International Symposium on Computer Graphics, Vision and Image Processing (SIBIGRAPI), Rio de Janeiro, Brazil, 1998, pp.270-277
A System for Automatic Extraction of the User–Entered Data from Bankchecks ALESSANDRO L. KOERICH1 LEE LUAN LING2 1
CEFET/PR – Centro Federal de Educação Tecnológica do Paraná Av. Sete de Setembro, 3125, 80230–015, Curitiba, PR, Brasil
[email protected] 2
UNICAMP – Universidade Estadual de Campinas Caixa Postal 6101, 13083–970, Campinas, SP, Brasil
[email protected] Abstract. This paper presents a system for automatic extraction of the user–entered data from Brazilian bankchecks. We have assumed that the layout structure of bankchecks is standardized, that any bankcheck can be identified through the MICR line and that a sample of the background pattern is available for every bankcheck. Based on these assumptions, a template is designed for extracting the user–entered items of any bankcheck, no mattering which financial institution has issued it. First, the bankcheck is digitized through a scanner and its skew angle is computed by an algorithm based on the Hough transform. Next a template is generated and used for extracting the user–entered data. The extracted data still shows the presence of the background pattern, character strings, and vertical and horizontal lines. The background pattern is eliminated by a morphological subtraction operation while the baselines are erased using an algorithm based on the projection profiles. The printed character strings are eliminated through a morphological subtraction between the image covered by the signature area and a sample of the character strings generated by the system. Finally, a post-processing algorithm is used for recovering some erased pixels. Experimental results show that this approach is robust and efficient for automatic extracting the user–entered items from Brazilian bankchecks achieving a moderate processing time, very good image quality and excellent accuracy rates. Keywords: Bankcheck Processing, Document Image Analysis, Image Processing
1 Introduction The automatic processing of bankchecks represents a challenging research area due to the difficulties in solving the problems related to the complexity of bankchecks. Millions of handwritten or machine printed bankchecks have to be processed everyday. Since the bankcheck processing is merely a repetitional task where a large amount of documents of the same kind must be handled, it is desirable to realize this processing in an automatic fashion. In spite of intensive research efforts, the degree of automation in processing bankchecks is still very limited. Recent papers have addressed the problem of automatic processing of bankchecks [Cheriet et al. (1995), Dimauro et al. (1997), Knerr et al. (1997), Lee et al. (1997), Liu et al. (1996), Liu et al. (1997), Okada–Shridhar (1996), Okada–Shridhar (1997), Santos et al. (1997), and Su et al. (1997)]. The problem of processing bankchecks can be divided in two main branches: the extraction and the recognition of
the user–entered (filled–in) data. In this paper the main goal is to provide an approach for automatic extracting the user– entered data from Brazilian bankchecks. Currently, two strategies are used for extracting the user–entered data: thresholding techniques and image subtraction. Different thresholding techniques have been suggested to isolate the user–entered data from the bankchecks [Dimauro et al. (1997), Liu et al. (1997), and Santos et al. (1997)]. These techniques have shown good results only if the bankchecks do not have complex background patterns. If these techniques were applied to bankchecks in which the background pattern has colorful pictures and draws, it would be very difficult to find a threshold value to segment the background from the other elements. Furthermore, it would be very difficult to segment the printed information from the user–entered data since they can have similar gray levels. On the other hand, the techniques based on image subtraction have shown more
robustness to segment the user–entered data. Okada–Shridhar (1997) have proposed a method for extracting the user– entered data from American bankchecks that have colorful pictures on the background pattern. They have suggested an approach similar to that proposed by Yoshimura (1994) for extracting the signature from Japanese traveler’s checks. Both approaches perform a subtraction operation between two images: one of the filled–in bankcheck image and the same bankcheck image without the filled–in information. From the practical point of view the approaches proposed by Okada–Shridhar and Yoshimura may not be feasible since a large database is needed to keep a sample image of each unfilled bankcheck. Recently we have proposed a feasible method for extracting the user–entered data from Brazilian bankchecks by using a subtraction operation between images [Koerich– Lee (1997)]. The first step of this operation is to subtract a background pattern sample from the filled–in bankcheck image. After the elimination of the background pattern, a template was applied and the user–entered data was extracted. This approach has shown very good results in terms of quality of the final images, but in terms of computational costs, it is not very efficient, since a large amount of data must be handled (images with 1370 per 600 pixels). In order to overcome this drawback we proposed another approach where the information in which we are not interested is eliminated in a stage before the information extraction. The main idea is to handle only sub–images which contain the user–entered data, said, the digit amount, the worded amount, the place and date, the payee’s name, and the signature. After digitizing the bankcheck image is position–adjusted and a template is used to extract the areas where the user–entered data is supposed to appear. For each of the five resulting sub–images, the corresponding background pattern is subtracted from a sample stored in a database. The baselines present in the sub–images are detected by using an algorithm based on the projection profiles. The baselines detected are erased by substituting the corresponding positions with white pixels. The character strings present below the baseline dedicated for signature are eliminated by other morphological subtraction operation between the corresponding sub–image and a generated binary image which contains every make–up character string to be eliminated. In this new approach we include a tracing recovering algorithm which recover some parts of the user– entered data that can be lost during the baseline erasing. This novel approach assumes that a sample of every background pattern is available in a database and that Brazilian bankchecks have a standardized layout.
