JPH06501803A - Character recognition method that involves separating and extracting specific parts from extracted data - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] Character recognition method that involves separating and extracting specific parts from extracted data Background of the invention The invention generally relates to an improved method for automatically recognizing data on documents. With respect to the steps and methods, and more specifically, the transfer of goods such as checks, invoices and remittance documents. RELATING TO IMPROVED MEANS AND METHODS FOR AUTOMATICALLY RECOGNIZING Amount Information on Financial Documents Ru.

Today's financial services industry faces a major challenge in efficiently processing vast amounts of documents. confronting. Predictions that the paper payment method would decline have not materialized. actual , document payment methods are expected to continue to grow and increase worldwide.

Therefore, it is essential to devise improved methods for processing such documents. There is a need.

The use of imaging technology as an aid to document processing has been demonstrated, for example, in U.S. Pat. No. 4,205,780, No. 4,264.808, No. 4,672,186, No. 4,888,812, which significantly improves document processing. It has been recognized as a method. Imaging typically involves scanning documents optically and electronically. processed and for later retrieval and display (magnetic disk drive and/or creating an electronic image that is stored on a mass storage medium (such as optical memory) including. Document imaging offers an opportunity to reduce document handling and movement. It is clear that these electronic images cannot be used in place of actual documents. This is because it can happen. For example, a document image is retrieved from storage and where the operator can view the document without directly inspecting it. Amount data and other information can be entered based on the captured image.

Can the use of imaging in document processing systems offer significant improvements? The need for operators to inspect and enter data from documents can be accomplished by document Continue to limit processing speed and efficiency.

Summary and purpose of the invention In accordance with this invention, further increases in the speed and efficiency of document processing are achieved by improved methods for automatically locating, extracting and recognizing data, and more specifically Required for use in financial document processing systems such as checks, invoices, and remittance documents. by providing an improved method that can advantageously operate at the required high speeds. be made into

U.S. Pat. No. 4,449.239, 4. 201. No. 978, No. 4,468 ．． No. 808, No. 4,918.740, No. 4,523,330, No. 4,685 , 141, 3°832.682 and European Patent No. EP-0.111,93 No. 0 discloses various automatic data recognition techniques known in the art.

The purpose, features, advantages and uses as well as specific nature of this invention are clearly illustrated in the accompanying drawings. Additionally, it will become apparent from the detailed description below.

Brief description of the drawing FIG. 1 illustrates a typical pear check widely used in the United States.

FIG. 2 generally illustrates a document processing system in which the present invention may be incorporated. .

FIG. 3 shows an automatic courtesy in accordance with the present invention.

unt) Flowchart generally illustrating the various operational steps performed by the reader. -This is a chart.

FIG. 4 illustrates a preferred embodiment for accomplishing the "$location" step 102 of FIG. It is an illustrative flowchart.

FIG. 5 shows a typical gray level provided by image module 14 of FIG. This is a le image.

FIG. 6 is a seed and mask binary image from the gray level image represented in FIG. FIG. 2 is a block diagram illustrating an apparatus for parallel generation and storage of .

FIG. 7 shows a typical [$j seed server established by step l02B of FIG. Cheeria $sA is exemplified.

FIG. 8 illustrates a typical “$J mask service” established by step 102B of FIG. Cheeria $MA is exemplified.

FIG. 9 shows a preferred embodiment for accomplishing the "face amount extraction" step 104 of FIG. It is a flowchart which illustrates.

FIG. 1O shows a typical par value seed established by step 104B of FIG. Search areas C, A, and SA are illustrated.

FIG. 11 shows a typical denomination mask established by step 104B of FIG. Search areas C, A, and MA are illustrated.

FIG. 12 shows a preferred method for accomplishing the "separate and categorize φ parts" step 110 of FIG. 2 is a flowchart illustrating a preferred embodiment.

FIG. 13 shows the heterogeneous connection configuration elements (prior to disambiguation in step 104J of FIG. 9). A typical extracted face value including prime groups 62 and 63 is illustrated.

14-16 show a typical extraction after disambiguation in step 104J of FIG. The following is an example of the face value.

Figures 17-18 show how the ``ψ'' character is made up of an underlined two-digit number. An example of what is extracted from the field is shown below.

Figure 19-21 shows how the ``φ'' character is extracted from the ``Uφ'' field consisting of a fraction. Illustrate how it will be served.

detailed description Like numbers and letters refer to like elements throughout the drawings.

For purposes of this detailed description, the present invention describes a document processing method for processing financial documents. The system automatically determines the dollar amount (specifically referred to as the “face amount”) on a check. It will be illustrated as being applied to dynamic recognition. However, this The invention extends to other types of documents as well as other types of financial and other data recognition applications. It should be understood that also applicable.

Referring to Figure 1 recently, there is an alliance check IO widely used in the United States. or exemplified. The check IO has a "$" currency symbol 10a and an associated amount 10b. However, it is typically referred to in the financial industry as the "face value." Accept this face value Recognizing readers are typically referred to as denomination readers (CARs). Face value 10 b may be machine printed or handwritten as shown in FIG.

The typical check lO shown in Figure 1 has a machine readable code encoded in the lower left corner of the check. It also contains possible data IOc, which includes the identity of the bank on which the check is drawn, the customer Helps provide identifying information such as the customer's savings account number and check number.

Typically this encoded machine readable data 10c is provided in magnetic ink. It is called by the acronym MICRJ (Magnetic Ink Character Recognition).

FIG. 2 generally illustrates a document processing system in which the present invention may be incorporated. place The documents to be processed are typically financial documents, such as checks of the type illustrated in Figure 1. include. As shown in FIG. 2, these financial documents 10 are provided to a document processor 12. The machine reads the encoded data from the document and, in a conventional manner, reads the encoded data from the document. image and process the document into pockets (not shown) do.

The document processor 12 in FIG. 2 captures an image of a document and converts the captured document image into process and compress the compressed document image, then store the compressed document image on a disk drive, etc. It includes an imaging module 14 that transmits data to a storage device 16 . Work station 19 for display and input of data by the workstation operator; A document image, such as a denomination from an autopsy image, is received from storage 16.

A processing unit (CPU) 20 controls the entire system and (via a storage device 16) ) transmitted thereto by document processor 12 and workstation 19 Also provides database maintenance for document information.

The document processor 12 of FIG. 2 further includes the amount of face value on the check as illustrated in FIG. a denomination coupled to an imaging module 14 for automatically recognizing the denomination; It includes a reader 18. Such face value reading 1 in the document processing system of Figure 1 An important advantage of providing 118 is that the amount or a successfully read check is Their face value can be read by examining the images at And it does not need to be input.

A face value reader (CAR) 18 provides automatic face value recognition in accordance with the present invention. In order to perform arithmetic processing on the document image provided from the image module 14, In addition to appropriate programming, multiple microprocessors, RAM, Typically includes ROM and other associated circuitry. I have set something like that in CARI8. The aspects to which the invention relates will be apparent from the disclosure herein.

Figure 3 shows the steps performed by CARI8 in Figure 2 in recognizing the face amount on a check. 1 is a flowchart generally illustrating various operational steps involved. This flowch The charts are given as examples and should not be considered as limiting the scope of the invention. It's not a good idea to be criticized; it's a good thing to be understood. For example, shown here some steps may be omitted, others may be added, and /or the sequence of steps may be changed.

As indicated by step 100, the CAR 18 is connected to the imaging module of FIG. Receive the gray level image of the check from Rule 14. CAR is ' in Figure 1. Step I O2) to locate $J10a, then the related face value 10 b is extracted (step 104). Is the face value extracted or machine printed? Or, a determination is made as to whether it is handwritten (step 106). If machine stamp If it is a printing, a relatively simple face value recognition is performed (step +08) and The result is output (step 118).

If it is determined that the extracted face value is handwritten (step 106), then Requires complex analysis. In such a case, "φ" portion 10b-1 (FIG. 1) First separated and classified (step 110), the "φ" character or Then, it is extracted (step 112). The "φ" character extracted as a result is that (step 114).

``After the φj character is successfully recognized (step 114), the dollar character is recognized (step 114). Step ll6). CAR18 (Figure 2) then recognizes the face amount or rejection. A dead signal is output (step 118). In the system of Figure 2, this CAR output Power is sent to CPU 20. If a rejection condition is detected during any of the steps in Figure 3. Once issued, a reject output is immediately given and the remaining steps are halted. Shown in Figure 3 As shown in Figure 2, the extraction and recognition of the “φ” portion of the face value is performed before the dollar portion. is done, because it tends to produce rejection. By CAR The recognized face value output given by It will be appreciated that an indication-based confidence value may also be involved. In Figure 3 The recognition steps 106.108.114 and +16 are described in the aforementioned patents. It is also understood that the information may be provided using known recognition techniques such as those disclosed. will be done.

A description of each of the steps illustrated in FIG. 3 is provided below.

Step 100 (Figure 3) During this step, the imaging module 14 in FIG. the portion of the check containing at least the "$" character 10a and the associated face amount 10b; (as illustrated in FIG. 5). Illustrated in Figure 5 It is understood that the sizes shown are examples only.

Step 102 (Figure 3) During this step, the "$" character 10a (FIG. 5) is located. clearly, Currency characters other than ``$'', such as an asterisk ``*'' or another suitable symbol, may also be placed. Can be used as a determining character.

Step 104 (Figure 3) During this step, the denomination 10b (FIG. 5) is transferred to the previously located "$" character 1 0a as a positioning guide.

A judgment is made as to whether the text is handwritten or handwritten. If so If either is machine printed, operation continues to step 108. If that If the is handwritten, operation continues to step 110.

Step 108 (Figure 3) If it is determined that the face value or machine-printed It is based on an old mechanical printing pattern.

Step 110 (Figure 3) If it is determined to be face value or handwritten, a more complex analysis is required; It begins with the separation of the φ portion 10b-1 (Fig. 5) from the dollar portion 10b-2. circle. The separated [ψJ parts are then sorted.

Step 112 (Figure 3) During this step, the "ψ" character is extracted based on the classification made in step 110. Served.

Step 114 During this step, the extracted "ψ" characters are recognized.

Step 116 (Figure 3) During this step, the "$" character 10b-2 (Figure 5) of the face value is recognized and the face value Complete the recognition of amounts.

Step 118 (Figure 3) During this step, CARl 8 (to CPU 20 in Figure 2) recognizes the face value Or output a rejection signal. A reject signal is sent if a reject condition occurs in any of the previous steps. provided by the CAR and, in such case, further Later steps are stopped. The recognized face amount may also be accompanied by a confidence value.

The various steps shown in FIG. 3 are now discussed in detail and in FIG. A preferred embodiment for achieving 02 is according to the present invention by following the steps in FIG. Steps 1028 through 102H will be discussed next. step The purpose of 102 is to locate the “$” character on the check IO in FIG. This may be something to keep in mind.

Step 102A (Figure 4) During this step, thresholding is provided by image module 14 of FIG. is used to obtain multiple binary images from the gray-level image (in Figure 5) . The derivation of these binary images is that a gray level image consists of each pixel having multiple gray levels. electrically typical as an X-Y matrix of pixels that have one of the following values: It can be understood by noting that it is expressed in a typical manner. and For example, each pixel corresponds to a binary number O through 15, where 15 is black and O or white. There may be 16 gray level values represented by 4 hits. of the obtained binary image each converts a different one of these gray level values into a gray level image or a binary image. It is created by using it as a threshold when converting to . For example, if If a threshold of 8 is used to generate a particular binary image, then that binary image is Mark pixels whose gray level value is 8 or greater as black pixels. However, other pixels of the binary image are always white.

For the particular embodiment of the invention considered here, three binary images or three Obtained from a gray level image (Figure 5) using different thresholds high, medium and low. It will be done. The high-threshold binary image is called a “$” sheet image, and the medium-threshold binary image is the face value It will be referred to as a forehead sheet image, and the low threshold binary image will be referred to as a mask image.

As explained below, the "$" sheet image has the "$" character 10a (Figure 5) The denomination binary image is used to locate and extract the denomination lOb. and the mask image is used for both purposes.

To increase the recognition speed, as illustrated in FIG. The seed and mask images are received from the image module 14 of the computer. can be generated simultaneously by each of the converters 34, 36 and 38, resulting in a binary Images are stored in respective random access memories (RAMs) 44, 46 and 48. stored in a searchable manner.

Step 102B (Figure 4) During this step, the search area on the "$" seed and mask image is "$" Established for use in positioning characters. Figure 7 is for the seed image. “$j Seed image search area $SA example is shown in Figure 8 for the mask image. "$J Mask search area $MA example. Figures 7 and 8 show the sheet and mask search area $MA. The effect produced by using different thresholds to obtain The results are also illustrated. In this regard, the “$” mask search area $MA in Figure 8 is (due to the lower threshold value used) than the “$” sheet search area $SA in Figure 7. Note that it contains extraneous black pixels (noise).

For the purposes of the particular embodiment considered, the desired “$” seed server of FIG. It will be assumed that the chain area $SA is known. For example, the location is CPU 20 (Fig. 2) or on the check 10 (Fig. 1). It may also be obtained from reading the machine readable line 10c of . instead," It is also possible to search the entire image until the $ character is located.

Steps 102c, 102D and 102E (FIG. 4) During step 102C, FIG. The "$" seed search area $SA in 7 is run for "new" black pixels. be inspected. As explained below, the new J black pixel is It is also not accounted for in A$SA. be. Typically, a vertical column-by-column scan is used, because it This is because it is an advantage in locating the "$" character that is encountered before the character. Ru. If no new black pixels are found during scanning (step 102D ), the last vertical column of the search area $SA was scanned. A decision is made (step 102E). In such cases, rejection or occurrence be done. Rejection will occur again if the maximum time allotted for the recognition process ends or It is important to understand that it can occur. This is one of the small The recognition process for stamps or time-consuming processes that conflict with check processing speed requirements. This is done to prevent this from happening.

If during step 102E it is determined that the vertical scan is not completed. If so, the operation returns to step 102C and “scan $j seed search area $SA”. Continue searching for new black pixels in the direction.

Steps 102F, 102G and 102H (Figure 4) If new black pixels If found during step 102D, operation proceeds to step 102F. Ste ・During whelk 102F, “discovered in $j seed search area $SA (Figure 7) The seeded black pixels (step 102D) are added to the "$" mask search area $ MA (Figure 8) is used to generate a connected group of pixels, which is will be designated by the symbol CC. The manner in which CC is generated will be explained next. cormorant.

The "$" seed search area $SA in FIG. 7 is first referred to. “$” character 10a is It appears to be approximately complete, with various parts as shown in 10'a. with a break, while the number ``8'' in the adjacent face value 10b is a larger and wider cut. It has eyes 10'b. This means that the "$" character is significantly more advanced than the normal face value character. High contrast and higher quality printing! produced using a process This is to be expected. A relatively high threshold (as explained earlier)? “$” The “$” sword in Figure 7 is used to obtain the sheet image. The search area $SA consists of only a few widely spaced heterogeneities such as 32s. Note also that it contains black pixels.

Next, referring to the "$" mask search area $~IA in FIG. ) is obtained using a lower threshold. lower threshold For processing, the $ character 10a is complete or the adjacent face value 10b is It will be seen that the "8" also includes some breaks 10"b. In addition, the “$” mask search area $MA in FIG. 8, such as 32m, in FIG. There are significantly more extraneous black pixels than in the $j sheet search area $SA. Ru. In addition, the “$” mask search area $MA is a black bixe with a face value frame of 33. Including files.

Steps 102D and 102F in FIG. 4 locate the "$" character and ``$'' seed and mask search keys (in Figures 7 and 8, respectively) to recognize Use both rear $SA and $MA. More specifically, the new black vixel or is found in the “$” sheet search area $SA in FIG. 7 (step 10 2D), with the corresponding position in the “$” mask search area $MA in FIG. The pixel is located. For example, if 34s in Figure 7 is If it is a new black pixel discovered in the search area $SA (step 102D), the corresponding positioning in the “$” mask search area $MA in FIG. The defined black pixel 34m is located. This corresponds to each “$” symbol in Figure 6. Seeds and masks stored in code and mask RAMs 44 and 48 It can be realized using images, for example, corresponding seed and mask pictures. A cell may be configured to have a corresponding address.

The next operation performed in the execution of step 102F is the “$” mask search Grow the black pixel 34m (Figure 8) in the rear $MA and connect it to 34m. The goal is to generate a CC consisting of all black pixels. For example, this is shown in Figure 6. This may be accomplished using the mask RAM 48 of .

Starting from black pixel 34m (Figure 8) and immediately adjacent to black pixel 34m (Figure 8) Determination as to whether there is a black pixel at the address corresponding to the pixel location A decision is made.

Similar judgments are made for each newly judged black vixel, forming the CC. The process is repeated again until all connected black pixels have been identified. These identification The address of the black pixel generated from black pixel 34rn constitutes the CC generated from black pixel 34rn. Ru. The mask RAM 46 in FIG. may be used to store the address of the cell.

Further referring to FIGS. 7 and 8, if the "$" sheet search area $S in FIG. The new pixel found at A (step 102D) or the "$" character 10a If the black pixel is 34s, enter the "$" mask search area $MA in Figure 8. The resulting CC created by growing the corresponding black pixel 34m (Step 102F) is CC-1, which is the "$" character 10a. this is all the pixels of the “$” character in the “$” mask search area $MA in Figure 8. This is because the files are connected.

On the other hand, if the "$j" character does not exist and the "$" seed search area $SA (Figure 7) It is assumed that there was a new black pixel or pixel with the number "8" 36s discovered in , the growth of the corresponding black vixel 36m in Fig. 8 generates CC-2, which is Because of the cut 10'b, it would be believed that it is only the upper part of the "8".

Following the occurrence of CC in step 102F, operation proceeds to step 102G, where Then, the size, geometry and position of the generated CC, or the appropriate value of the “$” character. To make a relatively quick judgment as to whether the candidate is a candidate or not, or whether the candidate is rejected. , thereby avoiding relatively more time-consuming recognition processes.

Only when it is determined that it is a CC or "$" candidate (step 102G), the operation is Proceed to step 102H, where it is determined whether it is traditional character recognition, CC or "$" character. It is executed to determine whether An example is the aforementioned U.S. Patent No. 4.449.2. The classifier approach described in No. 39 may be used for recognition. If C If C is not determined to be an appropriate candidate for "$" in step 102G, Or if the CC is not recognized as "$" in step 102H, The operation returns to step 102C, and the “$” seed search area $SA in FIG. and continues scanning new black pixels. However, if the CC If it is recognized as a "$" character in 2H, the "$" character is located. So If , no further scanning is performed and operation continues to step 104 of FIG. , extract the face value.

As mentioned above in connection with step 102E, when the end of the scan is reached, If the "$" character is not recognized, a rejection will occur. If additional search At the end of the scan, if it is desired that the rear be scanned for the "$" character, Instead of producing a rejection, the operation goes back to step 102B of FIG. 4 and creates a new sheet. Establish a code and mask search area. This scanning of additional search areas is desired. It may continue as many times as required or until an interruption occurs.

From the previous description of step 102F, the "new" referenced in step 102C Black pixels are identified during step 102F by As a result of the length of time, it was not discovered earlier, because the black pin that was identified earlier This is because there is no need to grow Quseru. Such previously identified seat pins Various possibilities to prevent growth in the ``$'' mask search area There is a way. In the described embodiment, in step 102F of FIG. During mask growth, the corresponding pixel in the “$” mask search area $MA is Once identified, the “$nono” (stored in “$” sheet images RA~144 in FIG. 6) Is it possible to achieve this purpose by removing sheet pixels from the sheet image? I found it to be advantageous. Therefore, the previous step in step l02F Black vixels identified during growth are not seen in the scan of step 102C and are therefore reduces the time required to locate the "$" character. This savings is based on the sheet image or relatively little "noise" black as a result of the high threshold used in its derivation. This is in addition to the time saved for including pixels. Like that The use of high thresholds is possible.

The seed/mask growth technique described above for generating CC is shown in check 10 (Fig. 1) It is understood that there is an additional advantage in locating the "$" character 10a above. , because the "$" character usually has high quality and high contrast. Create a cut in the “$J mask search area $MA (Figure 8)”. That's because it doesn't get blurred out like that. In this way, as explained above (step 102G and 102H), submit each of the generated CCs for recognition. ``$'' compared to other marks or letters (such as the number ``8'' discussed earlier) Significantly increases the chance that characters will be recognized.

Many sheet/mask growth techniques for generating CC are within the scope of this invention. It can be further understood that it undergoes the transformation of For example, generate CC The definition of connectivity used to can be varied in various ways to address the recognition of particular types of characters in different environments.

For example, the definition of "connectivity" (such as vertical, horizontal and/or certain diagonals) etc.) may be modified to be restricted to one or more specific directions. connectivity Another possible variation in the definition of One (or more) pixel breaks occurring between the pixels may be allowed.

Detailed description of step 104 A preferred embodiment for accomplishing step 104 in FIG. 3 is step 1 in FIG. 04A to 104J will be discussed next. Purpose of step 104 Keep in mind that is to extract the face value 10b shown in Figure I. It will be.

Step l04A (Figure 9) During this step, the operation switches to extracting the face value 10b (Figure 5). , the face value is located at the "$" character 10 in step 102 (FIGS. 3 and 4). It has been determined based on successfully locating a. Description of step 104 As the process progresses, step 102 describes how to locate "$". The basic seed/mask method described can also be used for face value extraction or It will be clear that it can be used in different ways.

Step! 04B (Figure 9) During this step, the sheet and mask search area is set to step 102 in FIG. Established for extracting the face value based on determining the position of the “$” character in be done. Figure IO shows an example of face value sheet searcheria C, A, SA, while Figure 11 shows an example of some larger face value mask search areas C, A, MA. show. C and A in Figure 11.

The MA is sufficient to include the face value portion that may protrude below the face value frame 33. Please note that it is a reasonable size. In this respect, the face value ``No. 7 must not all be included in the face value search areas C, A, and SA in Figure 10.'' Also, "7" is placed in the larger face value mask search areas C, A, and MA in Figure 11. Note also that all will be extracted as a result of sheet/mask growth. I want to be noticed.

In the preferred embodiment described herein, the The same mask image (stored in RAM 48 in Figure 6) as above or used for amount extraction. The face value sheet image (stored in RAM 46 in FIG. 6) is a “$” statement. “$” seed image (in RAM 44) used to position the character used for amount extraction instead of . This is the "$" seed image threshold As previously explained, the high-contrast “$” character is It is chosen that the face value character has a larger range of contrast changes It is done because it is extremely inappropriate. Figure 1O is the face value sheet search area An example of possible choices of thresholds for C, A, and SA is shown in which box 33 (Fig. 5 ) do not appear, as do low-contrast extraneous pixels (noise). In this regard , all parts of the face value are included in face value search areas C, A, and SA in Figure 10. It is understood that there is no need to do so. In C, A, and MA of Figure 11 to provide proper extraction of face value as a result of seed/mask growth , only a sufficient portion of the face value is required to be included in C, A, and SA in Figure 10. Ru.

Steps 104C, 104D, 104E and 104F (Figure 9) These steps occur after the end of the scan (step 10) for normal denominations. 2E) No rejection, the operation proceeds to step 106 (Figure 3) instead. Steps 102C, 102D, 102E, and 10 of FIG. It may be generally the same as that described above for 2F. Step 10 4C, 104D, 104E and 104F will therefore not be considered in detail. cormorant.

F new during scanning of face value sheet searcheria C, A, SA (Figure 10) Every time a black pixel is discovered, face value mask search areas C, A, MA Growth in (Figure II) creates a CC (as defined earlier).

Step l04G (Figure 9) Similar to step 102G in FIG. 4, this step occurs in step 104F. whether the CC is appropriate based on the size, geometry, and location of the CC; or test. For the purpose of face value extraction, this step in step 104G The test determines whether the CC generated is likely to be part of the face value. Ta For example, to determine whether the CC generated is likely to be part of the face value. A useful basis for The purpose of this is to determine whether or not it extends to the frame 52 (Fig. 11) of Ru. Such generated CCs are rarely part of the face value.

Step 104H (Figure 9) If the CC is generated, it seems that it is not part of the face value in step 104G. If so, the operation proceeds to step l04H and discards the generated CC. , the operation returns to step 104C, and the face value sheet search areas C and A in FIG. , the new black pixels in SA are scanned and the continuation is Ste! 1041 (Figure 9) If the CC generated appears to be part of the face value in step 104G, If so, operation proceeds to step 1041 where the generated CC ( (eg, in RAM memory 46 in FIG. 6) for later use. motion then returns to step 104C and searches the face value sheet search area C, A. Continue scanning new black pixels in SA.

Before leaving step 104r, the generated CC or (step 102 in Figures 3 and 4) and the method used for face value extraction and recognition. It may be helpful to note the differences between the methods used for " To locate the "$" character, each of the CCs generated is for recognition purposes. You may have it in mind that the character "$" is considered to be of high quality. This is because they are provided with high quality printing and usually have a high degree of contrast and do not have any breaks. . However, for some reason, the CC generated to extract the face value is a fragmentary part of the characters. It may not be too much, because the face value text, especially when written by hand, This is because there may be some kind of break. In this way, there are multiple face value characters. may consist of generated CCs. Therefore, extract and recognize the face value When locating the "$" character (in step 102H of FIG. 4) No attempt is made to recognize the generated CCs as is done. substitute Each CC determined to be likely to be part of the face value is added to the overall face value area. CC is stored until it is scanned (step 104I), and the CC is stored until it is scanned. It will probably happen. These stored CCs constitute the extracted face value. Ru. Figure 13 shows such stored par amounts extracted as explained above. This is an example. In this way, regarding the face value, the sheet/mass for generating CC The growth method is mainly used as a particularly advantageous method of extracting and storing face value for recognition. Dosage-wise step 104J (tests end of scan) step 104E The face value, seed search area C, A, SA (Fig. 10) can be scanned or completed through After the CC is generated and stored, it seems that it is part of the face value. be reached. The purpose of step 104J is to extract this stored extracted face amount ( 13) by removing extraneous CCs as exemplified by 62 and 63. It is important to make the information clear. One technique used is to use a heterogeneous CC such as 62 If the top and bottom of the face value area are separated by a predetermined amount The solution is to delete them. This may have a top and bottom border, for example. achieved by horizontally projecting the entire amount field to define the area may be made. above or below these boundaries, such as 62 in Figure 13. The CC is then deleted. If the projection yields multiple regions, use the appropriate The upper and lower borders correspond to the area containing the "$" character.

Removal of extraneous CCs such as 63 in Figure 3 located to the right of the face value is more complicated. pose a problem, because they may be part of the face value It is. A particularly advantageous way to determine whether these CCs are heterogeneous is to The horizontal distance between the right CC and the nearest black pixel to the left of the CC or the horizontal width of the CC Based on the condition that it is greater by a predetermined amount. If this condition is met For example, the CC is considered foreign and is deleted. How can this condition be realized? An example of this will be explained with respect to FIG. For the purpose of this example, we will use the face value area The area is divided into columns numbered from left to right, each column having a width equal to one pixel. It is thought that then. First, the positioning of the following columns is determined from the extracted face value be done, C1 = rightmost column with black pixels.

C2=rightmost column of area next to column 0 with minimum width W to the left of CI.

C3 = the column to the left of C2 containing black pixels.

If the following conditions are met, (C2-C3)>K (CI-C2) The remaining black pixel element 63 placed between CI and C3 is deleted. typical Generally, W may have a width corresponding to the width of three columns, and W may have a width corresponding to the width of three columns, and W may be Therefore, the above conditions are selected so that they are not satisfied. K typically has a value of 1.5 may exist. These values of W and K may be mistaken for face value or extraneous CC. selected to ensure that no The above is true as long as the conditions continue to match. repeated repeatedly. When the condition fails to match, the test ends and the operation proceeds as shown in Figure 3. Proceed to step 106.

Detailed explanation of step 110 (FIG. 12) Step 110 is performed if step 104 If it is determined that the face value extracted during This can be understood from Figure 3. The purpose of step 110 is to 1 (FIG. 1) from the torque portion job-2 of face value 10b. vinegar A preferred embodiment for accomplishing step 110 in accordance with the present invention is as shown in FIG. It will be discussed next with reference to steps llOA to llOH. of the face value The 87 part and the "φ" part are referred to as the "$" field and the "φ" field below. They will be called respectively.

Step ll0A (Figure 12) During step ll0A, the extracted face amount is Existence is searched. Such a period or decimal point is of course It shows the separation between the ``$'' and ``φ'' fields, and the detection of these used as a basis for separating fields.

A preferred method for detecting the presence of a period is described with respect to FIG. Dew. For this purpose, the extracted face value is stored in the RAM memory 46 of FIG. from left to right, such as by using column-wise scanning of the image of the extracted denominations deposited. will be investigated.

If a potential period candidate like PC in Figure 13 is discovered, the upper line UL and the lower line LL (Figure 14) or right next to the PC (like the number "9" in Figure 14) Determined for the face value portion on the left. Lines are numbered from top to bottom. potential A target period candidate PC is considered an actual period candidate if the following conditions are met. Conceivable.

(1) The height of potential period candidate PC is greater than 1/2 (UL-LL) do not have.

(2) Potential period candidate PC has less congestion W than a predetermined amount.

(3) The average number of lines of potential period candidate PCs is less than I/2 (UL + LL). do not have.

Typically, up to three period candidates can be identified based on the measurements above. It is. Operation then proceeds to step llOB of [N12.

Step 1 During IOB, up to three pyri determined in step 110A Odo candidates are known statistical classifications such as those disclosed in the aforementioned patents. be investigated using technology. If one or more period candidates are acceptable Once you know, the rightmost one is the one that separates the "$" and "φ" fields. 12, and operation proceeds to step 11OG of FIG.

However, if no period is identified, the operation proceeds to step ll0C. Proceed to step 3 and try to separate the "$" and "φ" fields on another basis.

Step ll0C (Figure 12) During step 110C, as shown in FIG. The subscripting is the “$” and “φ” fee of the face value. will be investigated as a basis for the separation of For this purpose, the extracted face value The amounts are again examined from left to right (for example, the number “5” in Figure 14). ) Detect the start of the superscripted character SC.

Similar to step 110B, which describes the search for potential period candidates, the upper line UL and the lower line LL (Figure 15) is the superscript of the candidate (like the number "7" in Figure 15). Determined for the face value portion immediately to the left of the written letters SC. The line is from above again numbered downwards. The character SC written in the superscript of the candidate is if the following conditions are met. If it is, it is considered to be an actual superscript.

(1) The line number at the bottom of the face value letters written in the candidate superscript is 1/3 (2LL +UL).

(2) The height of the character SC written in the superscript of the candidate is at least l/3 (LL- UL).

(3) The letters SC written in the candidate superscript (like the number “7” in Figure 15) are all separated from the face value portion to the left by at least one white column. Column 0 is a black pixel This is a column that does not have

(4) Face value portion SC immediately to the right of the face value letters SC written in the candidate superscript ’ (number “0” in Figure 14) is greater than the bottom line number of the SC plus half of its height. has a bottom line number that is not clear.

The first candidate that satisfies the above conditions is the start of the "φ" field written in superscript. It is believed that there is.

Operation then proceeds to step ll0D of FIG.

Step ll0D (Figure 12) Step ll0D is the “ψ” file written in the superscript executed in step ll0C. Receive the results of a field search.

If the superscript character is found, the "$" and "φ" fields The basis for separation is determined and operation proceeds to step ll0G.

However, if the superscripted ``ψ'' field is not found, then the The process proceeds to step ll0E in Figure 12, where the separation of the "$" and "φ" fields is performed. Find another basis for.

Step ll0E (Figure 12) During step ll0E, the composite "φ" field is The existence of the field has been investigated as a basis for the separation of the “$” and “ψ” fields. It will be done. Figure 16 shows the compound “ψnov” consisting of two underlined numbers written in superscripts. It can be seen that this indicates the first (seventh) alliance of the field. Figure 17 is “ψ” Indicates the second hardness of a composite "ψ" field given as an amount or a fraction.

The face value extracted to determine whether the composite "φ" field exists or not. The forehead is read from right to left as in steps 110A and 110C previously described. It will be investigated every time. When searching for a compound "φ" field, it is determined whether: (See Figures 16 and 17).

(1) Last occupied column C1゜ of extracted face value (2) Last occupied column C1゜ First 0 column to the left of column C1 C2゜ (3) Face value section located to the right of 0 column C2 The first line Ll occupied by minutes.

As shown in Figures 16 and 17, the values of CI, C2 and Ll are extracted for testing whether it is a good candidate for the 'φ' field. Determine the limits for a specific portion of the face value. Testing various possible "φ" fields field type and, in particular, the compound “ψ” field type shown in Figures 16 and 17. This is achieved using statistical classification techniques specifically designed to identify the The manner in which such statistical classification techniques may be implemented will be apparent from the above-mentioned patents.

If the compound "φ" field is recognized as shown in Figures 16 and 17, Then, column C2 is considered to be a separating column between the “$” and “φ” fields. Ru. Keep in mind that C2 is the first 0 column to the left of the last occupied column CI. It will be placed.

The result of the operation in step llOH is then sent to step ll0F in FIG. It will be done.

Step ll0F (Figure 12) Step ll0F is a composite "φ" field executed in step ll0E Receive the results of your search. If a compound "ψ" field is found, column C2 serves as the basis for the separation of the “$” and “φ” fields, and the operation is as shown in Figure 12. Proceed to step ll0G.

However, if the composite ``ψ'' field is not found, rejection will occur. , because the basis for separating the “$” and “φ” fields of the face value is because it cannot be found and there is no further basis for separation. In this regard , it is understood that further grounds for providing separation may also be provided. It is a place to be treated.

From the description above in Figure 12, find the period (steps 110A and ll0 B) Find the "ψ" field written in superscript (steps ll0c and and 110D) or find a composite “φ” field (steps ll0E and "φ" and "$" field of face value based on either As a result of finding a basis for separating the code, operation proceeds to step 110G. You will understand what I mean.

Therefore step ll0G is the specific basis (period) found for separation. , written in superscript or using a compound “φ” field) Separate the cables.

Also during step ll0G, the separated "φ" field is classified as one of several types using statistical classification techniques such as those disclosed. Ru. Given by step ll0G for the preferred embodiment described The categories obtained are two 0.2 digit numbers, an underlined two digit number, and a fraction. another A classification may also be given.

If an acceptable classification is determined in step 110G, the operation proceeds as shown in FIG. 2, otherwise a rejection occurs.

Step 112 (brief description of the drawing) A preferred embodiment for accomplishing step 112 in accordance with the present invention will now be discussed. There will be. The purpose of step 112 is to determine the classification determined for the "φ" field. It is kept in mind that the purpose is to extract the "φ) character based on .

Steps for the various categories given in the described preferred embodiment The operation at 112 is explained below. For this classification, the value of the ψJ field is 0. It is immediately known that Proceed to step 116. Typically, this classification is based on the basis of separation or period or It is only used where the detection of the superscript ``rφ'' field.

2 digit number Regarding this classification, [φ nofield numbers are directly available and therefore dynamic. The process proceeds to step 114 in FIG. 3 for recognition without further processing. " Similar to the ``Two Zero'' category, this category typically uses a period or It is only used where the detection of the ``φ'' field written in superscript or superscript.

two-digit number underlined For this classification, the ``φ'' field consists of an underlined ``36'' in Figure 1. 8 and 19, the operation removes the underline so that only the “φ” character remains. directed towards leaving. Preferred specific examples for achieving this objective are described below. It will be revealed.

First, the slope of the underline is determined as follows. For each column of the "φ" field, The number of white pixels relative to the first black pixel is counted from the bottom edge. if If the difference between these numbers in two consecutive columns is greater than 4 in terms of quantity, then the discontinuous Part exists. All parts of the discontinuity in the “φ” field and the two The largest column range between discrete sections is determined. This largest column range Underlined or underlined is also expected. The coordinate points (xi, yl) and (x2 ．． y2) is defined as follows.

xl = start column of column range.

YI=number of white pixels from the bottom edge to the first black pixel in column xl.

X2=end column of column range.

y2=number of white pixels from the bottom edge to the first black pixel in column x2.

The underlined slope SL is determined by the following equation.

SL=(y2-y1)/(x2-xi) to remove the found underline , a set of n straight lines of gradient SL and a vertical spacing of 1 are formed. The number of straight lines n is underlined is established as follows.

Other times n=32 Furthermore, at the starting point or above these straight lines, a scanning step width of 1 is applied along these straight lines. is established to scan the reJ field from the right and from the left.

For scanning from the left, XI(i)==rφ” field first column (for all straight lines) yl (i) = y start + 1-1 (for the i-th straight line) In this case, the y-start is ensured to occur at (xi yl) or the first straight line scan point. be erected.

For scanning from the right, xr(i) = last column of rφj field (with respect to all straight lines) yr (i) = y start + 1-1 (for the i-th straight line) In this case, the y-start occurs at (x2.y2) or the first straight line scan point. established.

The “φ” field is scanned along these straight lines and its purpose is to Below that straight line, either a complete underline or a straight line that no useful information occurs. This is to determine the line. For this purpose, the number of scanning points is As far as the scan point with the first black pixel of the card is concerned, the scans from the right and from the left are It is counted for every scan line in the scan. Then, scanning from the right Maximum of counted scanning points in the course and in the scanning course from the left A straight line with the number of pieces is determined. From this range of straight lines, the lowest one is selected. Ru. All parts of the extracted "φ" field below this lowest straight line are removed and produces the results shown in FIG. The above steps block the "ψ" character or the underscore. It also deals with situations where

After removing the underline (Figure 19), the remaining "φ" field component (r36 in Figure 19) ) is again ascertained using statistical classification techniques, with two zero or two-digit digit classifications. It is decided which one to enter. If the classification is 2 digits 0, the operation is step 11 Proceed to step 6, because the value of the ``φ'' field is known to be O. Ru. If the classification is a 2-digit number, the operation steps to recognize the 2-digit number. Proceed to l14. If neither category is found, a rejection occurs.

fraction For this classification, the operation begins with the denominator as shown in Figures 20.21 and 22. and then the fraction point, because the "φ" field is the numerator "80 ” and a fraction with a denominator r of 100J. achieve this purpose Preferred embodiments for this purpose are described below.

First, the field is examined to the fractional line or to the expected column range. This is For example, by determining the connected component group CC with the largest width. You may be

Once a fraction line is discovered, its slope is fl'! Found xLyl, x2, y2 and the underlined compound "φ" field described above. is determined by calculating the slope in the same manner as .

A dividing line is established here, where only the numerator and fractional lines are located as far as possible. child The dividing straight line is determined by the slope and the coordinates rxl, yl + offset with the following: It will be done.

Other times Offset = lO Establishing a dividing line in this way, the “φ” field component under this dividing line is For the fraction example shown in Figure 19, the denominator rloOJ is deleted. will be the result. In this way, the remaining "φJ field components are shown in FIG. It will be the underlined "80" shown in .

