Document analysis systems and processes

0001The present invention relates to document analysis systems and processes. It is concerned with systems and processes that identify a document and fields or informational elements within such document.

0002The ability to automate the process of data extraction from digital images of paper greatly increases the productivity and capacity of any business. By automating the data entry process, operating costs can be reduced due to reduced manpower requirements and capacity can be increased by maximizing the throughput of data entry operations. In businesses such as banking, there is a need for high speed processing of all types of documents. Banks usually receive many types of documents that must be entered into computers. In addition new services can be offered by automating data extraction from documents. As an example, extending Automatic Teller Machines (ATM) capabilities to include document processing would allow customers to cash cheques and receive funds back from the ATM. In the retail business there is a need to capture information from a customer at the point of sale. The information to be extracted could be from the customer's cheque or from his identification card (i.e. driver's license). Other applications for this invention is in wholesale and retail lockboxes. The more automated the process for handling payments, the lower the operating costs and the sooner the payments are credited.

0003To provide a robust solution, the system must be capable of processing a mixed stream of document types. Also, the system must be capable of processing fixed format documents as well as highly unconstrained documents.

0004To use a document analysis system the document is first scanned using one of several existing techniques such as a charged coupled device (CCD) to create a digital image representing a matrix of the black and white points on the page by a matrix of 0's and 1's. This matrix is then transmitted to a digital computer where it can be processed, displayed, identified and stored.

0005The requirements for processing the scanned document can be divided into two types: identification and decomposition of constrained documents and identification and decomposition of unconstrained documents.

0006The identification of constrained documents has been solved for various document types. For example the DP-ATM Model 5665 system produced by NCR in Dundee, Scotland, the HITC form identification system, available from NCR in Atlanta and the FIRST system developed by Quest, a division of Lucent Technologies are three systems that identify and process constrained documents. These systems classify the document as a specific type and then invoke an identification subsystem to process the type of document identified.

0007For unconstrained documents, some systems do locate and read specified fields. For example, to find the amount in figures, also known as the courtesy amount, some systems search a designated area for a specific currency character, for example the '$' character. However, in many cases a '$' or other currency character may not be present in the field and the courtesy amount may not be within the area designated. Furthermore, these systems may not have the capability of locating other fields that are not clearly delineated. These systems include the Courtesy Amount Locator (CAL) by Quest, a division of Lucent Technologies, and Scaleable Image Item Processing System (SIIPS) available from NCR, Waterloo, Canada. Other systems attempt a trial-by-error approach. These systems search for the field in a list of locations ranked by probability of success. This brute force approach is inaccurate and inefficient as it requires a large amount of processing resources. Furthermore, these systems cannot interpret data from different document types, i.e., cheques, deposit slips, and miscellaneous financial documents. All these existing systems focus on a specific task and do not analyze the overall structure of the document to derive a solution.

0008While it may be possible to search for a field, for example the courtesy amount, by processing each and every print field on the right side of the document, this results in a waste of processing resources since a large region of the scanned image must be searched and the region may include a large number of fields. Therefore, one of the problems that exist today in document analysis is to create a system that determines what information is useful in generating a solution for identifying a document and fields within the document. This solution would be based on information gathered from the scanned image, as well as the computational cost involved in generating the solution.

0009It is the object of the invention to provide such a solution.

0010According to the invention a system for analyzing a target document including at least one informational element, the system comprising: <ul id="ul0001" list-style="none" compact="compact"><li>(a) means for storing a digitized image of the target document;</li><li>(b) document feature extractor means for extracting low level features from the digitized image;</li><li>(c) document classification means for classifying the document based upon the extracted low level features; and</li><li>(d) document analysis means for analyzing the target document in order to extract informational data associated with the at least one informational element, responsive to the document classifying means.</li></ul>

0011Document recognition and data extraction is accomplished by an advanced document feature extractor that locates, identifies and extracts low level features within bitonal and grayscale images. Low level features include machine print text, hand print text, and vertical and horizontal lines. The low level features are passed to a document classification means which forms initial hypotheses about document identity. The document identity is passed to a document analysis means that handles both constrained and unconstrained documents. For constrained documents, the document analysis may sort through various models to determine the exact type of document and then extracts the relevant fields which are sent for optical character recognition (OCR). For unconstrained documents the system creates information and hypotheses to identify and locate relevant fields within the document. The fields located are then sent for optical character recognition (OCR).

BRIEF DESCRIPTION OF THE FIGURES OF THE DRAWINGS

0012<ul id="ul0002" list-style="none" compact="compact"><li>Fig. 1 is an overall illustration of the invention,</li><li>Fig. 2 is a breakdown of the Document Analysis</li><li>Means shown in Fig. 1,</li><li>Fig. 3 is a representation of a complex object.</li><li>Fig. 4 is an example of the hierarchy of information in the Knowledge Database shown in Fig. 2,</li><li>Fig. 5 is a sample business cheque,</li><li>Fig. 6 is an example of simple and complex boxes,</li><li>Fig. 7 is a graph of the height attribute in a document model file,</li><li>Fig. 8 is a hypotheses graph,</li><li>Fig. 9 is a flow chart of the control means execution cycle,</li><li>Fig. 10 is a flow chart of the event handler,</li><li>Fig. 11 is a business cheque with tokens,</li><li>Fig. 12 is a document model file for personal cheques,</li><li>Fig. 13 is a document model file for a business cheque,</li><li>Figs. 14A and 14B is an example of a discriminant matrix.</li><li>Figs. 15A and 15B is an example of a centre matrix.</li><li>Fig. 16 is an example of the hardware setup for developing the invention.</li><li>Fig. 17 is a flow chart for chequeing if a knowledge source requires more than one event.</li><li>Fig. 18 is a more detailed flow chart of the control means execution cycle.</li><li>Fig. 19 is a more detailed flow chart of the event handler.</li></ul>

0013A preferred embodiment of the invention is now described in detail. Referring to the drawings, like numbers indicate like parts throughout the views.

0014Referring to Fig. 1, the input to the system is the scanned image of an unknown document 100. Document 100 contains fields of data which are informational elements. For example the informational elements for a business cheque include the signature line, courtesy amount, legal amount, MICR and date.

0015In the preferred embodiment a TIFF format scanned image is used, grey-scale (8 bits/pixel) or binary, 200, 240, 254, 300, 400, 600, 1200 or 2400 dpi, both sides of the image should be scanned and any orientation of the image is supported. The image may be uncompressed (raw) image, or a compressed (binary or grey) image. However, any image type is readily usable. The image scanned may be stored on any storage media including hard disk.

I. Document Feature Extractor

0016From the scanned image 100 it is possible, using Document Feature Extractor 105 to locate and identify basic features of the document. These basic features may include machine print, hand print, vertical lines, horizontal lines and noisy regions within the digital image of the document 110.

0017The Document Feature Extractor 105 identifies the four corners of a rectangle that enclose individual document features. Features that are positioned close to one another or of the same type may be clustered into a single feature. For example, characters that compose an individual machine print word may be grouped into a single machine print feature, as well as multiple machine print words that are located close together within the document. The output of the Document Feature Extractor is the location and type indicator for each document feature. An example of a Document Feature Extractor 105 is the Net32K product available from AT&amp;T Laboratories in Holmdel, New Jersey.

0018Other features, systems and methods that can be used as the Document Feature Extractor or as part of the Document Feature Extractor are disclosed in <i>A High Speed Image Understanding System</i> in <u>Adaptive Analog VLSI Neural Systems</u> by M.A. Jabri, R.J. Coggins, and B.G. Flower, published by Chapman &amp; Hall 1996, ISBN 0-412-61630-0; <i>Analysis of Complex and Noisy Cheque Images,</i><u>Proceedings of IEEE International Conference on Imaae Processing,</u> pp. 316-319, published by IEEE Computer Society Press 1996, ISBN 0-8186-7310; <i>A Neural Network Accelerator for Image Analysis,</i><u>IEEE Micro,</u> vol. 15, number 3, June 1995, published by IEEE Computer Society Press 1995; <i>Net32K High Speed Image Understanding System,</i><u>Proceedings of the Fourth International Conference on Microelectronics for Neural Networks and Fuzzy Svstems</u> (Microneuro '95), pp. 413-421, published by IEEE Computer Society Press 1995; <i>Address Block Location with a Neural Net System,</i><u>Advances in Neural Information Processing Systems,</u> vol. 6, pp. 785-792, published by Morgan-Kaufman Publishers 1994.

0019The location and type indicator for each document feature is passed to Document Classification Means 115. Document Classification Means 115 identifies the general document class of the document scanned from the set of low level features extracted by Document Feature Extractor 105. A general document class is for example fixed format document with a high degree of linear features such as giros or single fixed format documents such as United States ANSI standard personal cheques or an unconstrained document such as a business cheque.

0020The output of Document Classification Means 115 is an ordered list of classification hypotheses, consisting of a document type identifier and a score which indicates the confidence level of the hypothesis 120. Details of the Document Classification Means are described below.

0021The output features 110 of the Document Feature Extractor 105 and the ordered list of classification hypotheses 120 from the Document Classification Engine 115 are passed to the Document Analysis Means 125. The Document Analysis Means 125 locates relevant areas or zones within the document. For example, for a U.S. business cheque 145 or personal cheque 130, the Document Analysis Means 125 outputs the location and print type of the courtesy amount which is shown in figures, the legal amount which is shown in words, MICR, date, and signature zone. Details of the Document Analysis Means are described below.

0022In the preferred embodiment of this invention, the code for the Document Classification Means and the Document Analysis Means may be written in Microsoft Visual C++. The software may run on an Intel 486, Pentium or Pentium Pro platform under Windows NT or Windows '95. It also may utilize the Win32 application subsystem.

0023One setup of a system for developing the invention is set forth in Fig. 16. A computer 1610 is shown which contains a monitor 1613, keyboard 1614, a mouse 1615, a central processing unit 1611 which can be an Intel 486, Pentium or Pentium Pro processor, memory 1612, and a hard disk 1616. The computer 1610 may be connected to a network. The network may contain hard disks 1620 and a scanner 1630. The scanner 1630 is used to scan the document into an image that can be stored either on hard disks 1620 or 1616. While this configuration is disclosed for the preferred embodiment, equivalent hardware systems or configurations may be substituted.

II. Document Classification Means

0024Document classification is divided into two parts: training and run-time. During system training, immediate features such as machine print, hand print, vertical lines and horizontal lines are extracted from a set of sample documents using the Document Feature Extractor 105. A set of indirect features are then computed based on the set of immediate features. Indirect features are created by putting together immediate features. A list of indirect features may include: image width, document boundary width, horizontal line total length, horizontal line total count, horizontal line average length, vertical line total length, vertical line total count, vertical line average length, machine print total area, machine print total count, machine print average area, hand print total area, hand print total count, hand print average area, machine print vs. hand print area ratio, machine print vs. hand print count ratio, machine print vs. hand print average area ratio, and total number of boxes.

0025Based on the set of immediate and indirect features, a set of discriminant vectors, one for each document class, is derived from optimizing a Fischer criteria using standard statistical procedures. The Fischer criteria is based on the concept that features from different fields on the document are spaced apart, while features from the same field are clustered together. Information from the discriminant vectors are stored as a file that is used at run-time.

0026At run-time, immediate features are read from the Document Feature Extractor 105 and indirect features are generated. Both the immediate features and the indirect features for the document read form a test set of features. The test set of features are put in a matrix and projected onto the discriminant vector for each class of document (implicitly the discriminant vector includes the features for that class of document). The result of the projection is a value indicating the distance between the test set of features and the features for each document class. The document is classified as belonging to the document class that has the smallest projection value.

0027For example, given the image vector of the form <ul id="ul0003" list-style="none" compact="compact"><li>1st column: Bounding box width of the image</li><li>2nd column: Bounding box height of the image</li><li>3rd column: horizontal line length</li><li>4th column: vertical line total count</li><li>5th column: vertical line average length</li><li>6th column: machine print total area</li><li>7th column: machine print total count</li><li>8th column: machine print average area</li><li>9th column: machine/hand print total area ratio</li><li>10th column: machine/hand print total count ratio</li></ul> image_vector = [ 1632 730 31727 8 151.88 259 41 6.34 63.9 66.3]

0028The values in the image_vector are values for each of the ten features of the business cheque defined above. These values represent the pixel values for those features.

0029Given the discriminant vector shown in Figs. 14A and 14B for the six pairs of possible combinations of four forms (Business Cheque, Deposit Slip, Giro, Personal Cheque): <ul id="ul0004" list-style="none" compact="compact"><li>1. Business Cheque vs. Deposit Slip</li><li>2. Business Cheque vs. Giro</li><li>3. Business Cheque vs. Personal Cheque</li><li>4. Deposit Slip vs. Giro</li><li>5. Deposit Slip vs. Personal Cheque</li><li>6. Giro vs. Personal Cheque</li></ul>    discriminant_vector = [Figs. 14A and 14B]

0030Given the centre vector shown in FIGS. 15A and 15B, a 12 x 2 vector for the six pairs of possible combinations of four forms:    centre_vector = [Figs. 15A and 15B]

0031For testing the first pair, the discriminant vector for the first pair (Business Cheque vs. Deposit Slip) is selected:<img file="EP0809219A2_D0001.tif" />

0032To get a projection vector, the image_vector is multiplied with the first_pair_discriminant_vector:    projection vector = image_vector ∗ pair_discriminant_vector.

0033The centre vector for the same pair (Business Cheque vs. Deposit Slip) is selected:    business_cheque_centre_vector = [-5.1658998e+001 -1.6968360e+002]    deposit slip_centre_vector = -423.03635e+001 -1.3201835e+002]

0034The distance from the current image to the business cheque is computed:    dist_business_cheque = norm(projection vector-business_cheque_centre_vector)    dist_business_cheque = 0.5759.

0035The distance from the current image to the deposit slip computed:    dist_deposit_slip = norm(projection_vector_deposit_slip_centre_vector)    dist_deposit_slip = 38.7298.<maths id="math0001" num=""><math display="block"><mrow><mtext>norm</mtext><msqrt><msub><mrow><mtext>(x</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext> - y</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msup><mrow><mtext>)</mtext></mrow><mrow><mtext>2</mtext></mrow></msup><msub><mrow><mtext> + (x</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><msub><mrow><mtext> - y</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><msup><mrow><mtext>)</mtext></mrow><mrow><mtext>2</mtext></mrow></msup><msub><mrow><mtext>•••+(x</mtext></mrow><mrow><mtext>n</mtext></mrow></msub><msub><mrow><mtext> - y</mtext></mrow><mrow><mtext>n</mtext></mrow></msub><msup><mrow><mtext>)</mtext></mrow><mrow><mtext>2</mtext></mrow></msup></msqrt></mrow></math><img file="EP0809219A2_D0002.tif" /></maths>

0036Since the distance to the business cheque is smaller, one credit is given to business cheque. By the same process, the distance to the possible six pairs is computed and the final results are:    Business Cheque = 3 credits    Deposit Slip = 0 credit    Giro = 2 credits    Personal Cheque = 1 credit

0037The system therefore concludes that the image is a Business Cheque.

III.    Document Analysis Means

0038The Document Analysis Means 125 processes constrained and unconstrained documents differently. If the document is identified by the Document Classification Means as constrained, the document analysis means sorts through various model documents of the type identified and extracts the fields that are relevant for the document identified. For personal cheques, the document model file is accessed, as explained below, and the relevant fields are located. For giros and deposit slips, the system sorts through various model files for the type of document and the relevant fields are located. The sorting through files for giros and deposit slips may be accomplished by systems that process constrained documents such as the DP-ATM Model 5665 produced by NCR in Dundee, Scotland.

0039For unconstrained documents, the problem is more complex. We therefore start by explaining the problem solving model used by the invention. Most problem solving models generally use a backward-reasoning process or forward-reasoning process in order to construct a solution. In a backward-reasoning approach, the system attempts to reason backward from a goal to be achieved to an initial data state. This type of system is often implemented as a rule based system, where if-then rules are defined and the system performs inference in order to work backwards from the goal state to the initial state. This problem approach is similar in nature to top-down model based problem solutions. In forward reasoning, the system attempts to reason forward from the initial data state to the goal state. Again, this is often achieved by inference rules, with the rules being applied to the current set of information until a goal state is found. This problem approach is very similar to bottom-up data driven problem solutions. Regardless of the model selected, both approaches require the system to determine when and how information should be applied to generate the problem solution.

0040Rather than be bound by either solution model, the document analysis system taught by this invention applies information in an opportunistic manner. Dependent upon the state of the system, the system dynamically determines which of the two approaches will yield the most useful information to derive a solution. To achieve this dynamic process flow, the document analysis system is implemented as a blackboard architecture. Referring to Fig. 2, the document analysis means 125 consists of three components: knowledge sources 231, 232 and 233, a control module 250, and a knowledge database 210.

0041The knowledge database 210 is the repository of all information and data in the system. The knowledge database 210 contains general a priori knowledge in the form of document models, specific information about the current image being processed, as well as hypotheses created by the system. The knowledge sources 231-233 communicate with each other via the knowledge database 210.

0042Knowledge sources 231-233 consist of a set of procedures or subsystems that are invoked to add to the current set of information contained within the knowledge database 210. Inputs to knowledge sources are stored in the knowledge database. The output of each knowledge source is stored in the knowledge database so that other knowledge sources can access the data. Knowledge sources may be simple procedures, or in some instances may be complete systems, such as the Net32K or field understanding engines such as optical character recognition systems.

0043A control module 250 is required to control the firing or invocation of knowledge sources 231-233. The control module monitors the state of the knowledge database 210, determining the best knowledge source to activate given the current state of the system. This scheduling activity is where the system selects to perform forward reasoning versus backward reasoning.

A. Knowledae Database

0044Information in the knowledge database 210 may be subdivided into the following types of information: image specific knowledge, a priori general knowledge, hypotheses, and system state information.

1. Image Specific Knowledge

0045Image specific knowledge is information that is derived from the scanned image such as machine print features or boxes. Information on the scanned image is subdivided into two sets: tokens and complex objects. Tokens are the smallest unit of information. Tokens are generated by Net32K and may divided into five types of objects: machine print tokens, hand print tokens, vertical line tokens, horizontal line tokens and graphical object or noise tokens.

0046Each token is represented as an object in the system and contains information that describes the size and location of the token, the type of token, and a confidence level generated by Net32K. The confidence level indicates the confidence of the token type. Associated with each machine print and hand print token is an additional field that contains the results of recognition operations performed on the token. This field is provided by the field understanding engines in the knowledge database.

0047Complex objects are structures that represent relationships between tokens or objects and other complex objects. As shown in Fig. 3, complex objects are represented as a directed graph, where each node defines a relationship that exists between other objects in the graph. In many instances, each node is a new object. As an example, Fig. 3 shows the data structure to represent a machine print object that is contained within a box. A machine print object 320 is represented as machine print 321 and a box 322. The box 322 is represented by its components, vertical lines 323 and 324 and horizontal lines 325 and 326. The data contained within each node, or object, consists of an object identifier, a confidence level, reference to the constituent objects or tokens that compose the new object, and information describing the size and location of the object.

0048The information contained within the knowledge database 210 is partitioned into a hierarchy of knowledge. Complex objects are structured from more primitive objects. Information can be therefore partitioned as shown in Fig. 4. Tokens 430 are at the bottom of the hierarchy. The next level up, 429, are objects formed by the spatial relationships between tokens. For example two horizontal lines .05 inches apart. The next level in the hierarchy 428 includes more objects formed by geometric relationships between objects or tokens. For example line intersections forming boxes. The highest level may be spatial relationships between tokens and objects 427 as described for Fig. 5.

0049The following information describing spatial relationships between tokens may be represented in the system. <ul id="ul0005" list-style="bullet" compact="compact"><li>groups of tokens based on horizontal location</li><li>groups of tokens based on vertical location</li><li>isolated tokens</li></ul>

0050In many cases, machine print tokens or hand print tokens that are located on the same line of text can be grouped together as a single token. In Fig. 5, the text    PAY *******158*DOLLARS AND *11CENTS may be identified as multiple machine print tokens by the Net32K board. However, with respect to the overall document composition, these fields are grouped into a single line of machine print text, since they are approximately the same font size, located on approximately the same position on the y-axis, and located closely together.

0051Similar to the grouping of tokens into a line of text, vertical grouping of text can be performed in order to create blocks of text. In Fig. 5, the text    NCR CORPORATION    ATTN: PAT DIXON, ROOM 8147    1200 P'TREE ST., PROMENADE I    ATLANTA,   GA 30309 may first be grouped horizontally to form four individual lines of text. Next, the relationship between the lines of text can be analyzed to determine if the text forms a block of text.

0052There can also be represented geometric relationships between tokens. These are restricted to information that can be derived from the horizontal and vertical line tokens created by Net32K. These relations listed can include line intersections, simple and complex boxes.

0053Line intersections are self explanatory. Simple boxes are boxes that are created by examining the endpoints of line segments. Complex boxes are boxes that are created by including line intersections, rather than only analyzing line endpoints for boxes. In general, simple boxes form the largest box that encompasses smaller, complex boxes, if complex boxes exist.

0054Fig. 6 illustrates simple and complex boxes. The simple box is denoted by the solid line, and is created by lines L1, L2, L3, and L4. The complex boxes, or boxes contained within the simple box, are created by the following set of lines.    <L1, L4, L5, L6>    <L1, L2, L5, L6>    <L3, L4, L5, L6>    <L3, L2, L5, L6>

0055After complex objects have been created, spatial relationships between objects are formed such as enclosed objects. Enclosed objects refer to objects or tokens that are enclosed by either a simple box or a complex box. Other types of spatial relationships may include underlined objects, or objects that have headers or labels.

2. A Priori General Knowledge

0056Given the variety of documents that can be processed by the system, a priori general knowledge can be represented in various forms. Prior to discussing the representation of this information in the system, the information that is represented must be defined.

0057To perform the initial classification process, general characteristics of the various document types must be identified. For example, giros can be described as documents that generally contain a mix of machine print and linear features,.with little or any hand print features present in the document. These documents often contain a large number of linear features.

0058U.S. personal cheques generally contain a mix of machine print and hand print features, with a fairly large number of hand print features. Since the document layout is controlled by ANSI standards, the location of specific hand print features can also be exploited to perform the classification process. In cases where the cheque has been typed, this definition will not classify the document as a personal cheque. This is actually not a problem, since the document would be classified as a business cheque and the correct information would be located by the business cheque processor.

0059Business cheques generally contain a mix of machine print and linear features. However, the linear features are not as prevalent in a business cheque. Also, most business cheques contain a signature in the lower right portion of the document to help differentiate between giros and business cheques.

0060Deposit slips generally contain a mixture of machine print, hand print, and linear features. However, in image ready deposit slips, the line features will drop out leaving primarily machine print and hand print features. The hand print features generally do not match the layout of a personal cheque, making the classification process viable.

0061The approach used for modeling a constrained document is simple since all the fields are at a predefined location. An example of a model file for a personal cheque is shown in Fig. 12. As shown in Fig. 12, the model file is broken into zones. These zones include a signature zone 1205, MICR zone 1210, legal amount zone 1220, courtesy amount zone 1230 and date zone 1240. Each zone contains fields which further identify the zone. For example the signature zone contains the following fields: fixed position, type hand print, and position 0.65, 0.9, 0.55 and 1.0 (these are representative of Ymin, Ymax, Xmin and Xmax: For example given 0,0 as the top left corner of the cheque, the signature should be contained in the box formed by the four points (0.55, 0.65) (1.0, 0.65) (0.55, 0.9) and (1.0, 0.9). This data is interpreted to mean that the signature zone of a personal cheque is a fixed position field, it is hand printed and in the box designated by the position field.

0062The approach used for modeling a business cheque is more complicated since this is an unconstrained document. The approach is to define the set of attributes or properties that can be used to distinguish a field within a business cheque from other fields or extraneous information. As an example, a courtesy amount field may have the following attributes: machine print field, right area location, short field length, enclosed by a box, field contains '$' or '*', field contains numeric data, font size is approximately equivalent to the legal amount field, field to the right of the legal amount field, field located above the signature field. A legal amount field may contain the following attributes: machine print, location in the right area, long field, '*", alphanumeric data, font size approximately equivalent to the courtesy amount field, field to left of courtesy amount, field higher than signature field. A signature zone may contain the following attributes: hand print, located in the right bottom area, large field size, field is below the courtesy amount, field below the legal amount. The data zone is a machine print field. The MICR zone may contain the following attributes: machine print field, bottom location, long length, numeric data, font size in pixels.

0063Given the set of attributes for each field, candidate zones are identified for each object based on local information kept by the object. A simple rule based processor that evaluates each attribute as true or false in order to determine if a token or object meets the field requirements is appropriate. However, in most cases not every attribute is true, For example, the courtesy amount may not be enclosed in a box or may not contain a '$' or '*'. As a result, the set of attributes should be viewed as a set of heuristic that are used to locate candidates for the individual fields.

0064Since the attributes are considered as heuristics, a weight or value is assigned to each attribute to determine the amount of supporting evidence each attribute contributes in identifying a zone. Also, given that some attributes must be true, it should be possible to assign a negative weight to attributes that are proven false. These weights are used to generate probabilities that correspond to the degree of match with the model.

0065The definition of true and false is fuzzy with some attributes. If the attribute specifies a location or size, there may be a range where the attribute should be considered true, false, or partially true. For example, assume the attribute is that the zone height is between 15 pixels and 30 pixels. If the height for a candidate zone is 14 pixels, it does not make sense to mark the attribute as false, since the zone is very close to meeting the requirement of the attribute. The solution is to treat attributes not as binary relations, but adjust the score according to the degree of match with the attribute. The graph in Fig. 7 shows an example for the height attribute. If the height constraint is met, the score is 0.3. If the height is within 8 pixels of the constraint, the score is adjusted based on the distance from the defined height constraint using a linear formula. If the height difference is greater than 8 pixels, the score is 0.0.

