CA2094706A1 - Handwritten digit recognition apparatus and method - Google Patents

Abstract

A computer system and a method for a mail sorting operation in which the computer system determines the location of the ZIP code within a digital image of an address block (20) from a piece of mail. An interstroke distance is calculated for the image and the strokes of the image are thinned (38) to enhance vertical separation between the lines of the address block. A medial axis for each line is determined and the medial axis is superimposed upon the digital image. A bleeding operation is conducted on the digital image from the medial axis at which data bits that do not connect to the medial axis are notated as punctuation and interlinear connected strokes are then divided between the two lines. The last line which is determined to be large enough to contain a ZIP code based on bounding box size is then selected. Alternate splits of words are formed and the best split is selected (50) in which the last formed group is detected to be the ZIP code.

Description

2 ~ 9 ~ 7 Q ~ PCl`/US91/03624 HANDWRITTEN DIGIT RECOGNmON APPARA~US & METHOD

Technical Field of the Invention The technical field of the present invention relates to optical character recognition and more 5 particularly recognition of handwritten digits.

Bac~ound o~ the InvQntion TherQ are many instances where it would be useful or desirable to provide a computer readable form of a document not available in a compatible computer readable 10 form. Normally it is the case that the document is not available in machine readable form because the document was handwritten or typewritten and thus no computer readable form exists, or because the computer readable form is not available. In some instances there is a "foreign"
15 document, i.e. an existing computer readable form but the document was produced on an incompatible computer system.
In some instances, such as facsimile transmission, a simple optical scan of the document can produce the required form.
In most instances the form most useful for later use and 20 decision making is a separate indication of each character of the document.
The field of optical character recognition deals with the problem of separating and indicating printed or written characters. In optical character recognition, the 25 document is scanned in some fashion to produce a electrical image of the marks of the document. This image of the 2 ~

marks is analyzed by computer to produce an indication of each character of the document. It is within the current state of the art to produce relatively error free indication of many typewritten and printed documents. The 5 best systems of the prior art are capable of properly distinguishing a number of differing typa fonts~
On the other hand, unconstrained handwritten characters have not been succassfully located and recogni2ed by present optical systems~ The problem of 10 properly reading unconstrained handwritten characters is difficult because of the great variability of the characters. One person may not write the same character exactly the same every time. The variability between different persons writing the same character is even 15 greater than the variability of a single person. In addition to the variability of the characters themselves, handwritten text is often not cleanly executed. Thus characters may overlap horizontally. Loops and descenders may overlap vertically. Two characters may be connected 20 together, strokes of one character may be disconnected from other strokes of the same character. Further, the individual written lines may be on a slant or have an irregular profile. The different parts of the handwriting may also differ in size. Thus recognition of handwritten 25 characters is a difficult task.
An example of a field where recognition of handwritten characters would be very valuable is in mail sorting. Each piece of mail must be classified by destination address. Currently, a large volume of ~W092/08203 2 ~ 3 ~ 7 0 ~ PCT/US91/03624 typewritten and printed mail is read and sorted using prior art optical character recognition techniques. Presently, approximately 15% of current U.S~ mail that is hand addressed. Present technology uses automated conveyer 5 systems to present these pieces of mail, one at a time, to an operator who ViQWS the address and enters a code for the destination. This is the most labor intensive, slowest and consequently most expensi~e part of the entire mail sorting operation.
Furthermore, it is expensive to misidentify a ZIP
code and send the piece of mail to the wrong post of~ice.
Once the mail is forwarded to the receiving post office, the receiving post office recognizes that there is no matching address or addressee in that ZIP code. The mail 15 must then be resorted and redirected to the proper post office. Because of the high expense associated with misdirected mail, it is more desirable to have an automated system reject a piece of mail if the system cannot determine the ZIP code with an extremely high degree of 20 accuracy. The rejected pieces of mail can then be hand sorted at the sending station or other measures can be taken to eliminate or reduce the cost of the misdelivery.
Sorting of handwritten mail is an area having a unique set of characteristics. First, due to the problem 25 of user acceptance it is not feasible to place further constraints on the address. Thus address lines or individual character boxes, which would be useful in regularizing the recognition task, are ruled out. on the other hand, therè already exists a relatively constrained 2 Q Q'~ ;3 '~
W092t08203 PCT/US91/03624 portion of the current address. The ZIP code employed in a majority of handwritten destination addresses provides all the information needed for the primary sorting operation~ Most handwritten ZIP codes are formatted with 5 5 digits while some handwritten ZIP codes use the longer 9 digit ZIP code. This information is relativel~ constrained because the 2IP code consists o~ only 5 or 9 digits~ In addition thQ ZIP code is usua~ly located at the end of the last line of the destination address or sometimes is by l0 itself on the last line.
Various systems have been devised to recognize handwritten digits. However, many of these systems assume that the digits are already located and isolated and the problem is only to determine which numeral the handwritten 15 digit represents. Often these systems require the digits to be written inside individual boxes.
In order for a computer to analyze and recognize the handwritten numerals in a hand-written ZIP code in an address block typically appearing on an envelope, the group 20 of numerals comprising the ZIP code must first be successfully located as a group.
Even though the above mentioned constraints on the ZIP code in the form of number of digits and location exist, previous attempts to locate the ZIP code have 25 encountered problems. The same problems that exist in general for successful recognition of handwriting also pose problems for locating the ZIP code. Previous attempts to count lines of a handwritten address block have been W092/08203 ~ Q . ~ 7 0 ~ PCT/US91/03624 stymied by loops, descenders, line slant or other line irregularities.
What is needed is a highly reliable system to correctly locate the ZIP code in an address block before 5 analysis of the digits of the ZIP code.

