Academia.eduAcademia.edu

Figure 6 – uploaded by bui hai phong

See full PDFdownloadDownload figure
figure 8 Flowchart of the classification of variable and word images (see online version for colours)

Figure 8 Flowchart of the classification of variable and word images (see online version for colours)

Related Figures (14)

Figure 4 (a) Examples of a text line (b) VPP of the text line (see online version for colours)
Figure 4 (a) Examples of a text line (b) VPP of the text line (see online version for colours)
In order to determine an extracted word from a text line is a variable or a textual word, a binary classification method is proposed. A key step in the classification is to extract the dominant features of observed words. The important feature that is used to discriminate variables from textual words is the italic font style of images. In scientific documents, variable is typically represented in italic font.
In order to determine an extracted word from a text line is a variable or a textual word, a binary classification method is proposed. A key step in the classification is to extract the dominant features of observed words. The important feature that is used to discriminate variables from textual words is the italic font style of images. In scientific documents, variable is typically represented in italic font.
Table 1 Features of VPP of variable and word images in Figure 5 ‘able 2 Comparison of VPP features between italic and non-italic styles of character ‘a’ of Arial font
Table 1 Features of VPP of variable and word images in Figure 5 ‘able 2 Comparison of VPP features between italic and non-italic styles of character ‘a’ of Arial font
The feature extraction of VPP and HPP is proposed by the observation that the peaks and valleys of projection profile of characters approximately obey the Gaussian (normal) distribution (Ghahramani, 2000; Roger, 2008). For a peak P; in vector P, the probability density function of the corresponding Gaussian distribution is described as follows:
The feature extraction of VPP and HPP is proposed by the observation that the peaks and valleys of projection profile of characters approximately obey the Gaussian (normal) distribution (Ghahramani, 2000; Roger, 2008). For a peak P; in vector P, the probability density function of the corresponding Gaussian distribution is described as follows:
Notes: Image of character ‘a’ in italic (a) and non-italic (c) styles in Arial font. Peaks and valleys of VPP of the character in italic (b) and non-italic (d) styles.
Notes: Image of character ‘a’ in italic (a) and non-italic (c) styles in Arial font. Peaks and valleys of VPP of the character in italic (b) and non-italic (d) styles.
Notes: Bounding boxes of characters of variable a; (a), peaks and valleys of VPP (b) and HPI (c) of the variable. classification problem. Especially, the method is very powerful for the cases that the decision boundary is irregular. There are several functions for the metric of the distance such as Euclidean, Minkowski. In the work, the Euclidean distance is used to identify the nearest samples in the training dataset. Besides, the value of k affects so much on the accuracy. In the work, different values of & (these values are 1, 3, 6, 9, 12, 15, 18, 21) are tested to find the optimal parameter.
Notes: Bounding boxes of characters of variable a; (a), peaks and valleys of VPP (b) and HPI (c) of the variable. classification problem. Especially, the method is very powerful for the cases that the decision boundary is irregular. There are several functions for the metric of the distance such as Euclidean, Minkowski. In the work, the Euclidean distance is used to identify the nearest samples in the training dataset. Besides, the value of k affects so much on the accuracy. In the work, different values of & (these values are 1, 3, 6, 9, 12, 15, 18, 21) are tested to find the optimal parameter.
Algorithm 1 Algorithm of detecting variable from document images
Algorithm 1 Algorithm of detecting variable from document images
For SVM, the linear and RBF kernel are examined. The obtained precision is 81.43% and 86.14% respectively. It indicates that the RBF outperforms linear kernel.
For SVM, the linear and RBF kernel are examined. The obtained precision is 81.43% and 86.14% respectively. It indicates that the RBF outperforms linear kernel.
Table 3. The classification results by using different machine learning algorithms
Table 3. The classification results by using different machine learning algorithms
Notes: Ital value indicates the highest scores of the proposed method. Table 4 Performance comparison of proposed and existing methods
Notes: Ital value indicates the highest scores of the proposed method. Table 4 Performance comparison of proposed and existing methods
Notes: Ital value indicates the highest performance of the proposed method. Table 5 Execution time of methods in the classification of variable and text in testing dataset The proposed method achieves the highest performance in time efficiency for the discrimination of variables from word images. The computation, feature extraction and classification of profile projection of image are faster than those of methods using orientation of gradient in Sun and Si (1997) and DWT in Zhang et al. (2004). The better performance of the proposed method is obtained because only discriminant values of VPP and HPP of images are used in the classification.
Notes: Ital value indicates the highest performance of the proposed method. Table 5 Execution time of methods in the classification of variable and text in testing dataset The proposed method achieves the highest performance in time efficiency for the discrimination of variables from word images. The computation, feature extraction and classification of profile projection of image are faster than those of methods using orientation of gradient in Sun and Si (1997) and DWT in Zhang et al. (2004). The better performance of the proposed method is obtained because only discriminant values of VPP and HPP of images are used in the classification.
Related topics:
EngineeringComputer ScienceArtificial Intelligence
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved