Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
  • G06V30/10—Character recognition
  • G06V30/32—Digital ink
  • G06V30/333—Preprocessing; Feature extraction
  • G06V30/347—Sampling; Contour coding; Stroke extraction
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F18/00—Pattern recognition
  • G06F18/20—Analysing
  • G06F18/25—Fusion techniques
  • G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
  • G06V30/10—Character recognition
  • G06V30/32—Digital ink
  • G06V30/36—Matching; Classification
  • G06V30/373—Matching; Classification using a special pattern or subpattern alphabet

Definitions

• the invention relates generally to the input of user information into computer systems, and more particularly to the recognition of handwritten characters input by a user.
• One of the biggest problems in handwriting recognition technology is reducing the error rate.
• One frequent type of error results when a user electronically enters a handwritten character, known as a chirograph, that closely matches two or more possible characters in a set to which the computer is trying to match the chirograph, i.e., a set of possible code points.
• Characters which cause the most errors are typically those which are identical to one another except for a single difference that humans can discern, but contemporary recognizers cannot. For example, certain Japanese symbols are substantially identical to one another but for a single, subtle difference.
• Another object is to provide a method and system of the above kind that can be automatically trained using sample data. Yet another object is to provide a method and mechanism of the above kind that is fast, reliable, cost- efficient, flexible and extensible.
• the present invention provides a method and mechanism for recognizing chirographs input into a computer system.
• a primary recognizer is provided for converting chirographs to code points, and secondary recognizers (e.g., CART trees) are developed and trained to differentiate chirographs which produce selected code points. Each such secondary recognizer is associated with each selected code point.
• secondary recognizers e.g., CART trees
• Each such secondary recognizer is associated with each selected code point.
• the chirograph is provided to the primary recognizer whereby a code point corresponding thereto is received.
• a determination is made as to whether the code point corresponds to one of the selected code points having a secondary recognizer associated therewith. If not, the code point provided by the primary recognizer is returned. If so, the chirograph is passed to the secondary recognizer, and a code point is returned from the secondary recognizer.
• FIGURE 1 is a block diagram representing a computer system into which the present invention may be incorporated;
• FIG. 2 is a block diagram representing functional components for training a primary handwriting recognizer according to one aspect of the invention
• FIG. 3 is a block diagram representing functional components for sorting chirographs as recognized by a primary recognizer into code point-based files to develop a secondary recognition system according to the present invention
• FIG. 4 represents the contents of an exemplary file sorted by the primary recognizer in FIG. 3;
• FIG. 5 is a flow diagram representing the general steps taken to sort the chirographs
• FIG. 6 is a block diagram representing functional components for generating the secondary recognition system from the file ⁇ of FIG. 3;
• FIGS. 7 - 9 comprise a flow diagram representing the general steps take to construct and train the secondary recognition system
• FIG. 10 is a block diagram representing functional components for optimizing the recognition mechanism of the present invention.
• FIGS. 11 - 13 comprise a flow diagram representing the general steps taken to optimize the recognition mechanism of the present invention
• FIG. 14 is a block diagram representing functional components for using the recognition mechanism of the present invention to recognize a chirograph
• FIG. 15 is a flow diagram representing the general steps taken when using the recognition mechanism of the present invention to recognize a chirograph.
• the computer system 20 includes a processor 22 operatively connected to storage 24, the storage including random access memory (RAM) 26 and non-volatile storage 28 such as a hard disk-drive, optical drive or the like.
• RAM random access memory
• non-volatile storage 28 such as a hard disk-drive, optical drive or the like.
• the non-volatile storage can be used in conjunction with the RAM to provide a relatively large amount of virtual memory via well-known swapping techniques.
• the processor 22 also connects through I/O circuitry 32 to one or more input devices 30, such as a keyboard and pointing device such as a mouse, and a pen-tablet, touch device or other means of getting electronic ink.
• the system 20 also includes at least one local output device 34 connected to the I/O circuitry 32 for communicating information, such as via a graphical user interface, to the user of the system 20.
• An operating system is loaded in the storage 24.
• those chirographs which often confuse a recognizer are provided to a secondary recognition process.
• a conventional (primary) recognizer outputs a code point. Instead of directly returning the code point, however, the code point is first examined to determine if it corresponds to a confusion set, i.e., one of two (or more) code points indicative of chirographs which are often confused for each other.
• the code point originally returned by the primary recognizer is returned by the mechanism.
• a secondary recognizer specifically developed to distinguish that particular confusion set, is given the chirograph.
• the secondary recognizer analyzes the chirograph using more directed tests than performed by the primary recognizer, and returns one of the two (or more) code points based on the results of the tests. Note that such often-confused chirographs are not limited to sets of two, but are often confused with two or more other chirographs.