Sum~y of the I~venti~n In accordance with one aspect of the invention, a computer system is designed to locate a 2IP code from an digitized address block. The address block can be derived lO ~rom an addressed envelope, postcard or other label through optical scanning. The digitized address block is comprised of pixels arranged in a matrix. Preferably the digital address block is binary with pixels either being part of the foreground image or part of the background.
The computer system includes a mechanism for computing horizontal distances between sequential but separated pixels of the foreground image that are in the same row. The computer system subsequently compiles and determines a first significant peak in occurrences of 20 distances which is designated as the interstroke distance.
In the foreseen application, the foreground image represents character strokes of the address block. The strokes are arranged into words based on the interstroke distance.
The words are then formatted into groupings, i.e.
blocks. The interline vertical connections between different lines of the address block are broken via horizontal erosion of the strokes. The word blocks are r~

3 PCI`/US91/03624 then skeletonized down to a horizontal skeleton. A
subsystem dilates the resulting horizontal skeleton vertically into boxes with each box having a uniform height and then dilates the boxes horizontally such that boxes 5 overlapping in the horizontal direction are merged together to form line images. The line images are labeled (i.e., numbered~ uni~uely from the top of the image, to produce line numbers. Another subsystem then determines each line imaye's medial axis and superimposes the line-numbered 10 medial axis onto the original digiti2ed address block.
Desirably, the computer system then bleeds the line number label from each medial axis to identify all strokes connected to the medial axis. Strokes that are connected to two horizontal axes are divided to either the 15 line above or the line below. ~he mechanism identifies foreground image pixels not connected to any medial axis and excludes these pixels from a subsequent line count.
The desired last line that is large enough to contain a ZIP
code is then selected and possible wording splits of the 20 last line are determined from interstroke distances and the identified foreground imaye pixels that do not touch any medial axis~ One wording split is selected and a word from the split is identified as the ZIP code location.
Preferably a mechanism for creating a bounding ~5 box of the digitized address block is provided and operations are directed to only pixels within the bounding box to reduce the computer operating time. Furthermore, the pixels within the bounding box are down sampled further W092/08203 2 ~ ~ ~ 7 Q ~ PCT/US9t/03624 reducing computer time while still rendering the needed calculations and processing.
Preferably the computer system incorporates a parallel processing system. Computational time is hence 5 reduced to acceptable levels while the expense of a sufficiently powerful general computer is avoided.
In accordance with a broader aspect of the invention, the invention relates to a computer system and m~thod for locating a predetermined group o~ pixels within 10 a larger selection of pixels forming character images derived from handwriting or machine printing characters.
The computer system calculates horizontal distances between separated foreground image pixels in the same row to determine a first peak of distance lengths that is labeled 15 the interstroke distance~ The computer system separates the address block image into separate line images using the interstroke distances to form blocks, erodes the blocks horizontally to break interline strokes, skeletonizes the blocks, and subsequently dilates the skeletonized blocks to 20 form lines of the handwritten image. A group of pixels in a particular line is then selected by use of the interstroke distances and identified foreground pixels that do not have a connection to any medial axis of any respective line.

