JP2579413B2 - Character recognition method and device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for recognizing characters of a document which has been scanned and digitized in a dot pattern. More specifically, the present invention is applicable to any character recognition task,
It is particularly useful for recognizing handwritten characters, such as, for example, characters handwritten on bank checks.

[0002]

2. Description of the Related Art In the past, character recognition technology has collated a character with a template or a key area of a scanned character, and searched for a character feature serving as a key for identifying the character. These techniques work well as long as the characters to be recognized are limited to particular shapes or key features. If this restriction is relaxed, as in the case of handwritten characters, the success rate in analyzing character templates and key features will decrease.

One technique for handling handwritten characters is taught in US Pat. No. 4,757,551.
The patent treats handwritten characters by detecting the slope of a character stroke and comparing the character information with information stored in two different character libraries. The two character libraries are different based on the slope of the reference character in the library. The US patent uses directional codes for character outline pels. The directional code simply indicates the direction from one pel of a given character to the next adjacent pel of the character. The histogram of these directional codes is obtained for a plurality of portions of the character, and the character is recognized by comparing it with a library of histograms of the characteristics of the reference character.

The difficulty with the technique of this US patent is that it is essentially a template matching technique, which requires a huge number of templates (reference histograms) which must match the character feature histogram. Therefore, this technique requires a large storage capacity and is very time consuming to execute.

[0005]

SUMMARY OF THE INVENTION An object of the present invention is to improve the accuracy of character recognition, particularly by improving the accuracy of machine-printed (omni) characters or handwritten characters with less restrictions. is there.

It is another object of the present invention to reduce the time required to accurately recognize characters, especially omni or handwritten characters.

[0007]

SUMMARY OF THE INVENTION In accordance with the present invention, the above object is achieved by using a plurality of radial view patterns from a selected pel in a digitized dot pattern of a character. , The remote stroke of the character (remote str)
okes).
More specifically, the radial pattern is a two radial pattern that includes remote stroke information in more than one radial direction. The selected pels are all pels of the character,
A pel that forms the outline of a character, or a core pel that represents the inside of a character stroke. The radiation pattern from one pel includes eight radiation directions. Furthermore, for each radial direction, 45
A direction rotated by degrees is added. In each of the encodings of the pattern, the stroke length or distance to a remote stroke in one particular direction is compared to the stroke length or distance to a remote stroke in an adjacent counterclockwise direction. This comparison gives an indication of the direction of either the stroke or the remote stroke segment containing the pel. Thus, the two radial direction patterns for each pel include the radial direction component of the character stroke information and the rotational direction component of the character stroke information, and are therefore named two radial direction patterns.

[0008] Counts of the same radial pattern from the character pel are accumulated. The counts from the radial pattern form a character vector. Next, the character vector is analyzed by a linear decision network or a neural network to determine the character represented by the character vector. The radial pattern counts are weighted and added so that the character vectors for a particular character shape converge in a direction that also has a weighted sum. Thus, the linear decision network can distinguish each character. Neural networks are used to collect weighted sums for the same character feature into one or more subclusters. These intermediate weighted sums are further weighted and added again to classify the same sum into one recognizable class as its associated characters.

[0009]

A major advantage of the present invention is that it is powerful because the comparison in encoding two radial patterns contains information about the direction of the character stroke.
The additional rotational component in the two radial patterns is combined with the accumulation of the same two radial patterns to show character features common to a given character. Furthermore, character vectors are quickly processed into character clusters by linear decision networks or neural networks. Therefore, there is no need to perform time-consuming matching of the character feature with the reference pattern.

[0010]

FIG. 1 shows a model 3890 from IBM Corporation.
Image Plus HighPerformance
1 illustrates an image acquisition system 10 in a document image processing system 11 such as an eTransaction system. Such systems typically include a transport mechanism 13
, The documents 12 are read one by one at a time from the input document (not shown) based on the contents of the information read at the time of passage from the document under the program control. Move through the processing station to one of several output hoppers or pockets. This transport mechanism 13 is capable of moving documents at high speed (model 3890 is at least 750 cm (300 inches) per second, and the system can process 2400 documents per minute).

An optical system 14 with a lens 15 uses an elongated array 17 to obtain an electrical representation of each line 16 of the document based on the black and white or gray characteristics of the document. The resolution of the system depends on the design and components, but in one example, in each of the horizontal and vertical directions, 2.5 cm (1 inch) is divided into 240 pixels (pels or pixels).

