Investigation of Off-Line Japanese Signature Verification Using a ...
Investigation of Off-Line Japanese Signature Verification Using a Pattern Matching Katsuhiko Ueda Department of Information Engineering, Nara National College of Technology Yamatokoriyama, Nara 639-1080, Japan [email protected] Abstract This paper deals with an off-line Japanese signature verification using a pattern matching method. An annoying problem encountered in off-line signature verification using a pattern matching is that variation of signature stroke widths and a registered signature selected from a set of reference samples affect the verification performance. In this paper, we propose a modified pattern matching method, which is independent of stroke widths, and an appropriate selection method of a registered signature to solve this problem. These methods were experimentally compared with a conventional pattern matching method and a random selection method of a registered signature. Experimental results showed that the proposed methods improve the verification performance.
1. Introduction This paper deals with an off-line verification of Japanese signatures on bank checks. Although many algorithms for off-line signature verification have been proposed up to the present [1], most of them cannot be applied directly to Japanese signatures, because Japanese signatures have different features from them of western signatures. Japanese signatures almost always consist of two to six Kanji, Hiragana and/or Katakana characters, and their component characters are spaced each other. Some researches of off-line verification for Japanese signatures have been proposed [2], [3], [4], [5], although an off-line Japanese signature verification is difficult due to a lack of stability and individuality. Yoshimura et al. showed that a pattern matching method is able to achieve a good verification performance for Japanese signatures [3], [4]. However, the similarity between two signatures obtained by a pattern matching method is affected by their stroke widths. The stroke widths vary with the pen used for signing. Even if signatures are written with the same pen, the stroke width may also vary with a normalization of a character size for the following processing. To solve this problem, it is effective to verify signatures after normalizing their stroke widths. In addition, a registered
signature selected from a set of reference samples for each signer affects the verification performance. Therefore it is necessary to find an appropriate selection rule of a registered signature. In this paper we propose a new pattern matching method and a selection method of a registered signature for Japanese signature verification, based on the above consideration. In our modified pattern matching method, the strokes of signatures are first thinned at one pixel width, and then thinned signatures are blurred by a fixed point-spread function. Successively the similarity between registered and examined signatures is calculated. Concerning selection method of a registered signature, a signature is extracted from a set of reference samples, and the mean value and standard deviation of similarities between the extracted signature and the other samples are calculated. The signature with the maximal ratio of the mean value of the similarity to the standard deviation is selected as the registered signature. We will also show that the proposed methods improve the verification performance, through the verification experiments.
2. Outline of signature verification system using a pattern matching Figure 1 shows the flow diagram of the signature verification process proposed in this paper. The input signature is a binary image. It is first segmented into component characters. The position and size of component characters are then normalized. Successively, the similarity between registered and examined signatures is calculated in a pattern matching process. Finally, an examined signature is classified into one of two categories: genuine or forgery.
2.1. Input signatures We deal with Japanese signatures on bank checks. Therefore, it is necessary for verification to extract signatures from bank checks. The input signature of this verification process is binary image, which is extracted
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
3. Modified pattern matching An annoying problem encountered in off-line signature verification using pattern-matching method is that the similarity between registered and examined signatures is affected by their stroke widths. Therefore, it is effective for stable and reliable verification to verify based on the structural similarity of signature strokes. To realize this idea, the stroke of signature (Figure 3(b)) is thinned at one pixel width without any change of stroke structure. In this work, we use the Hildich’s algorithm for thinning. Figure 4(a) shows the example of thinning result.
Figure 1. Flow diagram of signature verification process proposed in this paper.
Figure 2. Example of input signature images.
Figure 3. Segmentation and normalization of size and position.
from bank check by a color clustering method [6]. The example of input signature is shown in Figure 2. The size and the resolution of input image are 210x700 pixels and 300dpi respectively.
2.2. Segmentation and normalization Japanese signatures almost always consist of two to six kanji, hiragana and/or katakana component characters, and they are spaced each other. The input signature is segmented into component characters based on this characteristic [7]. The segmentation is performed referring the widths and the number of component characters of the registered signature. The example of segmentation result is shown in Figure 3(a). The size and position of every component character are normalized to a fixed size and a fixed position respectively. Figure 3(b) shows the result of size and position normalization. The normalized image size is 240x240 pixels.
Figure 4. Normalization of signature stroke width.
