequence and sum all of them. If the summation is larger than a threshold, we will decide that it is a curve; otherwise, it is a line.

Figure 3.9 4-way chain codes 3.3.2 Direction and Location Feature

The direction of a line primitive is defined as the chain code which appears most frequently in the primitive. If the primitive is an arc or a curve, we set 5 to be its direction. In order to record the directions of the extracted primitives as a feature vector, we should give an unique id to each primitive. The primitives get their unique ids based on the relative locations on the notation. Since some notations have some rotation varieties with 90, 180, and 270 degrees, we provide an algorithm to find relative location.

First, we extract the directions of primitives. Then the primitives with the same direction are collected and sorted according to their top left corner points. Finally, each primitive gets its unique id based on the sorted list. When all the primitives get their unique ids, we combine their directions into a direction feature vector. An example is shown in Figure 3.10; the blue number in the figure denotes the id of a primitive. The provided algorithm is stated below.

Figure 3.10 An example of the relative location of each primitive in a notation with the direction feature vector is (1, 1, 1, 1, 3, 3).

When the algorithm is finished, all the primitives have unique ids and we group the directions of primitives according to their ids into a vector, DIR, which is considered as the second notation feature. The notation in Figure 3.10 has (1, 1, 1, 1, 3, 3) as its direction feature.

3.3.3 Length Feature

The length feature is a binary value which represents that a primitive is long or short. To extract this feature, we first find the longest primitive in a notation. Then each primitive is compared to the longest one. If the length of the primitive is larger than half of the longest one, it is considered as a long primitive;

Algorithm to Find Unique Id 1. Setting variable i to 1.

2. Collecting the primitives with direction i to a temp list.

3. Sorting the temp list according to the top left corners point of primitives.

4. Giving a unique id to each primitive in the sorted temp list according to its order in the list. 5. Increasing 1 to i. If i is less than 6, go to step 2; otherwise, stop.

1 2 3 4 4 5 2 3 6 1

otherwise, it is a short one. The length feature is calculated by     < = otherwise, 2 len(j) max 2 1 len(i) 1 ) (i if j LEN (4)

where len(i) denotes the length of the ith primitive, and max len(j) denotes the length of the longest primitive in the notation.

3.3.4 Hollowness

The hollowness feature is the only feature extracted in the notation level. Hollowness means whether the shape is a solid one or not. A hollow shape has a property that there are no points near the gravity center of the shape. According to this property, we locate a rectangle with size 60% of the convex hull, and the center of the located rectangle is the same as that of the convex hull. If the number of points inside the rectangle is smaller than a threshold, the notation is considered as a hollow shape. Otherwise, the notation is not a hollow shape. Figure 3.11 (a) is a hollow shape, and Figure 3.11 (b) is a solid shape. The hollowness feature, H, is also a binary value and defined by

   < = otherwise, 2 t P If 1 rec H (5) where Prec denotes the number of points inside the located rectangle, and t is a threshold value.

Figure 3.11 Examples of hollowness. (a) A hollow shape. (b) A solid shape 3.4 Final Classifier

Feature vectors extracted from the operations described above, including N, DIR, LEN, and H, are taken for pattern matching at this phase. We use the inverse of sum-of-absolute-difference (SAD) as the similarity measure to obtain the most likely notation for the input notation. Let notations T and T’ be the database notation and the input notation respectively, the similarity between T and T’ is calculated by

∑

= − = 4 1 ' ) ( i i i i K F F T SAD , ( ) 1 ) ( T SAD T S = , (6)

where Fi (Fi’) denotes the ith feature vector of T (T’), and Fi∈{N, DIR, LEN, H}. Ki denotes the number of elements in the feature vector Fi. Due to the dimension of direction feature vector and the length feature vector are dependent on the number of primitives, we will pad zero to the smaller vector between Fi and Fi’ for computing SAD. Let max S(T)

T

arg *=

T , the input notation is considered to be notation T*.

3.5 MBSAS Algorithm for Database Creation

The final classifier step uses the inverse SAD to classify the notation. If we calculate SAD between the input notation and all the notations in the database which is described in Chapter 2, the processing time will be very long. Therefore, we use MBSAS to reduce the database and get some representative feature vectors for reducing the processing time. Modified Basic Sequential Algorithm Scheme (MBSAS) [13] is a clustering algorithm. More specifically, it is an algorithm to group the objects based on attributes. MBSAS does not need to know the number of clusters. It contains two phases. The first phase determines the number of clusters; the second phase is the pattern classification.

4. EXPERIMENTAL RESULTS

In order to evaluate the recognition rate of the proposed method, we invite 20 persons, with poor experience using tablet digitizer and tablet PC, to sketch 23 supported notations about ten times for each notation. We use a tablet digitizer, Wacom Graphire4 CTE-440, and a tablet PC, HP Compaq tc4200, to collect the ink data. In the experiment, we randomly choose half of the ink data for training and the rest

(a) (b)

gravity center

for testing. Table 1 shows the recognition rate of the proposed method. The first column shows that only the top one is chosen and the recognition rate is 84.62%. The second column shows that the top three ones are taken, and the recognition rate increases from 84.62% to 91.24%. We can observe that the notations belonging to the Line, Circle, and Diamond categories are classified very well.