Array 17 is a charge-coupled device, available from various manufacturers such as Reticon, which generates an electrical signal proportional to the amount of light incident on each sensor. This electrical signal reaches an analog-to-digital converter 19 via a line 18 where each pel is converted to a digital representation by a threshold. This is stored in memory 20 for storage and / or further processing if desired. For details of the image acquisition processing system, see US Pat.
See U.S. Pat. No. 8,812.

The document 12 contains various indices, which may be printed or handwritten.
Other documents may include typewritten or machine printed characters. The indicator includes a date area 22, a payer line 24, an amount area 26 (which, in most cases, includes numbers rather than letters), an area 28 for writing the amount in letters, a signature 30, and a MICR code line 32. In.

FIG. 2 shows the document 12 of FIG.
Is enlarged. This money area 26 is composed of a “$” identified by reference numeral 30 and a box 32 having a money area therein. The dotted line referenced by the number 34 indicates how the money area is divided.
This partitioning can be performed by any known technique.

As a result of the segmentation, each character is identified separately, if possible, and subjected to the character recognition algorithm of the present invention.

In this case, reference numeral 36 identifies a zero in the monetary area that occurred after the decimal point and before the one. 4, 5 and 9, the use of this character in the recognition process will be discussed later.

Radial Pattern FIG. 3 illustrates the logic or flow for identifying a character using a radial pattern. At block 50,
An area such as a money area is identified. If desired, the section can be copied to individual sub-regions for further processing. This is advantageously done by the above mentioned patent on region finding. Next, at block 52, the image is segmented into individual characters or indices to be recognized, such as individual characters to be recognized, such as "@", decimal point, and the like. This can be done by any known character classification method.

Next, the image is normalized at block 54, and the radial pattern is calculated at block 56. All of these are described in more detail below in connection with FIGS. 6-9 (especially FIG. 6). These radial patterns in the preferred embodiment include the process of taking the radial pattern in each of the black pels in the normalized image from block 54.

Next, at block 58, the radial pattern is used. A training set generates probabilities for each of the generated vectors (see FIG. 8). In training mode, these radial patterns are used to generate a look-up table.

The radial pattern of the unknown character X is entered into a previously created table and the probability that a given vector (FIG. 8) is associated with each of the characters in the character set is examined.

FIG. 4 shows the character 0 taken from the money area of FIG.
It is shown. As shown in this figure, there are a plurality of horizontal rows (R0-R25) and a plurality of rows (C0-C22), forming a rectangular array identified as containing the character 0.

Next, the character is normalized, and the line (R0-
R9) and the columns (C0-C7) to form the image of FIG.

FIG. 6 illustrates another letter (number 7) having a central pel 60 in a circle and eight rays 61-68 directed to each of the eight identified compass orientations. I have. That is, line 61 faces north from Pell 60, line 62 faces northeast, line 63 faces east, and so on.
Is facing northwest. The problem is, along each of these lines that penetrate the entire character box, whether or not, and if so, where, the black pel is.

As shown in FIG. 7, the description of a pattern from any given pel pointing in any particular direction is:
There are no black pels on the path (in this case, a pattern number of 0 is assigned), adjacent pels on the path are black (in this case, a pattern number of 1 is assigned), or adjacent pels are Either white and finally black on the path (in this case, a pattern number of 2 is assigned). Pattern number 2
Is also assigned when a diagonal passes between two black lines. These pattern numbers and pattern descriptions are arbitrarily assigned, and it is effective to use other description and numbering techniques.

Thus, in FIG. 8, the radial pattern for the pel 60 in FIG. 6 is identified. That is, in the direction along the line 61 facing north from the pel 60, the adjacent pel is white, followed by the black pel adjacent thereto, and the pattern number 2 is obtained. Line 62
In the northeast direction along, the adjacent pel is black,
Pattern number 1 will be assigned to the northeast direction. In the east direction along line 63, there is no black pel in the path from pel 60 to the eastern edge of the character, so the pattern number is zero. Similarly, in the southeast direction, here again line 65
As in the south direction along, there is no black pel in the path. Along the line 66 going southwest from the pel 60, the adjacent pel is black and will be assigned pattern number 1. In the direction west of pel 60, there is no black pel in the path, so the pattern number is 0, and in the northwest direction along line 68, there is an adjacent white pel, and finally a black pel, The pattern number is 2 for the direction.