0066In the current document analysis system, a field model, the attributes for a particular field as represented in the zones of the document model file, is implemented as a C++ object that is created and then instantiated with a set of attributes and weights from an input file. In this way, additional fields can be supported by simply creating a new model object and loading the appropriate model file.

0067An example of a model file for a business cheque is shown in Fig. 13. These zones include a signature zone 1310, an MICR zone 1320, a legal amount zone 1330, a courtesy amount zone 1340 and a date zone 1350. The signature zone for a business cheque is not in fixed position. It is probabilistic which means that it may or may not be found in the location or with the print type specified in the signature zone. Because of this uncertainty, confidence values are computed for each zone. For example, if the field found is hand print, a confidence value of .3 is assigned. If it is not hand print a confidence value of -.5 is assigned. For the position of the signature field it should be bounded by the box formed by 0.7, 0.9, 0.6 and 0.8 (Xmin, Xmax, Ymin and Ymax: Box designated by these four points (0.7, 0.6) (0.7, 0.8) (0.9, 0.6) (0.9, 0.8). Since this field is probabilistic, the box that encloses the Net32K token is also taken into consideration. If the Net32K box overlaps the box in the signature zone then a confidence value of 0.7 is assigned. If the two boxes do not overlap and the distance is greater than 0.5 inches, a confidence value of 0 is assigned. If the boxes overlap and their distance apart is from 0.1 - 0.5 inches then a confidence value between 0 and 0.7 is assigned. For the height, if it is 110 ± 80 pixels in height then a confidence value of 0.1 is given. If the field is 110 ± 90 pixels in height then a confidence value of 0.0 is given. For the width, if it is 350 ± 250 pixels in width a confidence value of 0.2 is assigned. If the width is 350 ± 270 a confidence value of 0 is assigned.

0068The process of whether an object matches a field is as follows: The object attributes are compared against the zones or field model attributes in the document model file; confidence values are calculated as explained for the example in Fig. 8; the confidence values for each attribute comparison are added to obtain a zone or field model confidence value; the object is matched to the field model or zone with the highest confidence value.

3. Hypotheses

0069From the image specific knowledge and from the priori general knowledge supplied to the knowledge database, various hypotheses can be generated. The structure and impact of hypotheses vary according to the type of hypotheses. Hypotheses can be classified as three types: (1) document classification hypotheses; (2) field level hypotheses; and (3) overall document composition hypotheses.

0070Document classification hypotheses are hypotheses about the current type of document being processed, such as a business cheque, giro; deposit slip, or personal cheque. These hypotheses are derived from the models of generic document composition as described for the priori general knowledge, and are used to guide the system in performing the document decomposition or identification step.

0071Field level hypotheses are hypotheses pertaining to individual fields within the document. The fields being considered are determined by the current document classification hypothesis being explored. If the highest ranking document classification hypothesis indicates a business cheque, the field level hypotheses are for the courtesy amount field, legal amount field, and other fields that must be located and recognized on the business cheque. For a giro, where the only problem is document identification, there are no field level hypotheses, since the location of fields is implicit in the identification of the document.

0072Field level hypotheses may consist of the following data. <ul id="ul0006" list-style="none"><li>1. <b>hypothesis identifier.</b> The hypothesis identifier indicates the hypothesis type, i.e. courtesy amount field hypothesis, signature field hypothesis, etc. The value of this field is static after hypothesis creation.</li><li>2. <b>object reference.</b> The object reference is a reference to the token or complex object to which the hypothesis applies. A hypothesis can only reference one object or token, and the reference is static after creation of the hypothesis.</li><li>3. <b>confidence value.</b> The confidence value is the probability that the referenced object is actually an instance of the field. This value is dynamic, since as new information is obtained relative to the object or token, the confidence value will change.</li><li>4. <b>hypothesis evidence.</b> The hypothesis evidence consists of the status of each attribute that is contained in the field level model, as well as the degree to which the status of the attribute impacts the confidence score. An attribute is either true, false, or unknown. As discussed earlier, an attribute may be false and still contribute to the confidence score. Hypothesis evidence is useful in determining what type of information is required to help provide evidence to support the hypothesis or prove the hypothesis false, as well as for understanding how the system determines field level hypotheses.</li></ul>

0073Hypotheses about overall document composition are generated by creating valid tuples, or combinations, of field level hypotheses. A valid business cheque tuple consists of field level hypotheses for each zone of interest that meets the constraints defined in the model of the overall document composition for a business cheque. In instances where not all fields exist within the image, the tuple would contain a NULL entry for that field. It should be realized that a single field level hypothesis may be included in more than one tuple, since a valid tuple may exist with more than one combination of fields. Fig. 8 illustrates the relationship between field level hypotheses, document level hypotheses, and tokens or objects.

0074In Fig. 8, there are two document level hypotheses that exist. Hypothesis H1 consists of the tuple {C1,L1,S1,M1} and hypothesis H2 consists of the tuple {C2,L1,S1,M1}. The hypotheses H1 and H2 share the same hypotheses for legal amount, signature, and MICR, with the only difference being in the courtesy amount zone. Note that there is a third possible tuple, {C3,L1,S1,M1}. However, this tuple fails to meet the constraints specified in the document model and a hypothesis is not created for this combination. Field level hypotheses that cannot be grouped with other fields, such as courtesy amount hypothesis 3, remain as field level hypotheses. Similar to field level hypotheses not associated with document level hypotheses, there exist image specific knowledge that is not associated with any field level hypotheses.

0075The information stored within a document level hypothesis is very similar to a field level hypothesis. The primary difference is that the object reference is to field level hypotheses, and consists of a variable number of field level hypotheses references.

0076Both hypothesis creation and hypothesis updates are implemented as a knowledge source. As a result, these tasks are managed by the control module that is responsible for scheduling the execution of knowledge sources. A description of the control module follows.

4. State Information

0077State information consist of data that is required to control and monitor the system. The state information is information that the control means or scheduler requires to determine what knowledge source to execute next. This information is explained as part of the control means.

B. Knowledge Sources

0078A knowledge source is a procedure or subsystem that is invoked to generate additional knowledge. The input to a knowledge source is information contained in the knowledge database, generally in the form of tokens, objects, or additional parameters that impact the results of the knowledge source.

0079Each knowledge source is an object, in the context of programming, that has the following attributes: set of activation events, set of preconditions, cost, benefit, and performance statistics.

0080For a knowledge source to become executable, a set of data must be created by the system that is compatible with the inputs to the knowledge source. By definition, the inputs to a knowledge source are a set of tokens and objects that are used by the knowledge source to generate new information. As a result, the set of inputs are well defined. Rather than have a knowledge source search the knowledge database for this input data, it is more efficient to define a set of events that represent the creation of specific information, tokens or objects, that are processed by the knowledge source. The system would only have to cheque if an event was generated that indicates the creation of data that can be processed by the knowledge source. To implement this scheduling mechanism, each knowledge source contains a list of events that cause the activation of the knowledge source.

