Stroke-Number-Free and Stroke-Order-Free on-Line Chinese Character Recognition Method

The present invention relates to an improved on-line Chinese character recognition method without constraints on either the stroke number, the stroke order, or both. More specifically, the present invention relates to a stroke-number-free and stroke-order-free on-line handwritten Chinese character recognition method which is free from the constraints of either the stroke number, the stroke order, or both, of the input script to be recognized, so as to accommodate the wide variations of handwritten Chinese characters that have been accrued over the years.

Recently, pen-based approach has attracted great attention in providing human-computer interfaces, especially as an input means. The pen-based approach is similar to the traditional way of using a pen and paper, by which characters or commands are entered into the computer by writing. This approach is particularly appealing for use with Chinese language-based applications because the Chinese language contains a very large number of character categories, and it is very cumbersome to enter Chinese characters via a key board. One of the key elements in providing a pen-based system for Chinese language-based computer applications is to develop an accurate and reliable on-line handwritten Chinese character recognition (OLCCR) method.

One of the major problems in the implementation of an on-line Chinese character recognition system is that there are multitudinous categories of Chinese characters, some with very complicated structures, and that there exist very wide variations in each Chinese character. Chinese characters may be written with large variations in the stroke order, stroke number, as well as shape distortions, and are still readily recognizable by the human eyes. Most of the character recognition methods for the Chinese characters have constraints on the stroke order and/or the stroke number imposed on them, in order to achieve high recognition accuracy on a real-time basis. It is desirable to develop Chinese character recognition methods that can relax or eliminate these constraints.

Wakahara and Umeda proposed a selective stroke linkage method for recognizing Japanese/Chinese characters. This method relieved the constraint on the stroke number; however, the constrain on stroke order still exists with regard to connected strokes. When a hand-written character shows significant pattern distortions, this method often fails to determine the correct stroke correspondences. Subsequently, they also proposed a method of stroke linage rules, based on which possible cursive patterns are generated as the reference patterns for character matching. Each character category may contain more than one cursive pattern; this could cause a substantial increase in the computational time required.

Lin et al. proposed a model-based character recognition approach, using stroke-based features of the characters. They constructed a deviation-expansion model to represent the reference pattern of a character. The reference pattern contains hypothetical knowledge of writing variations. During pattern matching, a matching tree is constructed by combining the knowledge of the reference pattern and input pattern (i.e., input script) together, and the dynamic programming matching (DP) technique and A* algorithm are adapted to obtain the recognition result. In this method, if certain deviations of the handwritten characters are not included in the hypothetical knowledge of writing variations, recognition errors would result. In the worst case, the character matching process can require exponential-time complexity.

One approach to cope with the variations in stroke order is to use the stroke re-ordering strategy. In this approach, the strokes of an input character are rearranged into a pre-determined order based on certain features or pre-defined rules. Ye et al. disclosed a method by which the stroke order of an input character is rearranged by sorting the input strokes according the stroke types. Hidal et al. disclosed a method by which the slants and locations of the strokes are utilized to obtain a stroke correspondence between the input script and a pre-defined template. The stroke re-ordering can be acquired from the result of stroke correspondence. However, due to significant variations in handwritings, there might be deformations in the stroke shape and the connected strokes may be mistakenly recognized as corresponding to another type of stroke or rejected. But most importantly, results from the technique of stroke re-ordering typically are not stable for handwritten characters with potentially large variations.

Ishii, by analyzing the stability of structural relationships among the strokes in handwritten Chinese characters, proposed a method which utilized representative points of strokes for decomposing radical(s) from a character and re-ordering the input strokes. In this method, each character is assigned a protocol for stroke re-ordering which was derived from handwritten samples. The resultant re-order strokes are more stable than the previous methods of re-ordering. Shiau et al. extended the types of representative points and applied Ishii's method to 5,401 character categories. Both methods, however, contain a constraint on fixed stroke numbers.