Applying this technique to the pattern of FIG.
The table shown in FIG. 9 is created. In this table, positions R1, C
The row number 1 corresponding to No. 5 is “0012210” in the directions of north, northeast, east, southeast, south, southwest, west, and northwest, respectively.
0 "vector.

Row 2 is a similar radial pattern for the next black pel located at R1, C6. This continues for each black pel, but only a few of them are listed here.

For each vector created (eg, vector 10012100 in row 1), an established probability of occurrence exists and is stored in the look-up table. The vector represents the pattern of the scanned character. For a set of numbers, the allowed characters are the digits 0-9, but the monetary area may contain other characters, for example, symbols that are allowed and recognized, such as "$", decimal point, etc. .

After all probabilities for rows 1 through N have been determined, the probabilities are checked against the known probability of occurrence of the vector for each character in the set of numbers. What is matched is the product of the probability of occurrence of the individual vectors of the character to be identified. This is checked against the product of its vector occurrence probabilities for each character in the set of numbers. In FIG. 9, the occurrence probability of each character in the set of numbers is P1, P2, P3,.
For example, the probability of occurrence of the first character of the set is A1 × B1 × C1 × D1 × E1 ×... × I1.
The probability that this character is the second character in the set is A2 ×
.. × I2.

These products can then be normalized:
The probability of an unknown character, one element, can be compared to the probability of the next element. Then, to consider a character successfully recognized using the appropriate recognition algorithm,
How large the absolute value probability is needed,
Second, a threshold limit is set on how large the first candidate must be to prevent it from being too close to be called.

Two Radial Patterns The logic or flow of FIG. 3, as described above, is very effective in achieving character recognition, but the logic or flow is
Further enhancement can be achieved by using radial patterns. FIG. 10 is a logic / flow diagram for using a two radial pattern. The input to the flowchart of FIG.
After block 54 of "Normalize (optional)" in FIG. As described in FIG. 1, the image of the character is digitized into pels. FIG. 11 shows a digitized image of the digit 3 and eight radial patterns identified by compass orientation from the black pel 80 near the center of the letter or digit 3 and from the pel 84 at the top of the digit 3. Things.

At block 72, a two radial pattern (described below with reference to FIGS. 12 and 13) for the selected pel is determined. The pel selected in the preferred embodiment is the pel that forms the outline or outline of the character.
Obtaining two radial patterns for all black pels gives most of the information for analysis. However, performing this process requires significantly more computer time than using only contour pels. A second alternative is to normalize the character, as in FIG. 3, and use a single pel-wide contour as the core of the character. Yet another alternative is to use pels inside the character (all pels except the outline pel of the character).

After the two radial patterns are generated for each pel of the contour in block 72, block 74 generates a character vector from the two radial patterns. This character vector is a count list or histogram of two radial patterns for the pel of the character.

A character vector can also be analyzed by any number of processes, shown as block 76 in FIG. In the preferred embodiment of the present invention, either a linear decision network or a neural network is used. In the case of a linear network, the count of each two radial pattern of the character is weighted and added to the count of the other weighted two radial patterns. The weight is determined by training the network with thousands of character samples. The network generates a weighted sum from the character vector for each possible character. The sum for each character is then compared to identify the character with the highest value of the weighted sum as the recognized character. Any number of algorithms for recognizing characters from weighted sums can be used. A simple algorithm makes the sum greater than the threshold the largest sum. Another algorithm is
If the difference between the largest sum and the next largest sum is greater than a second threshold, the sum greater than the first threshold is maximized. The analysis by the linear determination network will be described later in detail with reference to FIGS.

When block 76 is implemented as a neural network, there is an intermediate layer of added values generated by weighting and adding the two radial pattern counts from the character vector. This intermediate sum represents character features that tend to occur (ie, aggregate) within similar groups. Restricting and weighting these intermediate additions and adding the restriction again results in an addition near zero or one for each possible character in the character set. Only the sum of one class should have a value close to one. That class identifies the recognized character from the character vector.

To understand how the two-radiation pattern works to enhance character recognition, first consider the four possible two-radiation pattern codes 0, It is necessary to understand the encoding operations 1, 2, and 3. The conditions necessary to satisfy a given pattern code are illustrated in the logic diagrams of FIGS. 12-15 and are as follows. The following terms are defined for that pattern code; "V" is the given radial direction; "V-45" is the radial direction rotated 45 degrees counter-clockwise from V. " _Lv " is the stroke length in the radial direction, i.e., the number of black pels traversing to reach the first white pel; " _Lv-45 " is the stroke length in the V-45 direction. "D _v " is the distance to the remote stroke in the V direction,
That is, the number of black pels (if any) and white pels traversed to reach the first black pel of the remote stroke; "D _v-45 " is the distance to the remote stroke in the V-45 direction. .