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
For stable pattern matching, the thinned signature represented in the xy-plane is blurred using the following equation.
g ( x, y ) = ∫∫ f ( x − α , y − β )h(α , β )dα d β
(1)
where g(x,y), f(x,y) and h (α , β ) are the blurred image, the thinned image and the point-spread function respectively. α , β denote the shift amounts in the xyplane. Equation (1) is the well-known relationship called convolution of f and h. As the point-spread function h (α , β ) , we define the binary square function as shown in Figure 5. In this figure, the black square region denotes h=1, and zero otherwise. The size of square region is 19x19. The example of blurred image is shown in Figure 4(b). Successively the pattern matching is performed using blurred signature images. Let Xi (u), Xi(l) be the pattern vectors of i’th component characters of examined and registered signatures respectively. The elements of these pattern vectors are pixel values of the blurred images. The similarity Si between i’th corresponding component characters of examined and registered signatures is calculated using the following equation.
Si =
( X i(u ) , X i(l ) ) ( X i(u ) , X i(u ) ) ( X i(l ) , X i(l ) )
4. Selection of registered signature The selection of registered signatures may affect significantly verification performance, because examined signatures are compared with the corresponding registered signature in the reference database for verification. A signature is represented as a point in a certain feature space. The effect of signature selected as a registered one is shown in Figure 6 schematically. In this figure, circular points denote the genuine signatures, and triangular points are the forged signatures. The solid line circles are the boundaries of the genuine signatures. Figure 6(a) shows the case of optimal selection. Selecting the signature at the center of genuine cluster can maximize the similarity between the selected signature and other genuine signatures. In other case as shown in Figure 6(b), the verification error rate may increase. Figure 6. Effect of selecting a signature as a
(2)
i= 1, 2,……, m (m : the number of component characters) where (*,*) denotes a scalar product of vectors. The resultant similarity S between examined and registered signatures is obtained by taking an average of the similarities Si as follows,
S=
1 m ∑ Si m i =1
registered one.
Table 1. Example of similarities between genuine signatures.
(3)
where m denotes the number of component characters of the signature. An examined signature is verified based on the resultant similarity S. On the basis of above idea, we propose the following method to select the registered signature from a set of reference samples. A signature is extracted from a set of reference samples, and the mean value µ and standard deviation σ of similarities between the extracted signature and the other samples are calculated. The µ
Figure 5. The point-spread function h used in Equation (1).
signature with the maximal ratio σ of the mean value of the similarity to the standard deviation is selected as the registered signature. Table 1 shows an example of pattern
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
matching results for 5 genuine signatures. In this case, the signature No.2, which gives maximal µ σ , is selected as the registered signature.
5. Threshold definition The verification is based on the assumption that the similarities for an individual writer tend to cluster, while those of a population of writers are more widely scattered. Moreover, the cluster of forgeries cannot be estimated previously. Therefore, we determine the threshold value for verification based on only the statistical property of genuine clusters. Let {Sj (j=1, 2, ….. , n)} be the obtained similarities between a registered signature and its n genuine signatures for training. Moreover, let µ and σ be the mean value and the standard deviation of { Sj } respectively. The threshold value T of genuine signatures for each registered one is defined as the following equation.
T = µ − aσ
Figure 7. Example of signature images used in the experiments.
( 4)
where a is a certain value (called threshold coefficient). For unknown examined signature, if the similarity S(u) of an examined signature is larger than T, then the signature is classified as a genuine and a forgery otherwise.
6. Experiments 6.1. Experimental method In order to test the verification performance, we have collected 110 genuine signatures from each of 10 signers and 100 forged signatures for each genuine one. 10 reference samples are extracted from the each set of genuine signatures to select the registered signature and estimate mean value and standard deviation of similarities. Therefore a total of 100x10=1000 genuine signatures and 100x10 =1000 forged signatures were used for test. Although the small number of reference samples seems unusual, a large number of samples are generally not available in practical applications. An example of signature images used in the experiment is shown in Figure 7. In order to evaluate the verification performance of the proposed method, we have performed three kinds of experiments as follows. (1) Proposed Method: described in this paper. (2) Method-1: with the modified pattern matching described in Section 3. and signatures selected randomly as a registered one.
Figure 8. FRR and FAR for various values of threshold coefficient a (Type B).
(3) Method-2: with a conventional pattern matching and signatures selected optimally as a registered one described in Section 4. Each verification process was implemented with C on a 1.9 GHz Pentium-4 PC.