To cope with the variations in both the stroke number and stroke order, Liu and Tai also adopted the stroke re-ordering strategy. They utilized line segments as character primitives, and proposed a method which involved general rules for 17 types of spatial relationships between any two line segments in the Chinese characters. The stroke re-ordering strategy proposed by Liu and Tai has the advantage that the space required to store the general re-ordering rules is typically substantially less than that required to store separate re-ordering rules for each character. However, the re-ordering strategy developed by Liu and Tai was found to be unstable, especially for those input characters with large structural distortions.

The primary object of the present invention is to develop an improved on-line method for recognizing handwritten Chinese characters. More specifically, the primary object of the present invention is to develop a novel approach which allows a stroke number- and stroke order-free method for on-line recognition of handwritten Chinese characters to be developed. The method developed in the present invention utilizes a rule-based approach without having constraints on either the stroke number or the stroke order.

Chinese characters are constructed based on a number of basic strokes according to certain structural rules. The present invention thus takes advantage of these structural rules for machine recognition. The method disclosed in the present invention is based, as a part thereof, on the standard patterns of the Chinese characters. To accommodate the handwritten characters, each character "category" in the reference database associated with the character recognition system disclosed in the present invention further includes several handwriting variations (of the same character), such as shape distortion and variations in stroke order and stroke number. The character recognition system also contains certain rules which are predefined in a reference database to effectuate the stroke correspondence (i.e., stroke recognition) between a template character and the input character script to avoid the need for repeating combinatorial exhaustion. For each stroke contained in a template character, one rule is utilized to find its possible matching stroke in the input script.

In the method disclosed in the present invention, basic stroke types and the so-called "geometric feature measures" of strokes are utilized in the design of stroke correspondence rules. When an input character script is entered and to be recognized, all possible strokes in the input character are identified (i.e., isolated) before the invocation of the stroke correspondence rules. Thus, in the method disclosed in the present invention, a constituent stroke is selected, one at a time, from the input script. The selected input stroke is matched against the template strokes, using the stroke correspondence rule for each template stroke, until a matching stroke, if it exists, is found. The connected strokes are segmented, i.e., disconnected, according to their basic stroke types, and the geometric feature measures of the strokes become the invariant features of the strokes in 2-dimensional handwritten characters. By using these invariant features, which are based on the geometric feature measures of the strokes, handwritten characters with wide variations in both the stroke number and stroke order can be successfully recognized. In implementing the method disclosed in the present invention, the computation time required for finding the stroke correspondence between a template character and an input script is in the order of O(n), where n is the stroke number of the template. This is a very distinct advantage over most of the methods disclosed in the prior art, which typically requires a computation time that increases exponentially with the number of template characters. The templates selected during the preliminary classification are taken to match against the input script pattern, stroke by stroke, in the manner described above. The input character is finally identified by using suitable discrimination functions as the template character with the minimum distance (distance is a measure of error between the template character and the input script). When there are more than one template with the same minimum distance, the input character is further identified by traversing a logical tree composed of special structural features.

The method disclosed in the present invention comprises a recognition architecture illustrated in FIG. 1. The recognition architecture includes a set of processing modules and an accompanying database. Each handwritten input script character is first processed using the preprocessing stage. After the basic strokes are identified, the estimated range of possible stroke number and the statistical features of the input character are utilized in the preliminary classification to reduce the number of template characters to be processed in the later stage of structural analysis as disclosed in the present invention. The preprocessing stage is documented in the prior art, and thus will not be elaborated. The processing steps of structural analysis include: (1) stroke correspondence, (2) computation of discrimination functions, and (3) detail recognition. Chinese characters can be written (handwritten or printed) in various "styles" (a style can be considered as equivalent to a "font"). The architecture described in FIG. 1 can be applied to recognize Chinese characters written in various styles. When the style of character to be recognized changes, only the correspondence rules in the reference database that are associated with the strokes need to be changed.