The definition of the pattern code is as follows: pattern code 0; no remote stroke in the V direction; and no remote stroke in the V-45 direction; and the stroke length in the V direction. Is smaller than the stroke length in the V-45 direction; or the remote stroke is in the V-45 direction; and the stroke length in the V direction is smaller than the distance to the remote stroke.

Pattern code 1; no remote stroke in the V direction; and no remote stroke in the V-45 direction, and the stroke length in the V direction is greater than or equal to the stroke length in the V-45 direction. Or the remote stroke is in the V-45 direction; and the stroke length in the V direction is greater than or equal to the distance to the remote stroke in the V-45 direction.

Pattern code 2: remote stroke exists in the V direction; and no remote stroke exists in the V-45 direction; and the distance to the remote stroke in the V direction is the stroke in the V-45 direction. Less than the length; or there is a remote stroke in the V-45 direction; and the distance to the remote stroke in the V direction is less than the distance to the remote stroke in the V-45 direction.

Pattern code 3: remote stroke exists in the V direction; and no remote stroke exists in the V-45 direction; and the distance to the remote stroke in the V direction is greater in the V-45 direction. Greater than or equal to the stroke length; or the remote stroke is in the V-45 direction; and the distance to the remote stroke in the V direction is greater than or equal to the distance to the remote stroke in the V-45 direction.

With the above definition of the pattern code in mind, the process of determining the radial pattern for character 3 in FIG. 11 begins with measuring the distances _Lv , _Lv-45 , _Dv, and _Dv-45. Is done. Next, these distances are shown in FIGS.
To select the correct two radial pattern code.

To illustrate the method of determining the two radial pattern codes, a pattern number for pel 80 in FIG. 11 is generated. The pel looking at the pattern is called the starting pel here. In the north radial direction, decision 90 in FIG. 16 branches to the “positive” side to block 92 and 94
, The position of the pel 81 of the first remote stroke along the V (north) direction is detected, and the distance from the origin pel 80 to the pel 81 is measured. As shown in FIG.
The distance D _v-45 is 5 pels.

Next, decision 96 tests the remote stroke in the V-45 direction. Since pel 80 has a remote stroke in the NW direction, this decision branches to the “Yes” side to blocks 98 and 100. In block 98, V-4
The position of the pel 82 in the first remote stroke in the 5 (NW) direction is detected. Block 100 measures the distance between pels 80 and 82. This distance D is 5 pels.

The remote stroke information determined in FIG. 16 can be applied to the logic flow diagrams of FIGS. 12-15 to determine the two radial pattern codes from the origin pel 80 in the north direction. The presence of a remote stroke in the radial direction V is indicated by RS _v , the absence is indicated by NRS _V ,
Having a remote stroke in the V-45 direction is RS _v-45 ,
The absence is indicated by NRS _V-45 . Since there is a remote stroke in the radial direction (RS _v ), the logic flow diagram of FIG. 14 or FIG. 15 is applicable. In addition, because there is a remote stroke in the northwest direction (RS _v-45 ), AND gate 102
And 104 are enabled. However, D _v =
D _v-45 (the distance between them is 5 pels) holds,
Only AND gate 104 has an output. The output from AND gate 104 passes through OR gate 106 and AND gate 108. Therefore, the two radial pattern code for the northward pattern is "3".

The logic flow diagrams shown in FIGS. 12-15 are preferably implemented in software. Of course, these logic / flow diagrams, as well as other logic / flow diagrams of the present invention, can be implemented in hardware logic blocks, programmable array logic, or software. In each case, the preferred embodiment is software, i.e., a computer program running on a computer or data processing system.

The distance and remote stroke information required to determine the two radial patterns for pel 80 of FIG. 11 is summarized in the table of FIG. The bottom of each column,
Two radial pattern codes for the pattern in that direction from pel 80. Distance and radial pattern
The code can be reached by going through the process described above for FIGS. 16 and 12-15.

The two radial pattern for a given origin pel is defined as the combination of the eight two radial pattern codes obtained for that pel. The two radial pattern (SVP) for pel 80 obtained by listing the two radial pattern codes clockwise from north is 31033203. Converting each bi-directional pattern code for origin pel 80 to binary, the bi-radial pattern is a 16-bit number (2 bits for each pattern) or 110100111110.
0011.