• the primary recognizer can be trained to recognize shape classes that represent code points (or subsets of codepoints) that look alike. When provided with a chirograph, the primary recognizer thus returns at least one shape class index. The secondary recognizer then determines from the shape class index which code point the chirograph represents.
• a shape class index is a more general concept, i.e., a code point is a particular type of shape class index. However, for purposes of simplicity, the invention will be described with respect to a primary recognizer that returns code points, except where otherwise noted.
• a first aspect involves the development of the improved recognition mechanism of the present invention using handwriting sample data taken from a number (preferably a large number such as thousands) of users.
• a second aspect involves the use of a recognition mechanism, developed according to the first aspect of the invention, to convert a chirograph into a code point.
• the first aspect the development of the recognition mechanism, is ordinarily performed in a development environment on a relatively high powered computer system, which may be connected via a network connection or the like to large databases of sample data.
• the second aspect, the use of the recognition mechanism is typically performed on a hand-held (palm-top) computing device or the like.
• Such a device preferably runs under the Windows CE operating system loaded in the storage 24, and includes a touch- sensitive liquid crystal display screen for inputting handwritten characters (chirographs) .
• Other preferred systems include tablet-based desktop personal computers running under the Windows 95 or Windows NT operating systems .
• a first training set 40 of sample characters is used by a construction / training process 42 to develop and train a primary recognizer 44 (FIG. 3) .
• a training set is a file including chirographs stored in conjunction with their actual, correct code points, i.e., the code points identifying the character that the user intended to write.
• the primary recognizer 44 is preferably one which uses a nearest neighbor (KNN) approach.
• KNN nearest neighbor
• the recognizer 44 actually may return a (probability-ranked) list of alternative code points in response to a chirograph input thereto, however for purposes of simplicity the present invention will be described with reference to a single returned code point unless otherwise noted.
• the primary recognizer can also be of the type that returns any type of shape index, with shape codes used to train the primary recognizer. As can be appreciated by those skilled in the art, this technique will work equally well for any recognizer or pattern matching technique that returns discrete proposals .
• the primary recognizer 44 is used to begin constructing the secondary recognition mechanisms described above.
• a sorting process 47 sorts chirographs according to whatever code point (or other shape index) the primary recognizer 44 returns for that chirograph, whereby the way in which the chirographs are sorted with respect to their actual code points ultimately reveals the chirographs that the primary recognizer 44 tends to confuse.
• a second training set 48 containing sample chirographs, each stored along with its actual code point, is provided to the primary recognizer 44.
• the primary recognizer 44 Using its trained recognizer data 46, the primary recognizer 44 returns a code point to the sorting process 47, which sorts the chirographs and actual code points into various files 50 ⁇ - 50 n . Note that if other types of shape indexes are being used, the chirographs are similarly sorted into files for each shape index based on the shape index returned by the primary recognizer 44.
• the sorting process 47 first creates a separate file for each code point that is to be supported by the recognition mechanism.
• the first chirograph in the second training set 48 is selected at step 502, and sent to the primary recognizer 44 at step 504.
• a code point (which may, in fact, be incorrect) is returned by the primary recognizer 44 to the sorting process 47, and at step 508 written to the file that is associated with the returned code point, along with the actual code point known from the training set 48.
• steps 510 - 512 repeat the sorting process 47 with a subsequent chirograph until all chirographs in the training set 48 have been sorted in this manner. Note that the process is the same for shape indexes other than code points.
• FIG. 4 shows the contents of one such file 50 ⁇ , with two confused chirographs therein having different actual code points x and Ny.
• each distinct chirograph code points Nx or Ny
• the sorting process 47 wrote each chirograph and its actual code point into the file 50 ⁇ . Note that if the primary recognizer 44 made no mistakes, all of the files would contain only (chirograph, actual code point) pairs which matched the code point identifying the file. However, no primary recognizer has ever been found to have such accuracy when provided with suitably large training sets.
• CART Classification and Regression Trees
• CART trees are binary decision trees described in the text entitled Classifica tion and Regression Trees , Breiman, Friedman, Olshen and Stone, Chapman and Hall, (1984), and herein incorporated by reference in its entirety.
• one CART tree in a set 54 ⁇ - 54 n will be developed and trained for each code point (or shape index) supported.
• FIGS. 7 - 9 generally describe how each CART tree is developed.
• a list of questions which are believed to be relevant in distinguishing confusion pairs is assembled. Such questions are frequently based on handwriting strokes, such as, "how many total strokes in the chirograph?", w what is the length of the first stroke?" and/or ⁇ hat is the angle of the third stroke with respect to the first stroke?" .