25 Brief Description of the Drawings Reference now is made to the accompanying drawings in which:

W092/08203 2 ~ 3 ~ 7 0 ~ PCT/US9~/03624 ^~

! ` - 8 -FIGURE 1 illustrates, in flow chart form, the process of ZIP code location employed in the preferred embodiment of the present invention;
FIGURE 2 schematically illustrates a sample 5 address block and its bounding box;
FIGURE 3 is an enlargad viet~ of the address block in the bounding box schematically illustrating the horizontal distances between two sequantial but separated pixels o~ the ~oreground image:
FIGURE 4 schematically illustrates t~e arranging o~ strokes into word blocks based upon the interstroke distance shown in FIGURE 3;
FIGURE 5 schematically shows the breaking of the descenders and vertical connections between the lines and 15 the formation of convex hulls:
FIGURE 6 schematically illustrates skeletonization of the convex hulls;
FIGURE 7 illustrates the vertically dilated word box areas dilated horizontally into lines and the formed 20 medial axis of each line;
FIGURE 8 illustrates superimposing the medial axis of each line shown in FIGURE 7 onto the original bounded down-sampled address block shown in FIGURE 2 and bleeding the characters from the medial axes;
FIGURE 9 illustrates a selection of the last line as determined by the previous steps;
FIGURE 10 discloses one selected splitting of words;

~V092/08203 2 ~ 9 ~ 7 0 ~ PCT/US91/03624 FIGURE 11 discloses an alternative selected splitting of words;
FIGURE 12 discloses a preferred selected word chosen for the ZIP code;
FIGURE 13 discloses a 2IP code correlating to the word being chosen in FIGURE 12;
FIGURE 14 discloses in block diagram form the morphological computer employed in the preferred embodiment of the present invention; and FIGUR~ 15 illustrates in block diagram form one of the neighborhood processing stages of the morphological computer illustrated in FIGURE 14.

Detailed ~escription of the Preferred Embodiment A ZIP code location system of the present lS invention implementing the method illustrated in Figure 1 is capable of locating a ZIP code within an address block.
Figures 2-13 provide an exemplary illustration of the steps diagrammed in Figure 1 and reference will be made to this example as the description of Figure 1 proceeds.
The handwritten fictional sample shown in figure 2 illustxates several problems that may occur in a handwritten address. The handwritten sample is only an example and it should be understood that the same below descxibed method is equally applicable to other 25 handwritings that can be quite distinctive and very different from the one shown.
The shown example has several common problems that posed problems for various prior art systems. The "J"