In the preferred embodiment, two radial patterns are determined for each pel on the character outline. (2) As a second example of the determination of the radial direction pattern, the starting point
Are generated here.

This process is shown in FIG. 16 and for the northwest pattern, since there is no remote stroke northwest of pel 84, decision 90 branches the flow to the right, leading to blocks 110 and 112. In the northwest pattern, block 110 detects the location of the background pel immediately adjacent to pel 84. Therefore, block 112 measures L as a zero pel. Next, decision 96 determines that V
Examine the remote stroke in the -45 pattern (west). Since there is no remote stroke in the west direction, the flow branches to blocks 114 and 116. Block 114
Detects the position of the first background pel in the V-45 direction. Block 116 measures the stroke length in that direction. That is, L _v-45 is 8 pels. This information is shown in FIG.
By applying to the logic flow chart of FIG. 15 to FIG. 15, the logic of FIG. 12 is satisfied. There is no remote stroke in the V-45 direction (NRS _v-45 ), and L _v (zero pel) is L _v-45 (8
(Pel), and AND gate 118 has an output. The signal from AND gate 118 passes through OR gate 120 and AND gate 122. AND gate 122 is enabled by the absence of a remote stroke radial northwest (NRS _v). Therefore, the two radial pattern codes for the NW pattern from pel 84 are "0".

FIG. 18 is a table showing the distance to the pel 84, the two radial stroke information, and the two radial pattern SVP. When listed clockwise from the northward pattern, the two radial pattern for pel 84 is 11123010. 2 each for pel 84
Converting the radial pattern code to binary, the two radial pattern is a 16-bit number, ie, 010111
011000100.

Another pattern from the origin pel 84 is notable because it includes two remote strokes. This pattern is a pattern in the south direction from the pel 84.
This pattern intersects the middle and lower strokes of number 3 in FIG. In this situation, up to the first or second remote stroke can be measured. In the preferred embodiment of the present invention, this measurement is made for a first remote stroke that is a distance of 7 pels in a southward pattern from pel 84.

The result of the process of FIGS. 12-16 is a 16 bit number for each pel on the character outline. This 16 bit number is the two radial pattern for that pel. Therefore, 0000000000000000000
0 to 111111111111111, ie
There are two possible bi-directional patterns from 0 to 65535 (64k). Since two radial patterns are determined for the pel of the character to be recognized, some of these patterns are not used and some are repeated one or more times.

The next block in the character recognition process of FIG. 10 is block 74, the generation of the character vector from the two radial patterns. This block is shown in FIG.
In more detail. Block 12 in FIG.
Numerals 4 and 126 are used to count two radial patterns in the character under analysis. Block 124 identifies a common bi-radial pattern of characters and block 12
6 accumulates the number of all pels having the same two radial patterns. Upon completion of the iteration count for each of the two radial patterns, block 128 creates a character vector by tabulating the counts sequentially from pattern 0 to pattern 65535. In the generalized shape, the character vector is represented by C = (c0, c1, c2,..., CN,..., C65535), where cN is a count value for two radial patterns N.

Block 76 in FIG. 10 identifies a character by analyzing a character vector. Analysis Block 7
One of the execution examples of No. 6 is the linear decision network of FIG. The task of the linear decision network is to indicate the range indicated by the character vector indicating that the character is a predetermined character. The counts for each of the two radial patterns are loaded into the associated input registers of FIG. The count value c1 of the two radial patterns 1 is loaded into the input register 130.
Similarly, the registers from the count value c2 to the count value c65535 are loaded into the registers. Only input registers 132 and 134 for c2 and c65535 are shown. c
The count value for 0 is forcibly set to 0, and the count value of 1 is loaded into the register 129. In the linear judgment network, the register 129 is a constant (in this case, the count value of 1).
Need to be loaded in. This requirement does not adversely affect the recognition process since no radial pattern zero can occur.

Each of the counts is multiplied by a weighting factor, and this weighted count is added to each possible character identification. For example, the addition circuit 136 that indicates a classification whose output value is possible as a number “0” performs the following addition.

[0056]

## EQU1 ## Sum of class 0 = 1 × W (0,0) + c2 × W (0,1) +... + CN × W (0, N) +... + C65535 × W (0,65535) Where “cN” is a count value for the two radial patterns N, and “W (0, N)” is a weighting factor for cN in classification 0.