The organization of this paper is as follows. Section 2 gives a brief description of bankchecks. Section 3 describes, step by step, the extraction of user-entered data and the elimination of non–interested information. The experimental results are presented in Section 4. Finally, the conclusions are stated in the last section. 2 Description of Bankchecks In Brazil each financial institution is responsible for designing and providing bankchecks to its customers. The financial institutions must follow some general standard rules stated by the Central Bank of Brazil to manufacture checks. These standards determine the elements in a check, their location and size as well as specification for the paper. In a practical sense, these standards make that the Brazilian bankchecks have a uniform layout. In spite of the standardization, Brazilian bankchecks still present a complex layout structure, mainly due to the presence of the background pattern. Is spite of this standardization, the financial institutions are free to customize some parts of the bankchecks such as the background pattern. The background pattern is generally used to personalize the bankchecks, and they can assume different colors and have imprinted pictures and drawings. Other variations may occur among the bankchecks issued by different financial institutions such as different character fonts, special symbols, logos, lines, etc. Fig. 1 shows a sample of a Brazilian bankcheck. Fig. 2 shows the distribution of the information that the bankchecks may have, in accordance with the standards of Central Bank of Brazil. A more detailed description of the layout structure of the Brazilian bankchecks can be found in Koerich–Lee (1997). 3 User–Entered Data Extraction In a practical sense, the background pattern and the printed information only make the recognition of the user–entered data more difficult. The background pattern is only useful for security purposes. Generally, it is printed with special (secure) inks that can indicate forgery trials. The printed information is used to identify the bank, the customer, the account, the check serial number, etc. The printed information is also encoded and printed at the bottom of the bankchecks in a MICR line using CMC–7 character font. Thus, it is desirable that the redundant information, both the background pattern and the printed information, be eliminated from the digitized bankcheck images. In addition, this processing for redundant information elimination must preserve the physical integrity of the remaining information, i.e., the user–entered items.
(1)
BANK NUMBER, AGENCY NUMBER, ACCOUNT NUMBER, SERIAL NUMBER, AND INTERNAL PROCESSING CODES
(3)
WORDED AMOUNT
(4)
PAYEE'S NAME
(6) (5)
LOGO
(7)
NAME AND ADDRESS OF THE FINANCIAL INSTITUTION
(8)
(9)
Fig.1: A sample of Brazilian bankcheck
System Overview The system that we have designed for automatic processing bankchecks is composed by three main modules: data acquisition, database, and image processing as shown in Fig. 3. The data acquisition module includes two devices, an optical scanner and a MICR scanner. Actually there are many devices that includes both an optical and a MICR scanner. The optical scanner provides a digitized image of the bankchecks with 200 dpi spatial resolution and 256 gray levels for the image processing module, while the MICR scanner reads the MICR lines printed at the bottom of the bankchecks. Through the MICR line reading, the complete identification of the bankcheck is done. The database was designed to store financial institution information, the background pattern samples, customer’s information, bankcheck layout information, and other parameters needed for bankcheck processing. The image processing module is composed of several algorithms for extracting the user– entered data: position evaluation for adjusting the skew angle and evaluating the horizontal and the vertical position of the scanned bankcheck images; items extraction for extracting the user–entered data, background elimination for eliminating the background pattern present on the bankcheck images; baseline erasing and character elimination for eliminating the vertical and horizontal baseline, and the character string printed below the signature baseline. Finally, tracing recovering for recovering some pixels that were lost during the item extraction and baseline elimination. Database A database was designed to store images, parameters and data that will be used during the bankcheck processing steps. The images are related to the samples of the background patterns from the bankchecks. Several parameters are stored in the database such as the parameters used to adjust the template. The data represents the information printed on the bankchecks, such as customer's name, agency’s address,
(2) D I G I T A M O U N T
PLACE AND DATE
SIGNATURE, NAME, AND PERSONAL DOCUMENT NUMBER OF THE CUSTOMER
MICR CMC-7 LINE
Fig. 2: A bankcheck layout description
account number, etc. The information stored in the database must be indexed using adequate keywords that provide a unique identification for each account. The keywords used for indexing the information in the database are composed of the three fields present on the bankchecks, as follow: bank number
agency number
account number