In a preferred embodiment of the present invention, the strokes of the Chinese characters are classified into 14 basic stroke types, as shown in Table 1. A stroke template database is constructed for the recognition system which contains collections of sampled "points" for each stroke obtained from a digitizing tablet using a standard template and several handwriting variations thereof Each sampled point is described by a vector (x, y, f), wherein x and y are the coordinates of that point and f is a flag, which has a value of "on" or "off", to indicate whether the pen is pressed or lifted. An input stroke contains sequentially contiguous points all with "on" values and terminates at a point with the "off" value. The preprocessing steps include normalization, sampling, and line-segment approximation. Each stroke, or more specifically each basic stroke (i.e., isolated stroke), is represented by a set of string(s) for its underlying line segment(s) in the basic stroke database, which contains variations of the basic strokes to accommodate the handwriting variations in stroke shapes, etc. Each input stroke is approximated by a set of line segment(s) represented by a string, and is recognized by a string matching algorithm.

In the method disclosed in the present invention, before the determination of stroke correspondence, it is necessary to separate a connected stroke into individual basic strokes. This process is called "segmentation". All possible basic strokes existing in a cursive input stroke are recognized which constitute the candidates in the invocation of the stroke correspondence procedure. FIG. 2a illustrates the "standard pattern", and FIGS. 2b and 2c show two different "cursive patterns", respectively, of the character . The two cursive patterns shown in FIGS. 2b and 2c vary in both the stroke number and stroke connections relative to the standard pattern shown in FIG. 2a. FIG. 2a also shows the basic strokes of the standard pattern for the character . All the possible basic strokes that are contained in the cursive patterns are also shown in FIGS. 2b and 2c. The input cursive character shown in FIG. 2b contains three strokes (as described above, a stroke is considered as consisting of a collection of sequentially connected "on" points), which are composed, collectively, of eight possible basic strokes in a variety of combinations, and the input cursive character shown in FIG. 2c contains only one stroke, from which a total of 11 possible basic strokes can be independently extracted.

In the present invention, to handle the unique structural complexity of the Chinese characters, a Chinese character is considered to be represented by one or more "primitives", which are further classified into three types: fore strokes, back strokes, and points. Fore strokes are those strokes that actually appear in a character pattern (which can be either a template character or an input script). In the standard templates, the fore strokes should not be connected. However, in typical Chinese handwritings, the fore strokes are often connected by "pseudo segments"; these pseudo segments are called "back strokes" in the present invention, meaning that these strokes should not exist in the standard character. In a broader sense, the back strokes are provided in the present invention to allow for imaginary strokes connecting between fore strokes. When a connecting stroke exists as a result of hasty handwriting, a fore stroke may appear in the handwritten script which would correspond to a back stroke in template (i.e., the stroke does not exist in the template). This is shown as stroke "b" shown in FIG. 2b. On the other hand, a back stroke (the imaginary stroke connecting two fore strokes) may also be "simplified" and thus degenerated into an intersection "point", such as the intersection points of a' and b' shown in FIG. 2c. By these designations, the primitives in the character representation can be classified into the three types of fore strokes, back strokes, and points. The designation of these three primitives disclosed in the present invention greatly enhances the efficiency of the character recognition process, and allows clear explanations to be conveyed to the end-users.

The task of stroke matching involves finding a binary relation (designated as "q") between the set of template primitives and the set of input primitives. The binary relation q for stroke matching from set X to set Y is defined as follows:

where X denotes the set of primitives of a template character, and Y denotes the set of primitives of an input character. For any element x.sub.i .epsilon.X, if the mapped image (i.e., a matched primitive) y.sub.i (y.sub.i =q(x.sub.i), y.sub.i .epsilon.Y) exists, then there exists only one such image y.sub.i.