In fact, the output representing the sum of class 0 from the adder circuit 136 is the inner product of the character vector and the weight coefficient vector for class 0. Similarly, an adder circuit 138,
The sum of classes 1, 2 and 3 from 140 and 142 is the inner product of the character vector and the weighting coefficient vector for classes 1, 2 and 3, respectively. The same number of adder circuits and weighting coefficient vectors as character set patterns, that is, character patterns in which character vectors are classified or recognized, are provided.

FIG. 21 shows how the sum of the classes for the numbers 0, 1, 2 and 3 shows how the input counts cN-1 and cN
FIG. 4 is a simple dispersion diagram showing whether or not a set is gathered. In other words, in a two-dimensional space in which the X axis is cN-1 and the Y axis is cN, cN-1, c
The N coordinates tend to aggregate where they are "1" in FIG. Similarly, cN-1, for the numbers 0, 2, and 3,
The cN coordinates also tend to aggregate. Next, using the weighting factor, the count value cN-1 generated in the vector analysis of the number is calculated.
And cN to generate a separate class sum for each of the numbers. In the simple example of FIG. 21, the count values cN−1, cN
By using the weighting factors W (0,1) and W (0,2) for N, if the count is from a character pattern with the number zero, the sum of class 0 is increased and the sum is calculated. If the number is from one of the other character patterns,
The sum of class 0 decreases.

Schematically, each cluster in the scattergram can be separated from other clusters by one line.
In a linear decision network, this line is drawn by a weighting factor. Of course, the complete variance diagram for a character vector has 64k dimensions. If such a space can be illustrated, clusters still exist and they can be separated by a hyperplane (a plane in a 64k dimensional space). The weighting factor vector used in the linear decision network is to separate clusters for classification of character or symbol patterns,
You will draw these hyperplanes.

The weighting factors of FIG. 20 must reach the use of a training set of thousands of character sets. Each of the characters used for training is digitized into pels, two radial patterns are determined for the pels of the contour, and a character vector is generated. Each count cN of the character vector, as is known, generates a sample weighting factor that increases the sum for character identification and a weighting factor that reduces the sum for all other character or symbol patterns. Used for

FIG. 22 is a logic / flow diagram of the last part of character analysis based on the linear decision network. The final process is character recognition based on the class sum from FIG. Blocks 144 and 146 of the flow chart correspond to FIG.
Of the largest and second largest class sum (C
S1 and CS2) are obtained. At the same time, block 144 identifies the character class C having the largest class sum CS1.
Next, decision block 148 tests to see if the maximum sum CS1 is greater than the minimum absolute threshold TH1.
If CS1 is not greater than TH1, block 150 indicates that the pattern or character is not recognizable, and flow terminates the process for digitized characters and exits.

If CS1 is greater than the minimum threshold value TH1, the character represented by CS1 can be identified as a scanned character. However, for greater accuracy, the difference between CS1 and CS2 is taken at block 152 in FIG. Next, decision block 154
Then, a test for comparing this difference with the minimum delta threshold value TH2 is performed. If the difference is less than TH2, flow branches to block 150 and the character is displayed as unrecognizable. If the difference is greater than TH2, block 156 identifies the character as "C", the character represented by the class with the largest sum CS1. Here, character recognition is successfully completed.

When the characters to be recognized have almost no restrictions on the shape, such as in free handwriting printing or shape-free machine printing (omni characters), there are considerable changes within each class. For example, the upper part of the number 3 is curved or flat, and the right side of the number 4 is open or closed. In these cases, a sub-cluster of counts for the class for the character is formed. FIG. 23 is a dispersion diagram of the count values cN-1 and cN of the pel / radial direction patterns N-1 and N. In this scatter diagram, the counts for some character vectors have two subclusters (see double clusters for counts from characters "1", "2" and "3" in FIG. 23). These subclusters for a given class cannot be separated linearly as a pair. Therefore, the analysis is performed using the feed forward neural network shown in FIG. 24 instead of the linear decision network as shown in FIG.

The function of a neural network is similar to a linear decision network, but with two significant differences. First, there is an intermediate addition circuit layer between the input count value register and the final classification addition circuit.
In addition, the addition circuit has a limiter for limiting the addition value between zero and one. The function of the hidden layer is to classify sub-clusters or character features. The sum of the outputs for the classes of subclusters is weighted and restricted to arrive at the final classification of character recognition. In fact, the hidden layer is the inner product of the character vector and the subcluster weighting coefficient vector,
Generate a subcluster vector or a character feature vector. Thereafter, the final addition layer classifies the subcluster vector or feature vector, which is the inner product of the weighting coefficient vector and the subcluster vector, into a character class.