0081After a knowledge source has been activated by an event, the knowledge source may have a set of preconditions that must be true prior to executing the knowledge source. This is a case where the creation of new knowledge, resulting in an event, has activated the knowledge source. Execution of the knowledge source, however, may still not be possible because some additional data may not exist that is required by the knowledge source.

0082Associated with each knowledge source is a cost that indicates the computational requirements of the knowledge source. This cost is state information and used by the scheduler to determine the best knowledge source to execute at any point in time. In many instances, the cost of executing a knowledge source may be fixed, not requiring the knowledge source to recalculate the cost at run time. However, in some cases the cost of the knowledge source may not be fixed. As an example, a knowledge source may be very sensitive to the number of objects or tokens that are input to the knowledge source. In this case, the knowledge source must be capable of calculating the true cost at run time based on the current state of the knowledge database.

0083Along with the cost of a knowledge source, the scheduler requires an estimate by the knowledge source on the value of the information to be generated by the knowledge source. This score is called the estimated benefit of the knowledge source. The benefit of a knowledge source can vary given the current activity or goal of the system. For instance, if the knowledge source is a routine that constructs boxes from linear features found in the document, the benefit of this knowledge source would be higher if the system is focusing on a model that contains box features, rather than a model that does not contain boxes. As a result, a knowledge source must be capable of reevaluating the potential benefit based on the current state of the system.

0084Each knowledge source should maintain performance statistics that can be used in determining the cost and benefit attributes of the knowledge source. One statistic is the average CPU time required to execute the knowledge source. This data can be collected in a test environment and analyzed in order to accurately determine the cost of the knowledge source. In a more sophisticated implementation, the knowledge source can use this information to determine the true cost at run time. In this example, the knowledge source cost is adaptable to the live run time environment of the system.

0085Analogous to capturing data relative to cost, the knowledge source should also collect statistics pertaining to the ability of the knowledge source to produce information that contributes to the problem solution. Again, this data can be collected in a test environment or run time environment in order to estimate the benefit of a knowledge source.

0086The following is an example list of knowledge sources: <ul id="ul0007" list-style="none"><li>1. <b>Clustering.</b> This knowledge source partitions tokens into sets based on their location within the input image and token type. A data object providing access to a list of objects by region is the output of this knowledge source.</li><li>2. <b>Horizontal Grouping.</b> This knowledge source groups machine print and hand print tokens that are located close together on the x-axis, and are approximately at the same location on the y-axis. Objects representing lines of text or partial lines of text are the output of this knowledge source.</li><li>3. <b>Vertical Grouping.</b> This knowledge source groups machine print tokens, hand print tokens, and groups of machine print or hand print tokens that are related according to there location on the x-axis and y-axis. This knowledge source may eventually be implemented as multiple knowledge sources that search exclusively for different groupings of text, such as columns of text or blocks of text. Objects representing blocks of text or columns of data are the output of this knowledge source.</li><li>4. <b>Simple Box Builder.</b> This knowledge source locates simple boxes in the input image. Objects representing simple boxes in the input image are the output of this knowledge source.</li><li>5. <b>Complex Box Builder.</b> This knowledge source locates complex boxes in the input image. Objects representing complex boxes in the input image are the output of this knowledge source.</li><li>6. <b>Enclosed Object Builder.</b> This knowledge source locates objects or tokens that are enclosed in both simple and complex boxes. Objects representing this relationship are the output of this routine.</li><li>7. <b>Token or Object splitter.</b> This knowledge source dissects large machine print or hand print fields that are intersected by lines and boxes into multiple tokens or objects. This routine compensates for inconsistencies that occur when the Net32K groups machine print or hand print features that are located closely together, as well as inconsistencies created by horizontal grouping. Objects or tokens, derived from the input set of objects or tokens, are the output of this routine.</li><li>8. <b>Underline Object Builder.</b> This knowledge source locates objects or tokens that are underlined in the image. Objects representing this relationship are the output of this routine.</li><li>9. <b>Courtesy Amount Locator/Reader.</b> This knowledge source determines if a specified zone in the input image is a courtesy amount field, based on the presence of a `$', other currency character or `*' character. The knowledge source also attempts to read the field. The output of the knowledge source is a confidence value indicating if the field is a courtesy amount field, as well as the contents of the field.</li><li>10. <b>Legal Amount Reader.</b> This knowledge source is responsible for reading legal amount fields from cheques. The knowledge source outputs the results of the legal amount reader.</li><li>11. <b>OCR Engines.</b> This knowledge source provides an interface to OCR engines in order to perform character recognition. This knowledge source outputs the results of the invoked OCR engine.</li><li>12. <b>Contextual Analysis.</b> This knowledge source provides the ability to perform contextual analysis of the results returned from the field understanding engines. This gives the system the ability to perform reasoning based on the contents of a field. For example, if the word AMOUNT is read from the image, the system needs the ability to make use of this information in locating the courtesy amount zone. At the present time, the proper use of this knowledge source or knowledge sources has not be completely explored.</li><li>13. <b>Header Analysis.</b> This knowledge source provides the capability to search for combinations of objects and/or tokens that form a field header and field value relationship. For example, if the document contains the word AMOUNT or TOTAL, in many cases the courtesy amount can be found beside or below this text. This relationship may also be found based solely on the location of tokens or objects. Objects representing this relationship are the output of the knowledge source.</li><li>14. <b>U.S. Personal Cheque Verification.</b> This knowledge source is called to verify that the current document matches the layout of a U.S. personal cheque. If the document is a U.S. Personal Cheque, the knowledge source outputs the location of the relevant fields in the image.</li><li>15. <b>Document Identification.</b> The document identification knowledge source provides an interface to the document identification engine previously developed by the HITC. for identifying giros. This knowledge source is Used to identify documents that contain linear features. On successful completion, this knowledge source outputs the identity of the document and the relevant field locations. The document analysis system can also align and validate the presence of field locations based on the document layout and locations of tokens. If the document is not identified, an unidentified status is returned.</li><li>16. <b>Net32K System.</b> While the Net32K output is considered an input to the document analysis system, as performance data becomes available it may be possible to optimize performance of the Net32K subsystem based on the run time environment. It may be possible to treat the software component that translates feature maps to tokens as individual knowledge sources. In this manner, the system will request processing of feature maps as needed, avoiding unnecessary computation. For instance, if the document distribution indicates that the document stream contains a large percentage of giros, line feature extraction can be performed initially. If there are a large number of lines present in the image, the Document Identification knowledge source can be scheduled to identify the document. If the document is successfully identified, the additional step of processing machine print features and hand print features can be avoided. Tokens are the output of this knowledge source.</li></ul>

C. Control Means

0087At any point in time, the system consists of a set of knowledge sources that can be executed to create new information that is used to construct new hypotheses or support current hypotheses. Since knowledge sources are activated with respect to a data object or a set of data objects, it is possible that a single knowledge source can be ready to execute upon the creation of a single data object. As described earlier, the computational power required by knowledge sources varies by the function being performed. The job of the control means is that of a scheduler, to select the best knowledge source given the state of the system.