W092/08203 2 ~ 9 `~ 7 0 ~ PCT/US91/03624 ~i in the "John~ in the first line extends below the medial axis in the second line. Furthermore, the "P" in the middle initial actually is connected to the "n" in the line below. The "S" and the "t" in the middle line extend below 5 the top of the ZIP code in the line below~ The lower part of the "S" is furthermore connected ~ith the extendex of the ~irst "0" in the ZI~ code. The second "NY" is angled such that the top of the "NY" is vertically located above the bottom o~ the "S,`' "t," and the "r" in the word 10 "Street'` in the line above. The word ~Main~ is positioned signi~icantly higher with regard to the word "Street'` such that this line of the address is significantly wavy. The letters also vary in size, e.g., compare the two capital "M's." Furthermore, the letters have extra loops, e.g., 15 the "M" in "Main," the initial "P" and the first "O" in the ZIP code. Furthermore, the strokes are relatively jagged and the ZIP code is unevenly separated by the intrusion of the '`t" and the wider spacing between the second "0'` and the "7.`' A digitized address block 21 as shown in Figure 2 forms input 20 into the system. The block 21 can originally be obtained from an optical scan or can be previously digitized and stored on disk or other computer storage techni~ues. The address block 21 is inputted in 25 digitized form in an array of pixels that are horizontally and vertically spaced into rows and columns. The address block pixels can have background states or image states.
The image states can be the state of thinned components, end points, junctions and "flesh," i.e., the part of the ` `W092/08203 2 0 ~ ~ 7 Q ~ PCT/US91/03624 image removed during thinning. The address block 21 is first analyzed for its quality (step 24) which includes noise, dropouts, broken strokes and other image processing defects which can effect analysis of the image. The 5 defects are than quantified and the image is then classified as a certain quality 22. The incoming address bloc~ ~1 then has noise and dropouts filtered (step 26) to the e~tent desired. A bounding box 23 as shown in Figure 2 is then determined ~step 28). A bounding box is 10 determined by the most left and right, and the uppermost and lowermost extent of the strokes of the image as shown by example in Figure 2. The image is then down-sampled within the bounding box 23 as indicated by step 30. The bounding box is formed and downsampled to reduce computer 15 time. The address block 21 then has horizontal stroke separation computed (step 32) as schematically shown in Figure 3. The horizontal spacing between two sequential but separated and horizontally aligned is computed. The ~pacing between different sequential but separated are 20 illustrated by spaces 25, 27, 29 and 31 of differing lengths. Common printing and handwriting provides that there is a common range of spacing between most adjacent vertical strokes as indicated by the stroke distances labeled 31. Significantly smaller gap distances occur less 25 often than stroke distance 31. Distances approximately equal to the gap distance 25 between horizontally adjacent image words occur less often than smaller stroke distances.
In addition, horizontal gaps that are slightly greater than the stroke distance 31 decrease in number such that the W092/08203 2 ~ 9 ~ 7 Q ~ PCT/US91/03624 computer can calcùlate a first peak of stroke distances approximately equal to distance 31 which are labeled an interstroke distance.
With the calculated interstroke distance 31 for 5 address block 21 as a guide, the strokes are then arranged into words (step 34). The words 35 are determined from the interstroke distances 31 and the wider distances 25 and 29 in Figure 3. The result of the grouping step 3~ is shown in Figure 4. Blocks 35 schematically represent the 10 groupings of each word. The vertical stroke connections between the lines are then broken (step 36) via horizontal dilation and subsequent horizontal erosion of the characters where thin descenders and long extenders 37 of characters are eliminated. The elimination of descenders 15 and extenders 37 break the connections between the two adjacent address lines~
Each block of words 35 then is completely separated and these group blocks 35 ar`e then formed into convex hulls 33 about the eroded image. The hulls 33 are 20 then skeletonized as shown in step 38 into skeletal segments 39 as shown in Figure 6. As shown in step 40, the blocks 35 are then further grouped horizontally into address lines. This is accomplished by dilating the skeletonized segments 39 vertically from the skeletal 25 segments 39 to form boxed areas 41 of uniform vertical thickness~ As shown in Figure 7, the boxed areas 41 are then dilated horizontally such that any horizontal overlap of one area 41 with an adjacent area 41 is filled in as indicated in areas 43.