Referring to FIG. 24, input registers 159 to 1
Numeral 61 receives the count value of two radial patterns, that is, a character vector, in the same manner as the input register of FIG. Again, a count value of 1 is loaded into the input register of the count value c0. All other counts c1-c65535 are loaded into their respective registers (c1, c2 and c
Only 65535 registers are shown). Each of these counts is applied to a cluster addition / limitation circuit 162, 1
Weighted by a weighting factor associated with each of 63 and 164. Only three sub-cluster limiting circuits are shown. For each sub-cluster, there is a sub-cluster adder. Each subcluster addition / limitation circuit includes a limiter having a function approximating a step function. In other words, the outputs of the sub-cluster circuits 162-164 are close to zero or one.

Similarly, the class addition / limitation circuits 165-1
Each of 68 is a combination of an addition circuit and a nonlinear step function approximation circuit. The class 0 circuit 165 finds the inner product of the subcluster vector and the weight coefficient vector of class 0, and limits the result to near zero or one. Circuits 166, 167 and 168 perform the same function for classes 1, 2 and 3, respectively. Of course, the same number of class addition / restriction circuits as character recognition set patterns or characters are provided. If only one class adder / limiter circuit has an output value near 1 and all other class adder / limiter circuits have output values near zero, the character vector is classified (character is recognized). Otherwise, the character is rejected as unrecognized.

The weights used for the weighting factor vectors of FIG. 24 are generated from a representative training set by standard neural network training techniques. Such techniques are taught in the following two documents: David Ramelhart
hart) One other author, "Parallel distributed processing", Volume 1, MI
T Press, (c) 1988, pp. 322 to 330; Sara Sola et al., "Accelerated Learning in Hierarchical Neural Networks," Complex Systems, Vol. 6. Complex System Publication, (c) 1988
Year, pp. 625-639.

[0068]

As is apparent from the above description with reference to the embodiments of the present invention, the present invention can recognize characters with high flexibility such as machine-printed characters and handwritten characters in a short time. It can be performed, and furthermore, it has a special effect that its accuracy can be improved.

[Brief description of the drawings]

FIG. 1 is a diagram showing an environment of the present invention, that is, an image document processing system.

FIG. 2 is a diagram showing a money area from a check of the sample of FIG. 1;

FIG. 3 is a logic flow diagram using radial patterns for character recognition.

FIG. 4 is a diagram showing a matrix of a pel component of one character from the money amount area of FIG. 2;

FIG. 5 is a diagram showing a skeletonized and normalized version of the character of FIG. 4;

FIG. 6 illustrates a radial pattern recognition system useful for generating all the multi-directional vectors seen from a given pel.

FIG. 7 is a diagram showing a table in which observation results in a specific direction from a specific pel are described and pattern numbers are assigned to the observation results.

FIG. 8 shows a vector of samples for one pel of FIG. 6;

FIG. 9 is a diagram showing a table including pattern vectors and probabilities for the characters of FIG. 5;

FIG. 10 is a logic / flow diagram for the present invention implementing a two radial pattern in an image document system.

FIG. 11 is a diagram showing an example of a digitized character showing eight patterns from a character pel near the center of the character and a character pel at the top of the character.

FIG. 12 is a logic / flow diagram for determining a two radial pattern code.

FIG. 13 is a logic / flow diagram for determining a two radial pattern code.

FIG. 14 is a logic / flow diagram for determining a two radial pattern code.

FIG. 15 is a logic / flow diagram for determining a two radial pattern code.

FIG. 16 is a logic / flow diagram for measuring stroke features used by the two radial pattern code determination logics in FIGS.

FIG. 17 shows a table of stroke measurements and two radial patterns for pel 80 of FIG. 11;

18 shows a table of stroke measurements and two radial patterns for the pel 84 of FIG. 11;

FIG. 19 is a logic / flow diagram for the two radial pattern, character vector determination required in FIG.

FIG. 20 is a logic / flow diagram of a linear decision network for the character vector analysis required in FIG.

FIG. 21 is a scatter diagram of two counts in a character vector for the numbers 0, 1, 2, and 3.

FIG. 22 is a logic / flowchart for recognizing a character from a weighted sum generated for each character candidate by the linear determination network of FIG. 21;

FIG. 23 is a cluster diagram of character vectors for characters with few restrictions, such as hand-written characters or machine-printed characters with few restrictions.