• the questions may be tailored to the stroke count in the chirograph which is known to the system. As will become apparent, the order of the questions is not important.
• the primary recognizer may have provided some featurization information which the construction process can leverage in addition to its own featurization of the ink.
• the CART-building process 52 applies all of the questions to all of the samples (in each of the files 50 ⁇ - 50 n ) in order to determine and rank which questions best resolve the primary recognizer' s confusion for a given file.
• a preliminary test is performed by scanning the sample data at step 700 to determine if all of the actual code points in the given file are the same (and match the file) . If so, the data in the sample is pure, whereby secondary recognition will not improve the overall recognition. Accordingly, the CART-building process 52 terminates for such a sample file.
• the chirographs will have actual code points that do not directly match the code point (and thus the corresponding file) determined by the primary recognizer.
• the first question in the list is obtained, the first sample chirograph
• the question is applied to the sample, producing a result.
• the question may inquire as to the horizontal length of the first stroke, and result in a value of nine (highest x- coordinate minus lowest x-coordinate equals 9) for the first sample.
• the resulting value is saved in conjunction with the actual code point for that sample, e.g., (value, actual code point) at step 706, and at steps 708 - 710, the process repeated on the next sample in the selected file
• step 706 again saves whatever
• the steps of FIG. 8 are executed, in general to find out which of the values divides (splits) the chirographs in the file along the lines of their associated actual code points. It should be noted that it is possible, although generally impractical, to test every conceivable value with each question in a brute- force approach to determine the best split. For example, every length from 1 to 1000 may be tested for the length question, and so on with other wide ranges of values for the other questions. Instead, however, only the actual results obtained by the steps of FIG. 7 are used for this purpose, substantially speeding up the split-testing process of FIG. 8. Moreover, while each unique result can be applied as a binary question against each of the samples to determine the split, a more optimal way is to use the already existing result data to determine the best split.
• step 720 sorts the results obtained for the given question (in FIG. 7) into an ordered range of (values, actual code points) .
• the shortest lengths may have been forty and the longest one-hundred.
• each of the code points having a 'value equal forty" are moved into one (e.g., left) subset, and all code points having other values placed
• the quality of the split is evaluated according to some split criterion.
• a preferred way to determine the quality of the split is to test for homogeneity of the sets using the Gini diversity index.
• the Gini diversity index uses a sum of the squares method for the homogeneity (h) using the quantities of the code points in each of the left and right sets, i.e.,
• H(Q1, VI) [ (h Le ft) (cpiLeft + cp Left) + (hRight) (CpiRight + CP2Right) ]
• Step 726 tests the quality of the split against any previous results, if any, and if better, step 727 saves the homogeneity result H(Q1, VI) as the best quality split. Note that step 727 saves the best split over all the questions so far, including possibly the present question, whereby step 726 compares each subsequent split agains the result from the best (question, value) previously determined.
• Steps 728 - 730 cause the split for the next value to be tested and compared again, this time using the next value in the range, e.g., forty-one (41) .
• the sample is now effectively split with code points in the left subset being those having values less than or equal to forty-one. Note that the code-points associated with forty previously moved to the left subset remain there, since these are also less than forty-one.
• step 724 the next homogeneity H(Q1, V2) is computed, compared at step 726 (against the value for forty, which was the best so far) , and if an improvement, saved as the best value, along with the identity of its corresponding question, at step 727.
• the best value i.e., the value providing the most homogenous split
• the next question is selected (steps 732 - 734), and the process repeated on the samples in the file using this next question.
• the best (question, value) pair will continue to be saved for comparison against splits of other questions and values, and so on, until the overall best single (question, value) pair is known.
• step 742 the sample set (file 50 ⁇ ) is then split at step 742 into two subsets using this best question/value pair. Then, as represented by step 744, the process is iteratively repeated on each of these two subsets to find the next best question and value pair for most homogeneously splitting each of the subsets. The process is repeated recursively, (i.e., the process returns to step 700 of FIG.
• step 7 to optimally split each of the two subsets) , branching into more and more homogeneous subsets until a point is reached at which the homogeneity is no longer improved.
• the recursive operation at lower and lower levels establishes the best question/value pairs at each branch and level to further refine the distinction the confusion pairs.
• a CART tree is built from these question and value pairs at the various levels.
• the CART trees tend to be imperfect, especially at the lower levels. Moreover, the CART trees may be large, requiring a lot of storage that is not generally available in hand-held computing devices. Accordingly, at step 748 a new set of samples is applied to the CART to test which of its embedded questions are making the correct decisions. Those questions which are determined to be ineffective at resolving the confusion pairs are removed (pruned) from the tree at step 750. This leaves a more manageable CART in terms of size while not adversely affecting the recognition accuracy.