Table 1. The recognition rate of top 1 choice and top 3 choices.

Shape Top 1 Accuracy% Top 3 Accuracy%

Activity 73(73/100) 86(86/100) Aggregation 88.78(87/98) 91.84(90/98) Activationbar 87.78(79/90) 88.89(80/90) Actor 87.78(79/90) 92.22(83/90) Branch 90.91(90/99) 100(99/99) Class 84.44(76/90) 92.22(83/90) Component 73.81(62/84) 86.9(73/84) Communication 100(98/98) 100(98/98) Dependency 81(81/100) 86(86/100) End 92(92/100) 92(92/100) Fork 89.29(75/84) 89.29(75/84) Generalize 97.96(96/98) 98.98(97/98) Initial 77(77/100) 81(81/100) Interface 78.65(70/89) 85.39(76/89) Lifeline 100(89/89) 100(89/89) Node 72.22(65/90) 86.67(78/90) Note 70.79(63/89) 91.01(81/89) Object 87.78(79/90) 88.89(80/90) Package 75.56(68/90) 92.22(83/90) State 71(71/100) 84(84/100) Swimlane 80.9(72/89) 91.01(81/89) Transition 94(94/100) 97(97/100) Use Case 89.89(80/89) 96.63(86/89) Total 84.62 (1816/2146) 91.24 (1958/2146)

Table 2. Comparison with SkGs method Shape SkGs without Language

Recognizer (%)

SkGs with Language

Recognizer (%) Proposed Method (%)

Actor 76.92(10/13) 92.31(12/13) 92.31(12/13) Use Case 83.3(45/54) 90.74(49/54) 96.30(52/54) Communication 100(21/21) 100(21/21) 100(21/21) Dependency 72.73(16/22) 72.73(16/22) 95.45(21/22) Generalize 81.82(9/11) 100 (11/11) 100(11/11) Transition 88.89(8/9) 88.89(8/9) 100(9/9) Total 80.99(98/121) 91.74(111/121) 96.69(117/121)

We compare our method to SkGs method [12] to show that our proposed method has higher recognition rate than others. In SkGs method, there are five students to participate the experiment, and each student draw 20-25 symbols of Use Case diagram. The recognition method proposed in [12] has two parts. The first part only used the Grammar based method to recognize symbol, and the second part combined the Grammar based method and the language recognizer. Our results will be compared to these two parts. In the comparison, we also invite five persons drawing the symbols supported in SkGs method, and the recognition rate is shown in Table 2.The results of the proposed method are better than these two parts besides Actor in Table 2. Thus, our recognition rate is superior to SkGs method.

5. Conclusion

tree. First, some geometric features are extracted for classifying the input notation to the corresponding category. Then we extract several notation features in primitive level and notation level to create the feature vectors. Finally, the similarity measure based on SAD is calculated for getting the final result.

In the system, users can sketch UML diagrams using tablet computer, digital tablet, and mouse. Users can sketch any notation in any kind of order in the system. After sketching a notation, the standard notation will replace the hand-drawn one and be displayed with the correct position and size. We also support user self-definition function which allows user defining gestures representing the UML notations. Besides these characteristics, the most important property of the system is that it is relative efficient and simple to other methods mentioned above because we use decision tree and reduction database to reduce the comparison time.

Although the system provides many functions of sketching UML diagrams, it is still not enough. In the future, we will add more functions, such as forward/backward engineering, modularity, supporting the multi-layer diagrams, and supporting more UML notations to make the system become a practical tool. 參考文獻

參考文獻 參考文獻 參考文獻

[1] L. H. Chen and Y. P. Yin, “A System for On-line Recognition of Handwritten Mathematical Expressions,” Computer Processing of Chinese and Oriental Languages, Vol. 6, No. 1, pp. 19-39, 1992.

[2] G. Hutton, M. Cripps, D. Elliman, and C. Higgins, “A Strategy for On-line Interpretation of Sketching Engineering Drawings,” Fourth Intl. Conf. on Document Analysis and Recognition, pp. 771-775, 1997.

[3] Z. Lin, J. He, Z. Zhong, R. Wang, and H. Shum, “Table Detection in Online Ink Notes,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 28, No. 8, pp. 1341-1346, 2006. [4] T. Kimura, A. Apte, and V. Vo, “Recognizing Multistroke Geometric Shapes： an Experimental

Evaluation,” Proceedings of the ACM conference on User Interface and Software Technology (UIST’93), pp. 121-128, 1993.

[5] M. Fonseca and J. Jorge, “Using Fuzzy Logic to Recognize Geometric Shapes Interactively,” The Ninth IEEE International Conference on Fuzzy Systems, Vol. 1, pp. 291-296, 2000.

[6] UML in Wikipedia : http://en.wikipedia.org/wiki/Unified_Modeling_Language [7] Object Management Group : http://www.uml.org/

[8] C. Damm, K. Hansen, M. Thomsen, and M. Tyrsted, “Creative Object-Oriented Modeling： Support for Intuition, Flexibility and Collaboration in CASE Tools,” ECOOP 2000-Object-Oriented Programming： 14th European Conference, Vol. 1850, pp. 27-43, 2000.