FIG. 24 is a logic / flow diagram of the neural network for character vector analysis required in FIG. 10; the neural network exerts power in character identification when character vectors are aggregated as in FIG. 23; .

DESCRIPTION OF SYMBOLS 11; document image processing system 10; image acquisition system 13; transport mechanism 12; document 14; optical system 15; lens 17; array 19; analog / digital converter 20; memory 22; date area 24; User line 26; amount of money 159 to 161; input register 162, 163, 164; cluster addition limiting circuit 165 to 168; class addition / limiting circuit

 ────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Terry Anthony Will 9217, Forest Manor Court, Charlotte, North Carolina, USA 28215 (56) References JP-A-56-59374 (JP, A) JP-A-62-287386 (JP, A) JP-A-2-29085 (JP, A)

Claims (9)

(57) [Claims] 1. A character recognition method in a character recognition system, comprising: a scanner that scans a character and generates an image signal of the scanned character; and an image digitizer that digitizes the image signal into a pel character pattern. Scanning a character by the scanner to generate an image signal; digitizing the image signal into a pel pattern by the digitizer; selecting a plurality of starting pels from the pels of the character pattern;
For each of them, examine the stroke of the pattern of pels along the plurality of radial directions from the origin pel, and measure the length of the first stroke measured from the origin pel in the first radial direction, adjacent to the first radial direction The length of the initial stroke measured from the origin pel in the radial direction, the distance from the origin pel to the remote stroke measured in the first radial direction, and measured in the radial direction adjacent to the first radial direction Measuring the distance from the origin pel to a remote stroke; and, for each of the origin pels, comparing the length of the two strokes and the two distances with each other, and according to the comparison result, the distance from each origin pel. (2) generating a vector characterizing the radiation direction pattern; and solving the vector in a logical network. And classifies the scanned character, on the basis of that classification, character recognition method, which comprises identifying the said scanned character as a character of a predetermined character in the set. 2. The method of claim 1, wherein the step of generating a vector comprises: generating a count of the number of vectors having the same value among vectors generated in the scanned character. Representing said counts as a histogram for all origins of the character being scanned, thereby forming a vector. 3. The method of claim 1, wherein analyzing the vector comprises: generating a dot product of the vector with a weighted coefficient vector defined for each character in the character set; Comparing to a threshold and, when the threshold is exceeded, identifying the scanned character as a character associated with the largest dot product. 4. The method of claim 1, wherein the step of analyzing a character vector comprises generating a subcluster inner product of the character vector and a subcluster weighting coefficient vector for each subcluster in a character in the character set. Limiting each of the subcluster inner products to values near zero or one and forming a subcluster vector from all of the restricted subcluster inner products; and Generating a character inner product with a weighting coefficient vector for each character in the following: limiting each of the character inner products to a value near zero or one; and all other restricted inner products near zero. Once, 1
Identifying the scanned character as a character associated with a single restricted character dot product having a value in the vicinity of. 5. A scanner which scans a recognized character to generate an image signal, means for digitizing the image signal into a pel pattern by the digitizer, and a plurality of starting pels among the pels of the character pattern. Selected,
For each of them, examine the stroke of the pattern of pels along the plurality of radial directions from the origin pel, and measure the length of the first stroke measured from the origin pel in the first radial direction, adjacent to the first radial direction The length of the initial stroke measured from the origin pel in the radial direction, the distance from the origin pel to the remote stroke measured in the first radial direction, and measured in the radial direction adjacent to the first radial direction Means for measuring a distance from the origin pel to a remote stroke; and, for each of the origin pels, comparing the lengths of the two strokes and the two distances with each other and according to the comparison result, (2) means for generating a vector characterizing the radiation direction pattern, and analyzing the vector with a logical network, Classifies the scanned character, on the basis of that classification, the character recognition apparatus characterized by comprising, means for identifying the scanned character as a character of a predetermined character in the set. 6. The character recognition apparatus according to claim 5, wherein the selected starting point pel is a pel of the outline of the scanned character. 7. The character recognition device according to claim 5, wherein the selected starting point pel is a core pel of the scanned character. 8. The character recognition device according to claim 5, wherein the selected origin pel is all pels of the scanned character. 9. The character recognition apparatus according to claim 5, wherein the means for generating a vector generates a count value of the number of vectors having the same value among the vectors generated in the character being scanned. Means for displaying the counts as a histogram for all origins of the scanned character, thereby forming a vector.