To reduce the complexity of the database, it was designed to have three levels: bank–level, agency–level and account–level. Into the bank–level are stored the background pattern samples, and the parameters used to adjust the template, according to the financial institution characteristics. The data which is printed on the bankchecks regarding bank’s agency name and its address is included into the agency–level, and it is stored as ASCII text. In this level is also included a list of the agency’s customers and his/her account numbers. The account–level has the personal information about every customer and the data that is printed on the bankchecks, such as the customer’s name, his/her personal information, etc. This data is also stored as ASCII text. The size of the database is directly related to the number of the financial institutions, agencies and accounts. To store each background pattern, 800 Kbytes are used; however, each sample is common to every customer during the processing of the bankchecks. This is due to the fact that in general a financial institution does not have more than three different kinds of background patterns. Approximately 1.5 Kbytes are used to store the data of each customer. Position Evaluation The dominant orientation of baselines determines the skew angle of a bankcheck. It is expected that a bankcheck has a null skew angle, that is, the vertical and horizontal baselines are parallel with the respective paper edges. However, when a check is manually digitized, due to the influence of the
Real Bankcheck
DATA ACQUISITION
Optical Scanner
MICR Scanner
DATABASE
IMAGE PROCESSING
Background Pattern Samples Customer's Information Financial Institution Information Bankcheck Layout Information
Position Evaluation Items Extraction Background Elimination Baselines Erasing Characters Removal Tracing Recovering
Digit Amount
Worded Amount
Payee's Name
Date
Signature
Fig. 3: A system overview
operator, non–zero skew, as well as shift in horizontal and vertical direction may occur. Since the algorithms for image processing need to have a reference about the bankcheck position for correctly handling the images, a position evaluation and adjustment must be done. The Hough transform is useful for straight-line detection. It can be used to detect the horizontal baselines present on the bankchecks and their skews. This method is useful when the objective is to find lines that fit groups of individual points on the image plane. This method involves a transformation from the image coordinate plane to parameter space. When multiple points are collinear, their transformations will intersect at the same point on the transform plane. For bankcheck baselines, the interest is only about their skew angle; we are interested
neither in the coordinates of the lines nor in end points. We also supposed that the baselines have small skew angles. A practical investigation has shown that the digitized bankcheck images generally have skew angles lower than one degree. For these reasons, the Hough transform can be successfully used with bankchecks. A bankcheck digitized with a 200 dpi spatial resolution yields an image which has 1370 per 600 pixels Thus, a skew angle with one degree corresponds to a difference of 24 pixels between the first and the last column. This implies that one pixel represents a skew angle of 0.042 degrees. We select an area measuring 1370 x 40 pixels where the second baseline is supposed to appear, and use this area to evaluate the skew of the bankcheck. An affine transform is
used after the item extraction step for correcting the skew of the sub–images. The basic computational steps in the Hough transform are as follows. for (x between 0 and 1370) for (y between 0 and 40) if (pixel value is lower that 100){ for (θ between ± 2 degrees){ calculate ρ = x.cos θ + y.sin θ increment accumulator array at (ρ, θ ) } }
In spite of this skew angle adjustment, the digitized bankcheck images may also have some shift along the vertical and horizontal directions. A simple strategy is adopted to evaluate the horizontal and vertical positions. The gray level of the pixels of the first 10 rows and columns are analyzed. If the digitized bankcheck image presents any shift either in horizontal or vertical position, a row or a column with white pixels must be found. Thus, these shifts both in vertical and in horizontal directions are stored in two accumulators, VS and HS, respectively. Item Extraction Due to the standardization of the layout structure of Brazilian bankchecks, it is reasonable to design a template for extracting the interested data from any bankcheck, no mattering which customer has filled–in it or which financial institution has issued it, only using a prior knowledge about the domain. As we are only interested on the user–entered data, the template must include every area where this data is supposed to appear into the bankchecks. The other areas can be eliminated without compromising the understanding of the document. Nevertheless, these areas are not located exactly at the same position for every bankcheck. There are small position variations, depending on the financial institution that has issued the bankcheck. Thus, the template must be adapted for these small variations in order to avoid selecting wrong areas, what can cause the loss or the degradation of the interested data. From a basic template, we construct a database with the possible positions of the interested data. During the processing, the database is accessed and the convenient parameters are selected to adjust the template, according to the financial institution that has issued the bankcheck which is being processed. The template is generated considering the horizontal and vertical shifts of the digitized bankcheck image through the inclusion of the parameters HS and VS into the algorithm. The basic template is presented in Fig. 4. The
information presented into the white areas is maintained while the information that coincides with the black areas is eliminated. By applying the template we are able to eliminate the redundant information, that is, the information in which we are not interest, and segment the different user–entered data. Fig. 5 shows a resulted image after applying the template operation. The output of the item extraction algorithm are five sub–images representing the user–entered items: digit amount, worded amount, payees’ name, place and date, and signature. Background Pattern Elimination The sub–images representing the extracted items still have the background pattern and the printed information present, as shown in Fig. 5. The elimination of the background pattern is done through a modified morphological subtraction operation. Different from some previous works [Yoshimura (1994), Koerich–Lee (1997), Okada–Shridhar (1997)], the subtraction operation is performed only between the sub– images instead of the entire bankcheck image. Furthermore this operation also includes a tolerance band which tolerates small difference between the two images involved in the background pattern subtraction. In fact, different gray levels are expected mainly due to the different conditions during the scanning processes. Thus, the same template used for extracting the user–entered items is also applied in the background pattern sample for extracting the correspondent background pattern sub–images to be used in the subtraction operation. Assuming that CN = cn(x,y) is the output sub–image of the subtraction operation, AN = an(x,y) is a bankcheck sub– image and BN = bn(x,y) is a corresponding background pattern sub–image, the modified subtraction operation, which includes a tolerance band, is defined as: 1, c n ( x, y ) = 0,
if an ( x, y ) − bn ( x, y ) ≤ T otherwise
where T is a tolerance factor which was chosen to be 20. This modified subtraction operation also performs a thresholding operation, providing a binary image. Baseline Erasing The sub–images extracted from the entire bankcheck image may also contain baselines or bounding boxes that must be removed. The bounding boxes are generally present around the digit amount, while the baselines are more likely found
Fig. 4: The template
below the worded amount, the payee’s name, the place and date, and the signature. The elimination of the baselines is based on the projection profiles. Since the baselines generally have horizontal and vertical length close to the proper sub–image dimension, it is expected that their positions can be easily identified from the horizontal and vertical projection profiles. In other words, the maximum values of the horizontal and the vertical projection profile occur in positions where the baselines are present. First we compute the horizontal projection profile for cn(x,y) that is denoted by: PPhn ( y ) =
∑ c ( x, y ) n
x
The points where the projection profile has relatively high values indicate the position of the baselines. The position of each baseline is recorded, that is, the coordinates of the beginning and end points are stored. This information will be used further for baseline extraction and for tracing recovering. A similar approach is used for vertical line detection, but it is only applied on the digit amount sub–image, since it is the only image that has vertical lines present. The baselines for the sub–images are eliminated according to the following rule: 1 d n ( x, y ) = c n ( x , y )
if c n ( x, y ) ∈ line otherwise
where cn(x,y) represents the sub–image after the background elimination, dn(x,y) represents the output sub–image for baseline erasing, and line represents a set of pixels that are assigned as belonging to a baseline. Character Strings Removal The remaining printed characters are presented in the signature sub–images. It is very difficult to eliminate these
Fig. 5: The extracted areas
strings during the foregoing processing because the signatures generally have stylist traces that overlap these strings. These character strings comprises the name of the customers and his/her register identification number. To eliminate the printed character strings we use a morphological subtraction operation as used before to eliminate the background pattern. First, a binary image is generated. This image contains exactly the same character strings that we want to eliminate from the signature sub– image. In order to generate this binary image, the information previously stored in the database is used. This information consists of the customer’s name, his/her register identification number, the position of the strings, the font style and size, etc. A dilation operation is applied in the binary image to expand