[9] D. Rubine, “Specifying Gestures by Example,” Proceedings of SIGGRAPH'91, pp. 329-337, 1991. [10]E. Lank, J. Thorley, S. Chen, and D. Blostein, “On-line Recognition of UML Diagrams,” The Sixth

IEEE International Conference on Document Analysis and Recognition, pp. 356-360, 2001.

[11]Q. Chen, J. Grundy, and J. Hosking, “An E-whiteboard Application to Support Early Design-Stage Sketching of UML Diagrams,” IEEE Symposium on Human Centric Computing Language and Environments, pp. 219-226, 2003.

[12]G. Costagliola, V. Deufemia, and M. Risi, “A Multi-layer Parsing Strategy for On-line Recognition of Hand-drawn Diagrams,” IEEE Symposium on Visual Languages and Human-Centric Computing (VL-HCC’06), pp. 103-110, 2006.

[13]S. Theodoridis and K. Koutroumbas, Pattern Recognition, Academic Press, 2006.

計畫成果自評 計畫成果自評 計畫成果自評 計畫成果自評 本計畫的執行進度符合當初所提之計畫內容，亦完成計畫書所擬定之研究目標。在這一年當 中，提出了一個應用手寫辨識於 UML 之系統。本系統擷取使用者輸入圖形的幾何特徵，來進行第 一階段的分類。接著從輸入圖形擷取需要的特徵，和其所屬分類中各個圖形的特徵向量進行比對， 即可得到最後的辨識結果。本系統之優點在於可以接受使用者任意筆順的輸入，並且辨識的方法較 之前更為簡單有效，正確結果出現在前三名的辨識率為 91.24%。 本技術以手寫辨識的方式，輸入 UML 的圖形，達到使用者介面更人性化的目的，此技術可應 用於數位教學及個人行動裝置上。

可供推廣之研發成果資料表

□ 可申請專利 □ 可技術移轉 日期：99 年 5 月 31 日

國科會補助計畫

國科會補助計畫

國科會補助計畫

國科會補助計畫

計畫名稱：應用手寫辨識於 UML 與樂譜作曲之系統 計畫主持人：陳玲慧 教授 計畫編號：NSC 98－2221－E－009－117－MY2 學門領域：影像處理

技術

技術

技術

技術/創作名稱

創作名稱

創作名稱

創作名稱

應用手寫辨識於 UML 之系統

發明人

發明人

發明人

發明人/創作人

創作人

創作人

創作人

陳玲慧 中文：在計畫中，我們建構了一個 UML 線上手寫辨識系統。根據 我們的觀察，UML 的圖形多半為類似方形或是菱形的圖形，因此 在本系統中利用決策樹的方式，來達到辨識的效果。首先我們擷取 使用者輸入圖形的幾何特徵，來進行第一階段的分類。接著從輸入 圖形擷取我們需要的特徵，和其所屬分類中各個圖形的特徵向量進 行比對，即可得到最後的辨識結果。本系統之優點在於可以接受使 用者任意筆順的輸入，並且辨識的方法較之前更為簡單有效，正確 結果出現在前三名的辨識率為 91.24%。

技術說明

技術說明

技術說明

技術說明

英文：We construct an online handwritten recognition system of UML

diagrams. We use a decision tree to do recognition. According to our observation, the shapes of the notations of UML diagrams almost look like rectangles or diamonds. Based on this characteristic, an input notation is first classified to the correct category. Then some notation features are extracted from the input notation and used to do final recognition. The advantages of our system are that we can accept free style input and our method is simpler and more efficient than previous methods. The recognition rate of the top three choices is 91.24%.

可利用之產業

可利用之產業

可利用之產業

可利用之產業

及

及

及

及

可開發之產品

可開發之產品

可開發之產品

可開發之產品

個人行動裝置、數位教學。

技術特點

技術特點

技術特點

技術特點

1. 可以接受使用者任意筆順的輸入，並且辨識的方法較之前更為 簡單有效。 2. 使用者能自訂手勢，表示 UML 的符號。

3. 我們利用 decision tree 和 reduction database 減少符號的比較時 間。

推廣及運用的價值

推廣及運用的價值

推廣及運用的價值

推廣及運用的價值

本技術以手寫辨識的方式，輸入 UML 的圖形，達到使用者介面更 人性化的目的，此技術可應用於數位教學及個人行動裝置上。 ※ 1.每項研發成果請填寫一式二份，一份隨成果報告送繳本會，一份送 貴單位研發成果 推廣單位（如技術移轉中心）。 ※ 2.2.本項研發成果若尚未申請專利2.2.本項研發成果若尚未申請專利本項研發成果若尚未申請專利，本項研發成果若尚未申請專利，請勿揭露可申請專利之主要內容，，請勿揭露可申請專利之主要內容請勿揭露可申請專利之主要內容請勿揭露可申請專利之主要內容。。。。

※

3.本表若不敷使用，請自行影印使用。