0088First, a high level description of the control means is given below and in Fig. 9. Reference numerals in Fig. 9 are enclosed in brackets ([]) on the lines of pseudo code. In the sections that follow, a more detailed breakdown of each component in the control means is given. The following defines the high level control means for the system.<img file="EP0809219A2_D0003.tif" />

0089As shown by the pseudo code given above and Fig. 9, the control means is divided into three components. First, knowledge sources must be activated 911. Activation in this context refers to some event in the system that makes it possible to run a knowledge source. Second, from the set of activated knowledge sources, the system must select the best knowledge source to execute based on some selection criteria 912. This is a scheduling process. Finally, the knowledge source is executed, resulting in an update in the current state of the system 913. Each loop through the scheduler loop is termed an execution cycle.

0090In general, a knowledge source is a function that processes a set of input data, tokens or complex objects, and produces a set of output data, new or updated complex objects, as well.as hypotheses. By definition, a knowledge source cannot be executed until the set of input data required by the knowledge source exists in the knowledge database. Also, it is possible to define a set of conditions on the inputs or state of the knowledge database that must be met prior to executing the knowledge source. These conditions may pertain to the presence of specific knowledge concerning the inputs or specific state information relevant to the system. As a result, a knowledge source may exist in three states: idle, active and executing.

0091In the idle state, the knowledge source is not bound to a set of input objects, meaning there is no context for the knowledge source to process. In the activated state, the knowledge source has been triggered and bound to a specific context, but cannot be executed because all of the knowledge source preconditions have not been met. In the executing state, the knowledge source has been triggered and all preconditions have been met in the given context.

0092To manage the knowledge source states, two queues could be implemented, the activation queue (ACTIVE) and the executable queue (EXECUTE). An idle queue is not necessary, since each knowledge source is always in the idle state. Given the three states listed above, the primary requirement is to define an efficient mechanism that can be used to activate knowledge sources and evaluate their preconditions.

0093The simplest mechanism available for event activation is to scan the current knowledge database looking for new objects or objects that have been updated. If these objects are found, the scheduler can then search the set of knowledge sources to determine if a knowledge source can be activated within the context of the new or modified object. However, as the knowledge base and the set of knowledge sources grows, this search method becomes computationally inefficient.

0094The mechanism selected for event activation is an event queue (EVENT). When objects in the knowledge database are updated or additional objects are added, an event is generated and added to the event queue. Rather than have a knowledge source be activated by the presence of a specific piece of knowledge, knowledge sources are activated by events (triggers) that are generated by other knowledge sources in the system. An event consists of an event identifier and a pointer to the data object that was responsible for the event. Associated with each knowledge source are a list of events that cause a knowledge source to be activated. As shown below and in Fig. 10, the first step of the ACTIVATION step is to process the current set of events 1010-1012, resulting in the creation of knowledge source contexts (tasks) .1013. A task is an executable entity that is the result of binding a knowledge source and an object. After a task is created, the task is added to the active queue (ACTIVATE) 1014.<img file="EP0809219A2_D0004.tif" />

0095After a task has been created, the system must evaluate the task preconditions to determine if the task is executable. Since the state of the system changes each time through the scheduler loop, these preconditions need to be evaluated each execution cycle of the scheduler.

0096Similar to event activation, the simplest mechanism available for condition evaluation is to scan the ACTIVE queue evaluating the preconditions for each knowledge source context. If all preconditions for a knowledge source context are met, then the knowledge source context can be added to the EXECUTE queue. However, as the number of knowledge source contexts on the ACTIVE queue grows, the processing required to evaluate each precondition can grow out of control. Also, it is wasteful to reevaluate each precondition. In many cases, if no new knowledge has been created that can affect the precondition, there is no need to reevaluate the precondition. As a result, the goal is to create a mechanism that causes preconditions to be updated when the knowledge source context is updated, i.e. state information about the object is modified. Figs. 17 - 19 flow chart this mechanism and the pseudo code is given below with reference numerals enclosed in brackets ([ ]).

0097To meet these requirements, a link must be added so that given a specific object, the system can determine the knowledge contexts that are impacted by the object update. A link already exists from the knowledge source to the object, since a task is a pairing of a knowledge source and an object to be processed by the knowledge source. To achieve a link from the object to the set of preconditions associated with knowledge source, a link must exist between the object and the task. This link is initialized when the task is initially added to the ACTIVE queue 1750, 1840. When a task becomes executable, the link is removed so that preconditions are no longer evaluated 1970.

0098After a linkage has been created, the systems needs a mechanism to trigger the evaluation of the precondition. The evaluation of preconditions can be performed at two points. One possibility is when an object is updated, the knowledge source updating the object can determine if the object should trigger the evaluation of preconditions. If so, the knowledge source can follow the link to the task and invoke the routine to evaluate the preconditions. This method is not optimal, since if multiple objects are being updated by a knowledge source, the preconditions would be evaluated on each object update.

0099A more efficient method is to update a flag contained in the task to indicate that the precondition needs to be updated 1840. In this case, as the scheduler traverses the ACTIVE queue to determine if a task can be moved to the EXECUTE queue, only preconditions of knowledge sources marked out-of-date are updated 1940. This guarantees that the preconditions are evaluated only once per execution cycle, after the knowledge base update is complete 1950-1970.

0100Since objects and tokens are responsible for updating their own attributes, these objects will also be responsible for updating the list of tasks to which the object is linked. Also, knowledge sources can have more than one event that must occur before the knowledge source can execute.

0101In addition to the event handler and the evaluation of preconditions, a Knowledge Source may require more than one event before it can execute. If so, all events that trigger the Knowledge Source must exist or be on the EVENT queue. If all the events are on the EVENT queue then the Knowledge Source is ready to be executed and is placed on the ACTIVE queue 1750.