the edges of the characters. This is necessary to compensate some degradation or imperfect alignment between the images during the subtraction operation. It is expected that the subtraction of the signature sub–image from the generated binary image results in an image with only the signature trace. Tracing Recovering The elimination of the baselines and character strings also eliminates the points of the user–entered data where the intersection between two parts occurs. In other words, inevitably, a part of the user–entered data is lost. To solve this problem we have design an algorithm to recover the lost information. The basic idea is to use the positions of the extracted baselines and establish connections between the neighbors. For each pixel that may be fitted to the erased baseline pixel (N0), their 8–neighbors, as shown in Fig. 6, are analyzed. The pixel N0 is going to be filled if their neighbors were black. The algorithm that performs the recovering of the lost information is as follows: for an erased baseline for (x between 0 and 1370) if (N1 or N2 or N3 is 0 and N6 or N7 or N8 is 0){ fill N0 with 0 }
N1 N4 N6
N2 N0 N7
N3 N5 N8
Fig.6: The 8–neighbors of the central pixel N0
4 Experimental Results The proposed system was tested by real–life Brazilian bankchecks. One hundred and twenty images of handwritten bankchecks were used to test the performance of the system. The bankchecks were issued by 10 different financial institutions and were filled–in by 20 different writers. In the Tab. 1 are shown the results for extracting the user–entered data, that is, the digit amount, the worded amount, the place and date, the payees’ name, and the signature. The images are classified according to the visual quality, which is related to the physical integrity of the extracted items. Three classes were used for classifying the images, as proposed by Liu et al. (1996): correctly extracted items for images in which the data was correctly located and extracted. Images in which occur loss of data that does not permit its correct understanding, were classified as being incorrectly extracted items. The problem occurs mainly during the items extraction operation. Sometimes the handwriting tracing exceeds the boundaries of the areas assigned by the template and the information is cut–off. Finally, the unused items were those images that were correctly extracted but presented hard structural defects that do not allow their further use. Items
Correctly Incorrectly Extracted Extracted (%) (%) Digit Amount 99.2 0.8 Worded Amount 95.8 4.2 Payee’s Name 96.7 3.3 Place and Date 95.0 5.0 Signature 90.8 9.2 Table 1: Results for extracting items
Unused (%) 0.0 2.5 1.7 4.2 5.8
We have observed that this novel approach has reached similar rates as our previous approach [Koerich–Lee (1997)] in terms of accuracy in extracting the user–entered data, but in terms of processing time, the new approach has had a reducing of 300% in the time for processing each bankcheck. It is expected that the final prototype can reach less than 10 seconds for processing a bankcheck. The inclusion of a post– processing stage to correct the tracing of the extracted items, has decreased the rate of unused images. A comparison among the results achieved by the proposed method and the methods proposed by Cheriet et al. (1997), Dimauro et al. (1997), Heutte et al. (1997), Koerich e Lee (1997), Liu et al. (1996), Liu et al. (1997) e Santos et al. (1997) are shown in Tab. 2. In this table are related the size of the database and the percentage of the correct and incorrectly extracted items, where DA represents the digit amount, WA represents the worded amount, PN represents the payee's name, PL represents the place of issuing, DT represents the date of issuing, SG represents the handwritten signature, and NA represent no available results. The results provided by this comparison can not be analyzed only by the percentages, since the authors use different ways to represent the results. Furthermore, the databases and the application domains are different since the papers have focused different kinds of bankchecks such as American, Japanese, French, Brazilian, Canadian and Chinese bankchecks. Our system has addressed a complete solution for identifying and extracting the information from bankchecks. The other approaches focus mainly the extraction of the digit and worded amounts, and sometimes the extraction of the date. Neither of them has proposed a solution for extracting all the items of the bankchecks. By combining this system with digit recognition, word recognition and signature verification systems, an automatic bankcheck processing system for payment recognition might be feasible for practical applications [Lee et al. (1997)]. References
5 Discussion and Conclusion This paper presents a system for automatic extracting the user–entered data from bankchecks based on the prior knowledge about layout structure of the bankchecks. A complete solution for data acquisition, layout analysis, MICR data identification and recognition, user–entered data extraction, and tracing recovering are presented. The input of the system is a real–live bankcheck and the output provides digital images of the digit amount, worded amount, place and date, payee’s name and signature, as shown in Fig. 7.