W092/08203 ~ 7 ~ ~ PCT/US91/03624 The next step 44 is to label strokes according to line number. This is accomplished by determining a medial axis 4S for each line 1, 2 and 3 shown in Figure 7 and superimposing the line-numbered medial axis 45 onto the 5 original down-sampled image of the address block 21 as shown in Figure 8~ A line number bleeding process based on connectivity is per~ormed. The bleeding process starts at the line-numbered medial axis. Any lettering that is directly connected to the medial axis or any stroke that is 10 conn~cted to a stroke that is in turn touching the medial axis will be identified as belonging to that medial axis of either lines 1, 2 or 3. For example, the descender of the `'t" in the word "street," even though it is below the top of the "0" and the top of the "7" of the third line will be 15 identified as part of the second line. The "r" in "Nr.,"
even though situated below the ~edial axis 44 will be identified as part of the first line since it is connected to the "M" and the line number bleeding will occur from the "M" into the "r."
The bleeding also helps define letters of two different lines that are connected to each other. The extender for the "P" in the middle initial and the extender of the "n" in "Main" are connected as are and the "S" in the word "Street" and the "0" in the third line are 25 connected. These characters are divided apart by the bleeding of the respective letters. The bleeding of the different lines meet at points 51. By definition any image pixel above point 51 is identified with the upper character and any image pixel below the point 51 is W092/08203 2 ~ 3 ~/, n ~ PCT/US91/03624 `

identified with the lower character. The period 53 and comma 55 are not connected to any axis and are therefor be left unlabeled and designated as punctuation marks.
Based on the bleeding, the punctuation, and the 5 interstroke distances, the strokes are grouped into words within the lines shown in step 44. We no~ have line yrouped word candidates. Step 46 now either discards impossibly small lines or merges the small lines together as indicated in step 48 if there is a hori~ontal overlap 10 detected between these lines. The last or lowest line 57 as indicated in Figure 9 is selected as shown in step 49.
The last line however does not include the lowest extension of the "t" in "Street`' in the line above nor will it include any of the part of the first "0" above the point 15 51.
The last line 57 is split into words as shown in step 50. Fiqure 10 indicates a splitting into three words with "NY" before the punctuation mark 55 forming one word 59, the second "NY" forming a word 61 after the punctuation 20 mark 55 and before the relatively large gap 27 as shown in Figure 3. The "10073" forms a third word 63. Alternative splits are also performed as shown in Figure 11. The alternate four-word splitting has the "loo" in one word labeled 65 and the "73" is in a second word labeled 67.
25 ~he split is a result of the slightly larger distances between the second "0" and the "7." However, due to other constraints, for example, the size of the word 67 being too small for a ZIP code in and of itself based on interstroke distances, the word split shown in Figure lo is preferred ~V092/08203 ~3 ~ r~) Q ~ PCT/US91/03624 over the word split shown in Figure 11. The word 63 is then selected as being the location of the ZIP code and as shown in Figure 12.
Step 52 provides that the corresponding 5 characters 69 for word 63 are shown in full resolution as illustrated in Figure 13. Based upon the class of the image ~uality as indicated in step 22, noise, blobs and dropouts within the word 63 are repaired as indicated in step 70. The blobs and noise are labeled in step 72 and 10 the image pixels in word 63 are shown on a sc~een in step 74. The ZIP code is then transferred to a digit separator in step 76 in which the group of ZI~ code digits can then be segmented, and the digits are then analyzed recognized by step 78 with a digit recognizer. If per chance a 9 15 digit ZIP code is used, the 9 digit ZIP code is detected and the digit separator then determines 9 digits rather than 5 digits.
The computer system used with the ZIP code location process is preferably a morphological type 20 computer which is pre~erably constructed in accordance with U.S. Patent No. 4,167,728 issued September 11, 1979 to Sternberg and entitled `'Aùtomatic Image Processor" and is commercially known by the mark "Cytocomputer." The teachings of U.S. Patent No. 4,167,728 are incorporated 25 herein by reference. Briefly, the construction is described in conjunction with Figures 14 and 15. The overall construction of the morphological computer 70 is illustrated in Figure 14 and the construction of a single W092/08203 ~ 7 a~ PCT/US91/03624 neighborhood processing stage 80 is illustrated in Figure 15.
In general, the morphological computer 70 includes of a pipeline of a plurality of neighborhood 5 processing stages 80, 8~...84. The first neighborhood processing stage 80 receives as its input a data stream corresponding to individual pixels of a binary image as the incoming address block in a raster scan fashion. The image of the incoming address block includes data corresponding 10 to individual pixels arranged in a plurality of rows and columns. The raster scan data stream consists of pixels in order starting with the top row of the left-most pixel to the right-most pixel, followed by the next row in left to right order followed by each succeeding row in similar 15 fashion.
The neighborhood processing stage 80 in turn produces an output stream of data also corresponding to individual pixels of a transformed image in a raster scan sequence. Each pixel of this output data stream 20 corresponds to a particular pixel of the input data stream.
the neighborhood processing stage 80 forms each pixel of the output data stream based upon the value of the corresponding input pixel and the values of the 8 neighboring pixels. Thus, each pixel of the output data 25 stream corresponds to the neighborhood of a corresponding input pixel. The output of each neighborhood processing stage 80,82... is supplied to the input of the next following stage. The output of the last neighboring `` W O 92/08203 2 f~ 9 ~ 7 0 ~ PC~r/US91/03624 processing stage 84 forms the output of the morphological computer 70.
The particular transformation or neighborhood operation performed by each neighborhood processing stage 5 80,82 ..~ 84 is controlled by transformation controller 90~ Each neighborhood processing stage 80,~2...84 has a unique digital address. The transformation controller 90 specifies a particular address on addrass line 92 and then spQcifies a command corresponding to a particular 10 transformation on command line 94. The neighborhood processing stage 80,82... 84 having the specified address stores the command. Each stage then performs the transformation corresponding to its last received command.
Figure 15 illustrates in further detail the 15 construction of an exemplary neighborhood processing stage 80. The neighborhood processing stage 80 operates in conjunction with the delay line formed of pixel elements 100-108 and shift register delay lines llo and 112. Pixel elements 100-108 are each capable of storing the bits 20 corresponding to the data of a single pixel of the input image.
An eight bit or sixteen bit pixel in most foreseeable uses would suffice since standard dilation and skelètonization need to define each pixel in one of five 25 states; background, thinned components, end points, junctions, and "flesh." Shift register delay lines 110 and 112 have a length equal to three less than the number of pixels within each line of the image. The length of the shift register delay lines 110 and 112 are selected to W092/0~203 2a 9~ PCT/US91/03624 ` ``