• CART tree does not improve the recognition some threshold amount (which may be even a very slight improvement) , there is no reason to keep it, since a CART tree costs storage space. Similarly, even though CART trees are extremely fast, secondary recognition using a CART tree adds to the total recognition time, again adding cost.
• FIGS. 10 and 11 - 13 represent one process for optimizing the recognition mechanism by discarding unneeded CART trees.
• a first chirograph from a third training set 56 (FIG. 10) is selected.
• the chirograph is sent to the primary recognizer 44.
• a primary recognizer match count 62 for the CART tree (i.e., this file) is incremented at step 906.
• the appropriate CART tree corresponding to the code point returned from the primary recognizer 44, is selected.
• step 920 the same chirograph is now provided to the CART tree, whereby a decision is made by the CART tree and a code point returned therefor.
• step 922 if the code point returned by the CART tree 52 is the same as the actual, correct code point, a CART match count 66 for this CART tree is incremented at step 924. Steps 926 - 928 repeat the process until all chirographs in the third training set 56 are tested.
• FIG. 13 compares the primary and CART match counts for each CART tree to determine if the CART tree improved the recognition. More particularly, the first supported code point (there is one file per each) is chosen at step 940, and the CART match count 66 for this CART tree compared against the primary recognizer match count 62. If the CART match count is less than or equal to the primary match count, the CART tree is discarded at step 944 since it did not improve the recognition mechanism. Otherwise the CART tree for this code point is kept. Steps 946 - 948 repeat the comparison until all supported code points have been tested. Note that if desired, step 942 can be a more complex test so as to discard any CART tree that does not improve the recognition process by some threshold amount.
• the CART tree can be discarded.
• the CART trees (which may number several hundred) only add about 18 kilobytes to a one megabyte primary recognizer, so any memory savings resulting from discarding a CART tree that only rarely improves recognition is probably not worth a reduction in recognition accuracy.
• the combined primary and secondary recognition mechanism of the present invention has been thoroughly tested, and for certain confusion pairs has a 99.7 percent accuracy rate. The 0.3 percent error rate is believed to result from characters too poorly written even for humans to discern, and in fact is comparable to the recognition error rate of humans. Note that the present invention is highly flexible and extensible.
• the recognition mechanism may be used in a relatively low powered system, e.g., a hand-held personal computing device.
• FIGS. 14 - 15 shown how the system is used to recognize a character.
• the system receives a chirograph 80 (FIG. 14) from a user at step 1100 in a known manner, such as via pen input on a touch-sensitive screen.
• the recognition mechanism of the present invention submits the chirograph to the primary recognizer 44 and receives a code point (or shape index) 82 therefrom (step 1102) .
• the code point 82 is used (by a lookup process 84 or the like) to determine if the code point has a CART tree associated therewith. If not, the primary recognizer's returned code point 82 is returned by the recognition mechanism at step 1108 as the returned code point 88.
• a CART tree is associated with the code point 82, the appropriate CART tree in the set of available CART trees 72 is selected and the chirograph 80 submitted thereto at step 1106.
• a shape index code that is not by itself a code point has a secondary recognizer (CART tree) associated therewith, even if only a minimal one that converts the shape index to a code point.
• the code point returned by the selected CART tree is returned at step 1108 as the returned code point 88.
• the recognition mechanism repeats until the user is done writing, as detected by step 1110.
• the list can be scanned for code points having associated CART trees, and the secondary recognizer operated for one or more of the code points in the list.
• the secondary process reorders the list with the result from the CART tree placed on top, i.e., with the highest probability.
• CART trees can provide alternatives ranked by probabilities, all of which can be weaved into a composite, probability-ranked list.
• a plurality of CART trees can be associated with a single character.
• a first CART tree can be provided as a secondary process for differentiating two-stroke *A" -shaped characters, and a second, distinct CART tree for differentiating three-or- more-stroke *A" -shaped characters.
• the primary recognizer can be arranged to split strokes, e.g., a one- stroke "A" shaped character can first be split into two strokes by the primary recognizer prior to its analysis thereof.
• stroke count may similarly be used by the primary and/or secondary recognizers.
• the points in the character are received as coordinates of the form (x, y, time), i.e, the points in sequence along with pen-up and pen-down positions are known.
• Off-line chirographs are only x-y points in no particular order.
• the invention is valuable in either type of recognition, although the primary and secondary recognizer (e.g., questions therefor) will be rather different.
• the method and mechanism differentiates ordinarily-confused characters with a high rate of success, and can be automatically trained using sample data.
• the method and mechanism that is fast, reliable, cost-efficient, flexible and extensible.