Therefore, as illustrated in Figure 21, the fractional behavior has been underlined so far. It should result in two zeros or two digit numbers, leaving the remaining "φ" field components are ascertained using statistical classification techniques, and these remaining components are Does it actually correspond to the two underlined O's or the two underlined numbers? decide whether If so, the behavior is the underlined compound "φ" #! As explained above regarding the code classification #! Next, figure out how to extract the FφJ character. 22), if not, a rejection will occur.

This invention has been described herein with respect to specific preferred embodiments, specific examples, arrangements, etc. Many modifications and variations in sequence and use are possible within the scope of this invention. That's a good thing to be understood. For example, in addition to the size and number of classification categories, The seed and mask images used as well as the number of search areas and pears also change. You may. Additionally, seed and mask images are limited to binary (two-level) images. There is no need to be Face value extraction, “ψfield separation and/or recognition” For example, if the mask image itself is It may be a gray level (multi-level) image. Additionally, additional processing steps are required. Additional features may be added to give features, or steps described may also be removed. or rearranged.

Additionally, the invention may be adapted to a wide variety of adaptations other than those described herein. . Accordingly, the following claims are intended to cover all possible developments within the scope set forth therein. It is to be considered as including all modifications and variations.

FI port, 1 10 Work Nutage Con 19 Suna, ArooctB3Eρ1 et al. Stay, 7’102 units 9 Store (Obpira FIG, +3 FIG, +7 FIG 18 FIG 19 FIG 20 FIG, 21 FIG, 22 International Search Report international search report Continuation of front page (81) Designated countries EP (AT, BE, CH, DE.

DK, ES, FR, GB, GR, IT, LU, NL, SE), A U, CA, JP (72) Inventor: Krenner, Norbert, Federal Republic of Germany, Day 7752 Reichenala, Steadygase, 18.A.

Claims (23)

[Claims] 1. 1. An electronically implemented method for recognizing characters of financial amounts on documents, comprising: , said financial amount has an integer part and a decimal part with possible multiple fractional part representations. creating an image representation corresponding to the financial amount; at least one predetermined plurality of distinguishing parts for the integer and decimal parts; examining the image representation to find separation characteristics; the previous step based on finding one of the discriminating properties during the examining step; separating the integer and decimal parts; Determine which of the possible fractional part representations the separated decimal part is. The step of classifying into one of several categories based on whether the decimal part corresponds to and a step for extracting separated decimal characters based on the step of classifying. A method including steps. 2. 2. The method of claim 1, wherein the image representation is a stored digital representation. 3. 3. The method of claim 2, wherein the digital display is a binary display. 4. If the plurality of predetermined distinguishing characteristics are present during the examining step, Claim further comprising rejecting further recognition operations if not found. The method described in 1. 5. The step of creating locates a foreign image portion that is not part of the financial amount. examining said image display to determine the location of said foreign image portion; and removing based on their position with respect to recorded financial amounts. The method described in. 6. The plurality of predetermined distinguishing characteristics are the integer and fractional numbers of the financial amount. including the presence of a period between several parts, and the step of examining includes the presence of a period between parts. 2. The method of claim 1, wherein the method is determined based on its size and location. 7. The plurality of predetermined identification separation characteristics are superscripted in the decimal part. the presence of superscripted characters, and the step of examining includes the presence of superscripted characters. 2. The method of claim 1, wherein the method is determined based on its size and location. 8. the plurality of distinguishing characteristics include the presence of a complex character sequence in the decimal portion; , the step of examining the presence of the compound character sequence using statistical classification techniques. 2. The method of claim 1, wherein the method determines. 9. 9. The method of claim 8, wherein the composite character array includes underlined characters. 10. 9. The method of claim 8, wherein the composite character array includes fractions. 11. 5. The integer part corresponds to dollars and the decimal part corresponds to cents. The method described in 1. 12. 12. The method of claim 11, wherein the plurality of classifications includes two zeros. 13. 12. The method of claim 11, wherein the plurality of categories includes a two-digit number. 14. The plurality of categories are underlined by two adjacent numbers with an underline below. 12. The method of claim 11, comprising an inverted two-digit number. 15. The plurality of classifications include fractions including a numerator, a denominator, and a fraction line between them. 12. The method of claim 11. 16. The above multiple categories include a two-digit number, an underlined two-digit number, a numerator, a denominator, and and a fraction including a fraction line therebetween. 17. The extracting step includes determining the position and slope of the underline. 17. A method according to claim 14 or 16, comprising the step of removing underlines. 18. The extracting step is a step of determining the position and slope of the fraction line. and the step of deleting the denominator and fraction line of the fraction carryover. or the method described in 16. 19. A method for recognizing predetermined monetary data on a document, the method comprising: creating an image display corresponding to the predetermined amount data; , the image display may display foreign data that is not part of the predetermined amount data. and examining the image display to locate the foreign image portion. and The foreign image portions are based on their positions with respect to the predetermined amount data. and removing. 20. The predetermined data consists of adjacent characters with top and bottom borders , and the removing step is performed if the foreign image portion is above the border line and a predetermined amount, spaced from the bottom and removing steps; The method according to item 5 or 19. 21. The boundary line is formed by projecting the group of characters to create an area surrounding the characters. 21. The method of claim 20, wherein: 22. The financial amounts are arranged in rows, and the foreign image portions are adjacent to the sides of the rows. one or more extraneous sub-image portions located in the step is such that the distance between the closest part of the financial amount and the foreign sub-image part is the sub-image part. Extraneous sub-images based on the condition that they are a predetermined amount greater than the width of the image portion 20. A method according to claim 5 or 19, comprising the step of deleting image portions. 23. The deleting step is repeated until the condition is continuously satisfied. 23. The method of claim 22, wherein the method is repeated.