ensure that pixel elements 100, 103 and 106 store data corresponding to pixels vertically adjacent in the input image. Likewise, the data and pixel elements 101, 104, 107 correspond to vertically adjacent pixels and the data in 5 pixel elements 102, 105, 108 correspond to vertically adjacent pixels.
Pixel data is supplied in raster scan ~ashion to the input o~ the neighborhood processing stage 80. The pixel is first stored in pixel element 100. Upon receipt 10 of the following pixel, the pixel stored in pixel element 100 is shifted to the pixel element 101 and the new pixel is stored in pixel element 100. Receipt of the next pixel shifts the first pixel to pixel element 102, the second pixel to pixel element 101 and the just received pixel 15 stored in pixel element 100. This process of shifting data along the delay line continues in the direction of the arrows appearing in Figure 15. Once the pixel reaches pixel element 108, it is discarded upon receipt of the next pixel.
The neighborhood processing stage 80 operates by presenting appropriate pixel data to combination circuit 114. Note that once the shift delay lines 112 and 110 are filled, pixel elements 100-108 store a 3 x 3 matrix of pixel elements which are adjacent in the original image.
25 If pixel element 104 represents the center pixel, then pixel elements 100, 101, 102, 103, 105, 106, 107, 108 represent the eight immediately adjacent pixels. This combination circuit 114 forms some combination of the nine pixels. Such combinations could take many forms. The 2 ~ 3 ~ rt ~ ~
`~.V092/08203 PCT/US91/03624 pixel output data may have more or fewer bits than the input data depending on the combination formed. It is also feasible that combination circuit 114 may form comparisons between one or more of the pi~els or between a pixel and a 5 constant received from transformation controller so. The essential point is that combination circuit 114 forms an output pixel from some combination of the pixels stored in the pixel elements 100-108.
The advantage of the arrangement of Figures 14 1~ and 15 for image operations is apparent. Each neighborhood processing stage 80,82 ... 84 forms a neighborhood operation on the received image data as fast as that data can be recalled from memory. Each stage requires only a fixed delay related to the line length of the image before 15 it` is producing a corresponding output pixel stream at the same rate as it receives pixels. Dozens, hundreds or even thousands of these neighborhood processing stages can be disposed in the chain. ~hile each neighborhood processing stage performs only a relatively simple function, the 20 provision of long chains of such stages enables extensive image operations within a short time frame. As known from the above description, location of the ZIP code does require a complex and extensive computation due to the number of problems that are inherent in handwritten ZIP
25 codes such as descenders, slanted lines, and jagged lines, irregular spacings, broken strokes, interconnected characters and lines and ink blots and other image noise.
The hardware system such as the one described is needed to provide the computational capacity to work on a typical W092/08203 2 0 9 ll 7 Q g PCT/US91/03624 handwritten address block to locate the ZIP code within that block.
Variations and modifications of the present invention are possible without departing from the scope and 5 spirit of the invention as defined in the appended claims.