Landscapes

• Engineering & Computer Science (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• Multimedia (AREA)
• Data Mining & Analysis (AREA)
• Bioinformatics & Cheminformatics (AREA)
• Bioinformatics & Computational Biology (AREA)
• Artificial Intelligence (AREA)
• Evolutionary Biology (AREA)
• Evolutionary Computation (AREA)
• Life Sciences & Earth Sciences (AREA)
• General Engineering & Computer Science (AREA)
• Character Discrimination (AREA)

Priority Applications (2)

Applications Claiming Priority (2)

Publications (1)

Family

ID=25355649

Family Applications (1)

Country Status (5)

Families Citing this family (26)

Citations (1)

Family Cites Families (17)

• 1997
  • 1997-06-06 US US08/870,559 patent/US6061472A/en not_active Expired - Lifetime
• 1998
  • 1998-06-04 WO PCT/US1998/011642 patent/WO1998055958A1/en not_active Ceased
  • 1998-06-04 AU AU78194/98A patent/AU7819498A/en not_active Abandoned
  • 1998-06-04 JP JP50292899A patent/JP4233612B2/ja not_active Expired - Fee Related
  • 1998-06-04 CN CNB988011077A patent/CN1163840C/zh not_active Expired - Fee Related
• 2000
  • 2000-03-20 US US09/528,889 patent/US6973215B1/en not_active Expired - Fee Related
• 2005
  • 2005-12-05 US US11/294,260 patent/US7379597B2/en not_active Expired - Fee Related

Patent Citations (1)

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events