Among the fore strokes, back strokes and points defined in the present invention, there are eight possible types of matches, namely: (1) fore.fwdarw.fore, (2) back.fwdarw.back, (3) back.fwdarw.fore, (4) back.fwdarw.point, (5) back.fwdarw.null, (6) null.fwdarw.back, (7) fore.fwdarw.null, and (8) null.fwdarw.fore, where x.fwdarw.y indicates a matched pair, and fore represents a fore stroke primitive, back represents a back stroke primitive, point represents a point primitive, and null indicates no matched primitive. It should be noted that in this preferred embodiment, the point primitive only exists in the input script. Furthermore, a back primitive (i.e., an imaginary stroke) in the template can be corresponded to a fore stroke, a back stroke, a point in the input script, or no match. This illustrates the wide variations of Chinese handwritings.

The fore strokes of a standard pattern (i.e., a template pattern) are the essential primitives of a character and, generally speaking, they should always exist in the input handwritten patterns, especially for those long or major strokes. A standard pattern is described as a template in the reference database. When strokes are connected in the handwritten characters, back strokes in the templates may appear as fore strokes or may even be reduced to "points" in the input patterns (i.e., scripts). These variations are commonly observed in the handwritten Chinese characters, and thus are accommodated in the method of the present invention for character matching. During the matching process, an input script is matched against the templates of candidate characters selected during preliminary classification. For explaining in the process of the present invention, it can be considered that an input script is matched against a template of the input category, which contains possible variations of a character. The fore strokes in the template characters are first matched with the fore strokes in the input script. The back strokes in the template pattern are obtained by tracing the matched pairs following the input stroke order, which can be obtained as an "on-line" recognition process. A back stroke in the template character may be mapped to a back stroke, a point, or a fore stroke in the input script. FIG. 2b shows the stroke mapping between the standard template and an exemplary cursive pattern. In FIG. 2b, the matched pairs .alpha..fwdarw.a, .beta..fwdarw.c, .gamma..fwdarw.d, and .delta..fwdarw.e are the matched fore strokes. A back stroke (i, j) is identified between the two fore strokes "i" and "j". Thus, the back strokes (.alpha., .beta.), (.beta., .gamma.), and (.gamma., .delta.) are identified after the above matched pairs of fore strokes are found. By tracing the matched pairs of the fore strokes following the input stroke order, the back strokes (.alpha., .beta.), (.beta., .gamma.), and (.gamma., .delta.) can be mapped to b, (c, d), and (d, e), respectively. FIG. 2c shows another example of such mapping process. In the steps of priliminary identification just described, the back strokes are not contained in the reference template database, rather, they are derived from list of mapping pair of fore strokes. The required information used in such recognition is stored in the database of spatial relationships between strokes. This will be discussed in subsequent sections.

FIGS. 3a-3c illustrate the process of applying the stroke correspondence rule using the examples shown in FIGS. 2a-2c. FIG. 3a shows the stroke correspondence rules for the character . The stroke correspondence rules consist of (1) basic stroke types that are "allowed" (i.e., commonly encountered) in the handwriting, and (2) the geometric feature measure(s) of the strokes used in the stroke matching process. Each character category (which includes variations of a standard character) has a set of rules so as to accommodate a number of possible variations. The sequence of applying the rules coincides with the sequence of finding the matched pairs in a template. A handwritten input stroke may consist of one or more basic strokes due to stroke connections (i.e., connections between the basic strokes as a result of handwriting). These connected strokes in the input script are first segmented. The process of stroke correspondence is to find a possible corresponding stroke in the input script for each of the rules (strokes) in the template character. If a possible basic stroke in the input script is found to match with one rule, then this selected basic stroke is removed from the input pattern, and the next stroke correspondence rule is applied onto the remaining strokes of the input pattern until all the rules in the template are exhausted. It is quite likely that some rules in the template character may not find matched strokes in the input character. FIGS. 3b and 3c show the process of applying the stroke correspondence rules for the two cursive patterns in FIGS. 2b and 2c, respectively.

The present invention utilizes the geometric feature measures, which are geometrically related characteristic measures (which can be x and/or y coordinates, lengths, or associated distances) of the strokes, to establish the stroke correspondence rules for character recognition. Chinese characters contain two-dimensional arrangements of basic strokes based on predefined and recognizable geometric configurations. By using one or more of such geometrically related characteristic measures, the variations in stroke order of the handwritten characters can be accommodated in the method disclosed in the present invention.