Claims (19)

AMENDED CLAIMS
[received by the International Bureau on 24 February 1992 (24.02.92);
original claims 1-19 replaced by amended claims 1-19 (9 pages)]

The embodiments in which an exclusive property or privilege is claimed are defined as follows:

1. A computer system for locating a predetermined group of pixels chosen from a digital pixel image consisting of foreground image pixels and background pixels set forth in an array of columns and rows, said digital pixel image forming characters arranged in plurality of lines, said computer system comprising:
means for computing horizontal distances between horizontally aligned foreground image pixels separated by at least one background pixels and determining the first significant peak distance in a histogram of occurrences of distances, said peak distance referred to as the interstroke distance;
means for horizontally dialating and then horizontally eroding said digital pixel image to enhance vertical separation of said characters in said plurality of lines;
means for grouping the characters together into blocks based on the interstroke distance and wider distances between said characters;
means for skeletonizing said blocks into lines extending the horizontal length or each block;
means for dilating the resulting skeletonized image in a vertical direction to create box areas of uniform vertical thickness;
means for dilating said resulting box areas horizontally such that box areas overlapping in the horizontal direction are merged together to form line images;
means for determining and labeling the medial axis of each respective line image;
means for simultaneously bleeding the foreground image pixels from said each medial axis to identify foreground image pixels connected to a medial axis directly or via other foreground image pixels such that two characters that are connected to two different medial axes and are connected together a-e divided where the bleeding from the two medial axes meet;
means for identifying a desired line of said characters and associating possible wording groups from interstroke distance; and means for selecting said predetermined group of foreground image pixels from said possible wording groups by using interstroke distances.

2. A computer system as defined in claim 1 wherein:
said characters form an address block; and said predetermined group of foreground image pixels form a ZIP code within the address block.

3. A computer system as defined in claim 2 wherein:
said identifying means identifies a last line of said address block large enough to include a ZIP code, and said selecting means selects a last wording group in said desired last line sized large enough to be a ZIP
code.

4. A computer system as defined in claim 1 further comprising:
means for identifying foreground image pixels not connected to any medial axis and thereby not bled by said bleeding means; and said means for selecting said predetermined group of foreground image pixels using said foreground image pixels not connected to any medial axis as punctuation.

5. A computer system as defined in claim 1 further comprising:
means for determining a minimum bounding box completely enclosing the foreground image pixels of said digital pixel image; and means for down-sampling said digital pixel image within said bounding box.

6. A computer system for locating a desired group of characters chosen from a digital pixel image consisting of foreground image pixels and background pixels set forth in an array of columns and rows, said digital pixel image forming characters arranged in a plurality of lines, said computer system comprising:
means for determining the medial axis of each respective line;
means for superimposing each respective medial axis onto said digital pixel image and simultaneously bleeding foreground image pixels forming said digital pixel image from each medial axis to identify foreground image pixels connected to the medial axis either directly or via other foreground image pixels such that two characters that are in two different horizontal lines and are connected together are divided where the bleeding from the two corresponding medial axes meet;
line selecting means for selecting a desired line;
group selecting means for selecting said desired group of characters within said desired line.