6.2. Experimental results Figure 8 shows the example of verification results by the proposed method. As known well, two types of errors, false-rejection error rate (FRR) and false-acceptance error rate (FAR), hold a trade-off relationship for change of the threshold coefficient “a” in equation (4). The performance of three methods was evaluated using the error rate at which FRR and FAR are equal (the error rate of cross-point of FRR and FAR curves in Figure 8), because the determination of the optimal threshold coefficient “a” for all authentic signatures is a difficult problem. The performance of the proposed method, together with the performance of the other two methods is compiled in Table.2. Although there exist some
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
differences in the error rates according to the signer, the average error rates yield 9.1% by the Proposed Method, 14.8% by the Method-1 and 19.2% by the Method-2. Comparing these results, we can say that the modified pattern matching method with stroke width normalization and the optimal selection of a registered signature are able to reduce the average error rate of about 10.1% and 5.7% respectively. The processing time of the proposed method was 15 to 20 seconds per signature. Table 2. Error rates in % at FRR=FAR.
8. References [1] F. Leclerc and R. Plamondon, “Automatic signature verification: The state of the art-1989-1993”, International Journal of Pattern Recognition and Artificial Intelligence, 8, 3, pp. 643-660, (1994). [2] I. Yoshimura and M. Yoshimura, “Off-line verification of Japanese signature after elimination of background patterns”, International Journal of Pattern Recognition and Artificial Intelligence, 8, 3, pp. 693-708, (1994). [3] I. Yoshimura, M. Yoshimura and T. Tsukamoto, “Investigation of an automatic verification system for Japanese countersignatures on traver’s cheques”, Proceedings of the 7th IGS Conference, pp.86-87, (1995).
7. Conclusion In this paper, we proposed the modified pattern matching method, which is independent of signature stroke width and the selection method of a registered signature for Japanese signature verification. Through the experiments, it was confirmed that both methods proposed in this paper are effective for improving the verification performance. We are now working towards improving the verification performance and reducing the processing time. The determination of optimal threshold value is the future problem. We will also investigate more precise evaluation of the verification performance with large database and an extension of this work to Western signatures in the future.
[4] M. Yoshimura and I. Yoshimura, “Investigation of a verification system for Japanese countersignatures on traver’s cheques”, Transactions of the IEICE, J80-D-II, 7, pp.17641773, (1998) (in Japanese). [5] S. Ando and M. Nakajima, “An active search of local individualities for an off-line signature verification”, Transactions of the IEICE, J84-D-II, 7, pp.1339-1350, (2001) (in Japanese). [6] K. Ueda, T. Mutoh and K. Matsuo, “Automatic verification system for seal imprints on Japanese bankchecks”, Proceedings of the 14th ICPR, pp.629-632, (1998). [7] M. Yoshimura and I. Yoshimura, “Segmentation of Japanese signatures into component characters”, Technical Report of IEICE, PRMU97-3, pp.17-24, (1997) (in Japanese).
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
1. Introduction This paper deals with an off-line verification of Japanese signatures on bank checks. Although many algorithms for off-line signature verification have been proposed up to the present [1], most of them cannot be applied directly to Japanese signatures, because Japanese signatures have different features from them of western signatures. Japanese signatures almost always consist of two to six Kanji, Hiragana and/or Katakana characters, and their component characters are spaced each other. Some researches of off-line verification for Japanese signatures have been proposed [2], [3], [4], [5], although an off-line Japanese signature verification is difficult due to a lack of stability and individuality. Yoshimura et al. showed that a pattern matching method is able to achieve a good verification performance for Japanese signatures [3], [4]. However, the similarity between two signatures obtained by a pattern matching method is affected by their stroke widths. The stroke widths vary with the pen used for signing. Even if signatures are written with the same pen, the stroke width may also vary with a normalization of a character size for the following processing. To solve this problem, it is effective to verify signatures after normalizing their stroke widths. In addition, a registered
signature selected from a set of reference samples for each signer affects the verification performance. Therefore it is necessary to find an appropriate selection rule of a registered signature. In this paper we propose a new pattern matching method and a selection method of a registered signature for Japanese signature verification, based on the above consideration. In our modified pattern matching method, the strokes of signatures are first thinned at one pixel width, and then thinned signatures are blurred by a fixed point-spread function. Successively the similarity between registered and examined signatures is calculated. Concerning selection method of a registered signature, a signature is extracted from a set of reference samples, and the mean value and standard deviation of similarities between the extracted signature and the other samples are calculated. The signature with the maximal ratio of the mean value of the similarity to the standard deviation is selected as the registered signature. We will also show that the proposed methods improve the verification performance, through the verification experiments.
2. Outline of signature verification system using a pattern matching Figure 1 shows the flow diagram of the signature verification process proposed in this paper. The input signature is a binary image. It is first segmented into component characters. The position and size of component characters are then normalized. Successively, the similarity between registered and examined signatures is calculated in a pattern matching process. Finally, an examined signature is classified into one of two categories: genuine or forgery.