In a preferred embodiment of the present invention, 27 types of geometric feature measures of strokes are utilized to establish the stroke correspondence rules. These geometric feature measures are listed in Table 2. Before applying the stroke correspondence rules, all possible basic strokes are sorted based on the 27 measures in an increasing order. The sorted sequence for each measure is recorded in an array. Each stroke correspondence rule would contain the information that the matching stroke must have a maximum or minimum value of certain types of geometric feature measures. The matching stroke can then be directly acquired from the array storing the sorted results.

These 27 types of geometric feature measures are summarized in Table 2 and are explained in more detailed descriptions that follow. First, each stroke is considered to be bounded by a minimum bounding rectangle (MBR). The x and y coordinates of the four boundaries and the center point of the MBR of a stroke are designated as geometric feature measures numbered 1 through 6, respectively, i.e., MBR.sub.-- Xmin, MBR.sub.-- Ymin, MBR.sub.-- Xmax, MBR.sub.-- Ymax, MBR.sub.-- Xcenter, and MBR.sub.-- Ycenter. The x and y coordinates of the starting point and end point of the strokes are designated as geometric feature measures numbered 7 through 10, i.e. Sx, Sy, Ex, Ey, respectively. These geometric feature measures are often adequate for relatively simple characters. For complicated characters, additional or auxiliary geometric feature measures can be provided to facilitate speedy and accurate stroke correspondence.

FIG. 4 shows the eight reference points of the bounding rectangle of a character. These eight reference points are: LT (left-top corner point of the rectangle), TC (top-center point of the rectangle), RT (right-top corner point of the rectangle), LC (left-center point of the rectangle), RC (right-center point of the rectangle), LB (left-bottom corner point of the rectangle), BC (bottom-center point of the rectangle), and RB (right-bottom corner point of the rectangle). From these eight reference points, 14 auxiliary geometric feature measures, numbered from 11 through 24, can be derived. They are defined as the Euclidean distances from the starting and end points of a stroke, and the center point of a stroke's MBR to these eight reference points, respectively. Additionally, the length of a stroke is considered as a geometric feature measure, numbered 25. Finally, the geometric feature measure numbered 26 is defined as the "city block" (or block distance) from the left-bottom corner point of a character to the left-bottom corner point of an individual stroke's MBR, and the geometric feature measure numbered 27 is defined as the city block from the left-bottom corner point of a character to the right-top corner point of that same stroke's MBR. Some, if not all, of these additional measures are useful for ensuring stable stroke correspondence for some complicated cases. As it will be discussed below, most of the stroke correspondence rules developed in the present invention require only one geometric feature measure. In other rules, two geometric feature measures are utilized.

After all the possible basic strokes in an input character are identified, the 27 geometric feature measures for each possible basic stroke are computed. All the possible basic strokes are respectively sorted based on the values of the geometric feature measures in an increasing order. During the stroke correspondence step, the matching stroke is selected from these sorted possible basic strokes. Therefore, the computation time required for performing stroke correspondence is in the order of O(n), where n is the number of strokes of the matched template.

In a preferred embodiment of the present invention, stroke correspondence rules are classified into two types: type one rules and type two rules, denoted as R1 and R2, respectively. Both types of rules also contain the information of the possible basic stroke types which are allowed (i.e., often encountered or expected) in the handwriting variations. By using basic stroke types, those strokes in the input script that are impossible as candidates based on stroke shapes are eliminated. The basic stroke information can also be used to dissect connected strokes into separate strokes. A type one rule utilizes one predefined geometric feature measure of strokes to find the matching stroke. In some cases, more information is required to find stroke matching. A type two rule utilizes two geometric feature measures to find the matching stroke. The first geometric feature measure is used to select candidate strokes from the input stroke. Then the second geometric feature measure is further utilized to determine the matching stroke from the candidate strokes.