A. L. Koerich, L. L. Lee, ”Automatic Extraction of Filled–in Information from Bankchecks Based on Prior Knowledge About Layout Structure”, First Brazilian Symposium on Document Image Analysis (1997), 322–333. G. Dimauro, S. Impedovo, G. Pirlo, A. Salzo, “Bankcheck Recognition Systems: Re-Engineering the Design Process”, III Int. Workshop on Frontiers in Handwriting Recognition (1996), 419–426
Fig. 7: The final sub–images for the user–entered data Method Proposed Method Koerich & Lee 97 Liu et al. 96 Liu et al. 97 Heutte et al. Dimauro et al. Santos et al. Cheriet et al.
Database 120 200 494 400 3374 300 20 25
Correctly Extracted Items (%) Incorrectly Extracted Items (%) WA PN PL DT SG DA WA PN PL DT 95.8 96.7 95.0 95.0 90.8 0.8 4.2 3.3 5.0 5.0 95.3 97.5 95.2 96.1 88.7 1.7 4.7 2.5 4.8 3.9 97.7 NA 98.9 NA 0.4 2.2 NA 1.0 97.0 NA 97.0 NA 2.5 3.0 NA 3.0 NA 2.0 NA 95.4 4.6 95.0 5.0 80.0 20.0 Table 2: A comparison of different extraction methods
DA 99.2 98.3 99.6 97.5 98.0
G. Dimauro, S. Impedovo, G. Pirlo, A. Salzo, “Automatic Bankcheck Processing: A New Engineered System”, Int. Journal of Pattern Recognition and Artificial Intelligence 11 (1997), 467–504. H. Su, B. Zhao, F. Ma, S. Wang, S. Xia, “A Fault–Tolerant Chinese Check Recognition System, Int. Journal of Pattern Recognition and Artificial Intelligence. 11 (1997), 571–594. J. E. B. Santos, F. Bortolozzi, R. Sabourin, “A Simple Methodology to Bank Cheque Segmentation”, First Brazilian Symposium on Document Image Analysis (1997), 334–343. K. Liu, C. Y. Suen, C. Nadal, “Automatic Extraction of Items from Cheques Images for Payment Recognition, International Conference on Pattern Recognition, (1996), 798–802. K. Liu, C. Y. Suen, M. Cheriet, J. N. Said, C. Nadal, Y. Y. Tang, “Automatic Extraction of Baselines and Data from Check Images”, Int. Journal of Pattern Recognition and Artificial Intelligence 11 (1997), 675–697. L. Heutte, P. B. Pereira, J. V. Moreau, B. Plessis, P. Courtellemont, Y. Lecourtier, “Multi-Bank Check Recognition System: Consideration on the Numeral Amount Recognition Module”, Int. Journal of Pattern Recognition and Artificial Intelligence 11 (1997), 549–570.
SG 90.8 11.3 NA NA
L. L. Lee, M. G. Lizarraga, N. R. Gomes, A. L. Koerich, “A Prototype for Brazilian Bankcheck Recognition”, Int. Journal of Pattern Recogn. and Artif. Intelligence 11 (1997), 549–570. M. Cheriet, J. N. Said, C. Y. Suen, “A Formal Model for Document Processing of Business Forms”, 3rd Int. Conference on Document Analysis and Recognition (1995), 210–213. M. Okada, M. Shridhar, “A Robust Approach to Extract User Entered Information From Personal Bank Checks”, III Int. Workshop on Frontiers in Handwriting Recogn. (1996), 459– 466. M. Okada, M. Shridhar, “Extraction of User Entered Components from a Personal Bankcheck Using Morphological Subtraction”, Int. Journal of Pattern Recognition and Artificial Intelligence 11 (1997), 699–715. M. Yoshimura, I. Yoshimura, “Off-Line Verification of Japanese Signatures After Elimination of Background Patterns”, Int. Journal of Pattern Recognition and Artificial Intelligence, 8 (1994), 693–708. S. Knerr, V. Anisimov, O. Baret, N. Gorski, D. Price, J. C. Simon, “The A2iA Intercheque System: Courtesy amount and Legal Amount Recognition for French Checks”, Int. Journal of Pattern Recognition and Artificial Intelligence 11 (1997), 505– 548.