2.1. Input signatures We deal with Japanese signatures on bank checks. Therefore, it is necessary for verification to extract signatures from bank checks. The input signature of this verification process is binary image, which is extracted
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
3. Modified pattern matching An annoying problem encountered in off-line signature verification using pattern-matching method is that the similarity between registered and examined signatures is affected by their stroke widths. Therefore, it is effective for stable and reliable verification to verify based on the structural similarity of signature strokes. To realize this idea, the stroke of signature (Figure 3(b)) is thinned at one pixel width without any change of stroke structure. In this work, we use the Hildich’s algorithm for thinning. Figure 4(a) shows the example of thinning result.
Figure 1. Flow diagram of signature verification process proposed in this paper.
Figure 2. Example of input signature images.
Figure 3. Segmentation and normalization of size and position.
from bank check by a color clustering method [6]. The example of input signature is shown in Figure 2. The size and the resolution of input image are 210x700 pixels and 300dpi respectively.
2.2. Segmentation and normalization Japanese signatures almost always consist of two to six kanji, hiragana and/or katakana component characters, and they are spaced each other. The input signature is segmented into component characters based on this characteristic [7]. The segmentation is performed referring the widths and the number of component characters of the registered signature. The example of segmentation result is shown in Figure 3(a). The size and position of every component character are normalized to a fixed size and a fixed position respectively. Figure 3(b) shows the result of size and position normalization. The normalized image size is 240x240 pixels.
Figure 4. Normalization of signature stroke width.
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
For stable pattern matching, the thinned signature represented in the xy-plane is blurred using the following equation.
g ( x, y ) = ∫∫ f ( x − α , y − β )h(α , β )dα d β
(1)
where g(x,y), f(x,y) and h (α , β ) are the blurred image, the thinned image and the point-spread function respectively. α , β denote the shift amounts in the xyplane. Equation (1) is the well-known relationship called convolution of f and h. As the point-spread function h (α , β ) , we define the binary square function as shown in Figure 5. In this figure, the black square region denotes h=1, and zero otherwise. The size of square region is 19x19. The example of blurred image is shown in Figure 4(b). Successively the pattern matching is performed using blurred signature images. Let Xi (u), Xi(l) be the pattern vectors of i’th component characters of examined and registered signatures respectively. The elements of these pattern vectors are pixel values of the blurred images. The similarity Si between i’th corresponding component characters of examined and registered signatures is calculated using the following equation.
Si =
( X i(u ) , X i(l ) ) ( X i(u ) , X i(u ) ) ( X i(l ) , X i(l ) )
4. Selection of registered signature The selection of registered signatures may affect significantly verification performance, because examined signatures are compared with the corresponding registered signature in the reference database for verification. A signature is represented as a point in a certain feature space. The effect of signature selected as a registered one is shown in Figure 6 schematically. In this figure, circular points denote the genuine signatures, and triangular points are the forged signatures. The solid line circles are the boundaries of the genuine signatures. Figure 6(a) shows the case of optimal selection. Selecting the signature at the center of genuine cluster can maximize the similarity between the selected signature and other genuine signatures. In other case as shown in Figure 6(b), the verification error rate may increase. Figure 6. Effect of selecting a signature as a
(2)
i= 1, 2,……, m (m : the number of component characters) where (*,*) denotes a scalar product of vectors. The resultant similarity S between examined and registered signatures is obtained by taking an average of the similarities Si as follows,
S=
1 m ∑ Si m i =1
registered one.
Table 1. Example of similarities between genuine signatures.
(3)
where m denotes the number of component characters of the signature. An examined signature is verified based on the resultant similarity S. On the basis of above idea, we propose the following method to select the registered signature from a set of reference samples. A signature is extracted from a set of reference samples, and the mean value µ and standard deviation σ of similarities between the extracted signature and the other samples are calculated. The µ
Figure 5. The point-spread function h used in Equation (1).
signature with the maximal ratio σ of the mean value of the similarity to the standard deviation is selected as the registered signature. Table 1 shows an example of pattern
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
matching results for 5 genuine signatures. In this case, the signature No.2, which gives maximal µ σ , is selected as the registered signature.
5. Threshold definition The verification is based on the assumption that the similarities for an individual writer tend to cluster, while those of a population of writers are more widely scattered. Moreover, the cluster of forgeries cannot be estimated previously. Therefore, we determine the threshold value for verification based on only the statistical property of genuine clusters. Let {Sj (j=1, 2, ….. , n)} be the obtained similarities between a registered signature and its n genuine signatures for training. Moreover, let µ and σ be the mean value and the standard deviation of { Sj } respectively. The threshold value T of genuine signatures for each registered one is defined as the following equation.