7. A computer system as defined in claim 6 wherein:
said input characters form an address block; and said desired group of characters form a ZIP code within the address block.

8. A computer system as defined in claim 7 wherein:
said line selecting means selects the last line of said address block, and group selecting means selects a last group in said desired last line.

9. A computer system as defined in claim 6 further comprising:
means for determining a minimum bounding box completely enclosing the foreground image pixels of said digital pixel image; and means for down-sampling said digital pixel image within said bounding box.

10. A computer system as defined in claim 6 further comprising:
means for identifying foreground image pixels not connected to any medial axis and thereby not bled by said bleeding means; and said means for selecting said desired group uses said foreground image pixels not connected to any medial axis as punctuation.

11. A computer system as defined in claim 6 further comprising:
means for computing horizontal distances between horizontally aligned foreground image pixels separated by at least one background pixel and determining the first significant peak distance in a histogram of occurrences of distances, said peak distance referred to as the interstroke distance; and said means for selecting said desired group of characters uses said interstroke distance as a factor to form groups of which one is said desired group.

12. A computer system as defined in claim 11 further comprising:
means for determining a minimum bounding box completely enclosing the foreground image pixels of said digital pixel image; and means for down-sampling said digital pixel image within said bounding box.

13. A computer system as defined in claim 11 further comprising:
means for grouping characters together into blocks based on said interstroke distance;
means for skeletonizing said blocks into lines extending the horizontal length of each block;
means for dilating said lines in a vertical direction, creating box areas;
means for dilating said resulting box areas horizontally such that box areas overlapping in the horizontal direction are merged together to form line images of each line.

14. A computer system for locating a predetermined group of characters chosen from a digital pixel image consisting of foreground image pixels and background pixels set forth in array of columns and rows, said digital pixel image forming characters arranged in a plurality of lines, said computer system comprising:
means for computing horizontal distances between horizontally aligned foreground image pixels separated by at least one background pixel and determining the first significant peak distance in a histogram of occurrences of distances, said first peak distance referred to as the interstroke distance; and means for selecting said predetermined group of characters within one of said plurality of lines by using interstroke distances.

15. A computer system as defined in claim 14 wherein:
said input characters form an address block; and said predetermined group of characters form a ZIP
code within the address block.

16. A computer system as defined in claim 14 further comprising:
means for determining a minimum bounding box completely enclosing the foreground image pixels of said digital pixel image; and means for down-sampling said digital pixel image within said bounding box.

17. A method for locating a desired group of characters chosen from a digital pixel image consisting of foreground image pixels and background pixels set forth in an array of columns and rows, said digital pixel image forming characters arranged in 2 plurality of lines, said method comprising the steps of:
computing horizontal distances between horizontally aligned foreground image pixels separated by at least one background pixel and determining the first significant peak distance in a histogram of occurrences of distances, said peak distance referred to as the interstroke distance;
horizontally dialating and then horizontally eroding said characters to enhance vertical separation of said characters in said plurality of lines;
determining the medial axis of each of said plurality of lines;
simultaneously bleeding foreground image pixels from each axis to identify all foreground image pixels connected to a medial axis directly or via other foreground image pixels such that two characters that are in two different lines and are connected together are divided where the bleeding from the two corresponding medial axes meet; and selecting said desired group of characters by using said interstroke distance.

18. A method as defined in claim 17 further comprising the steps of:
determining a minimum bounding box completely enclosing the foreground image pixels of said digital pixel image; and down-sampling said digital pixel image within said bounding box before said computing step.

19. A method as defined in claim 17 further comprising the steps of:
identifying foreground image pixels not connected to the medial axis of any of said plurality of lines; and said selecting step uses said identified foreground image pixels not connected to any medial axis as punctuation to assist in selecting said predetermined group of characters.