0102The following pseudo code contains the modifications to cheque if a knowledge source requires more than one event 1710-1750, the modifications to the event handler 1810-1840, along with the additional pseudo code required to evaluate preconditions 1910-1980.<img file="EP0809219A2_D0005.tif" /><img file="EP0809219A2_D0006.tif" />

0103Scheduling a knowledge source for execution consists of selecting the appropriate task from the execution queue (EXECUTE) based on a cost/benefit analysis, deleting the task from the execute queue, and invoking the task or knowledge source. In systems that support parallel processing capabilities, task scheduling can be more complex, since it may be possible to select more than one task for execution. In a multiprocessing environment, the scheduler must be capable of selecting tasks on the basis of computational resource availability, potential problems related to concurrent data access, as well as cost/benefit. As a result of the added complexity, and no current requirements for parallel processing, the task selection is based completely on a cost/benefit analysis.

0104Since cost and benefit are an attribute of knowledge sources, the scheduler calculates a cost/benefit score for each knowledge source and then selects the best knowledge source to execute. The current system calculates the score as benefit - cost, with the highest score being the best knowledge source to execute. The score is calculated from the benefit and cost values assigned to each knowledge source. The benefit and cost values are assigned during development of the system. The benefit value represents the value the information computed by the knowledge source has towards obtaining an answer. The cost is a value representing the cost of running the knowledge source. The score is computed by the benefit minus the cost.

0105The system recalculates the cost/benefit score each execution cycle of the control module, rather than maintaining an ordered list of tasks based on the cost/benefit score. In some instances, it may still be necessary to reevaluate the cost/benefit scores due to changes in the state of the system, <i>i.e.</i> the focus of attention (explained below), even if an ordered list is implemented.

0106On each execution cycle through the scheduler loop, the control module selects the next activity to pursue by the system based on the cost/benefit score. This implies that the scheduling mechanism is data driven, since knowledge sources become executable on the creation of specific data or knowledge. This process is basically a bottom-up problem solving approach. While this scheduling method is satisfactory for generating a solution, it is not the most efficient mechanism available for guiding the system to the problem solution.

0107A more efficient method is to design the system so that it is goal driven, where the current goal of the system impacts the set of executable knowledge sources, as well as the cost/benefit analysis of each executable knowledge source. This is called the focus of attention (FOA). The FOA includes fine tuning the benefit value based on the knowledge source context. For example, rather than simply calculating the score using the benefit and the cost for each knowledge source, a third value which would represent any existing knowledge the system has in the knowledge database can be introduced to the equation for calculating the score.

0108As knowledge is generated by the system, various types of hypotheses are created by the system. The obvious goal of the system is to generate additional information that provides evidence to either support the hypotheses or prove the hypotheses false. As previously described, hypotheses are generated by mapping information from objects and tokens to models of fields. Therefore, if the goal of the system is to locate the courtesy amount field, the system should place a priority on the knowledge sources that generate information required by the courtesy amount field model. By adding these capabilities to the scheduler, the system will evolve from a bottom-up, data driven reasoning system to a top-down, model based reasoning system.

0109Hypotheses, the creation and updating of individual hypotheses are implemented as knowledge sources in the document analysis system. Since knowledge sources can be triggered based on the occurrence of one or more events, the knowledge sources that manage hypotheses can be triggered when new knowledge is created by any of the knowledge sources in the system. Specifically, since the system supports three types of hypotheses, field level, document composition, and document classification, each type of hypothesis has a corresponding knowledge source for creating and updating the corresponding types of hypotheses.

0110Field level hypotheses can be created, updated, or deleted as soon as any new information becomes available to the system. For instance, if the knowledge source for identifying machine print tokens enclosed within boxes executes and generates new information, i.e. new objects, then the new objects are candidates for possible hypothesis creation or update. If.the system has already created a hypothesis about the machine print field contained within the object, then the hypothesis must be updated based on the new information available. The new information can increase the confidence value of a hypothesis, or can result in a decrease in the confidence value for the hypothesis. If the decrease in the confidence value is great, the hypothesis can actually be discarded. Even if there is a hypothesis associated with the object, the system also attempts to create new hypotheses based on the new information. This is because the system must be capable of managing multiple hypotheses associated with a single object.

0111Document classification hypotheses are created based on very general document features, such as the percentage of hand print tokens in the image or the number of lines in the image. Since this information is generally available on system startup from the Net32K, these hypotheses can be made early in the processing of the image. Also, information generated during the processing of the image may have little impact on these hypotheses.

0112An example of how the system would locate the relevant fields of an unconstrained document is given below using Fig. 11, a business cheque. Tokens returned by Net32K are numbered and enclosed in boxes for ease of reference only. <ul id="ul0008" list-style="none"><li>1. MICR Location Knowledge Source - Looks at the Net32K tokens and rules out unlikely candidates (<i>i.e.</i> hand print tokens). Takes the likely tokens and compares them against MICR zone of the business document model file. Although three tokens were returned by Net32K (14 - 16), this knowledge source puts the tokens together and compares tokens 14-16 against the MICR zone in the business document file. These tokens are returned with the highest confidence value.</li><li>2. Simple Box Builder Knowledge Source - Looks for intersections of horizontal and vertical lines. Starts looking at any relationships between machine print or hand print enclosed in boxes. For example the box formed by lines 17-20 in Fig. 11 forms a box. Token 13 is partly enclosed within the box. This knowledge source therefore formulates information regarding the machine print token 13 positioned within a box.</li><li>4. Clustering Knowledge Source - Groups same type of tokens based on proximity. For example the tokens 7, 8, 9, 10, and 11 would be grouped into a cluster because of their proximity, as well as tokens 1, 2, 3, 4, and 5.</li><li>5. Horizontal Grouping Knowledge Source - Generates horizontal groupings of machine print tokens that are in close proximity. For example tokens 21 and 22 would be grouped together.</li><li>6. Vertical Grouping Knowledge Source - Generates vertical groupings of the horizontal groupings. Tokens 7, 8, 9, 10, and 11 would be a vertical cluster.</li><li>7. Hypotheses - Takes groupings and compares against the zones in the document model file for a business cheque and gets back confidence values.</li><li>8. Hand Print Grouping Knowledge Source - Horizontal grouping of hand print tokens. For example, if the signature token 12 was represented by two tokens, these tokens would be grouped together.</li><li>9. Hypotheses - Takes hand print groupings and compares against the zones in the document model file for a business cheque and gets back confidence values.</li><li>10. The groupings with the highest confidence values for the date, courtesy amount, signature, MICR and legal amount are output.</li><li>11. The groupings are sent for optical character recognition (OCR). If the character recognition engine cannot interpret the grouping sent then the grouping with the next highest confidence value is sent.</li></ul>

0113While the invention has been described in detail with specific reference to preferred embodiments thereof, it is understood that variations and modifications thereof may be made without departing from the spirit and scope of the invention.