T = µ − aσ
Figure 7. Example of signature images used in the experiments.
( 4)
where a is a certain value (called threshold coefficient). For unknown examined signature, if the similarity S(u) of an examined signature is larger than T, then the signature is classified as a genuine and a forgery otherwise.
6. Experiments 6.1. Experimental method In order to test the verification performance, we have collected 110 genuine signatures from each of 10 signers and 100 forged signatures for each genuine one. 10 reference samples are extracted from the each set of genuine signatures to select the registered signature and estimate mean value and standard deviation of similarities. Therefore a total of 100x10=1000 genuine signatures and 100x10 =1000 forged signatures were used for test. Although the small number of reference samples seems unusual, a large number of samples are generally not available in practical applications. An example of signature images used in the experiment is shown in Figure 7. In order to evaluate the verification performance of the proposed method, we have performed three kinds of experiments as follows. (1) Proposed Method: described in this paper. (2) Method-1: with the modified pattern matching described in Section 3. and signatures selected randomly as a registered one.
Figure 8. FRR and FAR for various values of threshold coefficient a (Type B).
(3) Method-2: with a conventional pattern matching and signatures selected optimally as a registered one described in Section 4. Each verification process was implemented with C on a 1.9 GHz Pentium-4 PC.
6.2. Experimental results Figure 8 shows the example of verification results by the proposed method. As known well, two types of errors, false-rejection error rate (FRR) and false-acceptance error rate (FAR), hold a trade-off relationship for change of the threshold coefficient “a” in equation (4). The performance of three methods was evaluated using the error rate at which FRR and FAR are equal (the error rate of cross-point of FRR and FAR curves in Figure 8), because the determination of the optimal threshold coefficient “a” for all authentic signatures is a difficult problem. The performance of the proposed method, together with the performance of the other two methods is compiled in Table.2. Although there exist some
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
differences in the error rates according to the signer, the average error rates yield 9.1% by the Proposed Method, 14.8% by the Method-1 and 19.2% by the Method-2. Comparing these results, we can say that the modified pattern matching method with stroke width normalization and the optimal selection of a registered signature are able to reduce the average error rate of about 10.1% and 5.7% respectively. The processing time of the proposed method was 15 to 20 seconds per signature. Table 2. Error rates in % at FRR=FAR.
8. References [1] F. Leclerc and R. Plamondon, “Automatic signature verification: The state of the art-1989-1993”, International Journal of Pattern Recognition and Artificial Intelligence, 8, 3, pp. 643-660, (1994). [2] I. Yoshimura and M. Yoshimura, “Off-line verification of Japanese signature after elimination of background patterns”, International Journal of Pattern Recognition and Artificial Intelligence, 8, 3, pp. 693-708, (1994). [3] I. Yoshimura, M. Yoshimura and T. Tsukamoto, “Investigation of an automatic verification system for Japanese countersignatures on traver’s cheques”, Proceedings of the 7th IGS Conference, pp.86-87, (1995).
7. Conclusion In this paper, we proposed the modified pattern matching method, which is independent of signature stroke width and the selection method of a registered signature for Japanese signature verification. Through the experiments, it was confirmed that both methods proposed in this paper are effective for improving the verification performance. We are now working towards improving the verification performance and reducing the processing time. The determination of optimal threshold value is the future problem. We will also investigate more precise evaluation of the verification performance with large database and an extension of this work to Western signatures in the future.
[4] M. Yoshimura and I. Yoshimura, “Investigation of a verification system for Japanese countersignatures on traver’s cheques”, Transactions of the IEICE, J80-D-II, 7, pp.17641773, (1998) (in Japanese). [5] S. Ando and M. Nakajima, “An active search of local individualities for an off-line signature verification”, Transactions of the IEICE, J84-D-II, 7, pp.1339-1350, (2001) (in Japanese). [6] K. Ueda, T. Mutoh and K. Matsuo, “Automatic verification system for seal imprints on Japanese bankchecks”, Proceedings of the 14th ICPR, pp.629-632, (1998). [7] M. Yoshimura and I. Yoshimura, “Segmentation of Japanese signatures into component characters”, Technical Report of IEICE, PRMU97-3, pp.17-24, (1997) (in Japanese).
Proceedings of the Seventh International Conference on Document Analysis and Recognition (ICDAR 2003) 0-7695-1960-1/03 $17.00 © 2003 IEEE