The application of the stroke correspondence rules can be illustrated by FIGS. 5a and 5b. FIG. 5a shows a standard pattern (template) of character , with its strokes identified and labeled. FIG. 5b lists the stroke correspondence rules. Each stroke correspondence rule contains the rule type, the basic stroke types allowed in considering handwriting variations, and the geometric feature measure(s) of the strokes to be used to find the matching stroke. In FIG. 5b, the geometric feature measure is denoted as Fi (1<i<27), where i is the geometric feature number (i.e., the type of geometric feature measure). For a template stroke, the matching stroke must have the minimum (denoted as S0) or maximum (denoted as S1) value of certain geometric feature measure Fi in the possible basic strokes of the input script. A type one rule utilizes one geometric feature measure denoted by (Fi, Sj) (1<i<27 and j=0 or 1). A type two rule, on the other hand, utilizes two geometric feature measures denoted by (Fi, Sj, Ck; Fi', Sj'), where the meanings of i, i', j, j' are similar to those defined for type one rules, and the value Ck indicates that k candidate strokes, of the basic stroke types allowed in the rule, have been selected from the remaining possible basic strokes of the input script using the first geometric feature measure. The second geometric feature measure is used to determine the matching strokes from the k candidate strokes. The order of applying the stroke correspondence is based on a predefined rule sequence. The rules of strokes .alpha. and .beta. in FIG. 5b are taken as examples for explanation. The rule of stroke .alpha. belongs to type two, and stroke type 6 is allowed. Based on the first geometric feature measure, (which, in this case, is the stroke length), strokes .alpha. and .delta. are selected as the candidate stokes from FIG. 5a. Next, the matched stroke, with start point of minimum x coordinate (F7), is selected from the two candidate strokes. The handwritings of strokes .beta., .gamma., .epsilon., and .eta. in the character , are very close to "written dots", or short strokes. A handwritten short stroke is usually unstable and the stroke type recognized often is not the correct (i.e., original) stroke. Therefore, a number of possible variations (in this case, stroke types 0, 1, 2, and 3) should be taken into consideration and are accommodated in the rule. The rule for stroke .beta. belongs to type one. The only geometric feature measure is the Euclidean distance from the center point of the stroke's MBR to the left-top corner point of the character. From the remaining strokes with the allowed stroke types in the rule, the matched stroke is the stroke with the least value of the geometric feature measure.

After the stroke correspondence finds the matched strokes, the method disclosed in the present invention then utilizes the spatial relationships between strokes, as well as the information of lost strokes, superfluous stokes, and connected strokes to compute the character distance between a template character and an input script, for character discrimination. After the stroke correspondence between a template character and an input character script is accomplished, the matching relation between template strokes and input strokes is then considered to have been obtained. Based on the matching relation, input strokes can be re-ordered following the stroke order of the template character. For an input script, the spatial relationships between the strokes based on the template stroke order are used in character discrimination.

As discussed above, in the present invention, the matching relation between the template and input primitives is a binary relation q: X.fwdarw.Y, where X denotes the set of template primitives and Y denotes the set of primitives of the input script. In other words, one of the key elements in the stroke correspondence process is to exhaust the rules (i.e., stokes) of the template character. When the input category is the same as the template category, the matching types of fore stroke.fwdarw.fore stroke and back stroke.fwdarw.back stroke are considered correct matchings for fore strokes and back strokes; the matching types of back.fwdarw.point and back.fwdarw.fore can exist in the case of connected strokes; the matching types of back.fwdarw.null and fore.fwdarw.null occur in the case of lost stroke(s); and the matching types of null.fwdarw.back and null.fwdarw.fore can occur in the case involving superfluous stroke(s).

The distance functions used for character discrimination are defined based on the matching type(s), and are explained below. For the matching type of fore.fwdarw.fore, the matched pairs of fore strokes can be re-ordered according to the stroke sequence in the template character. In the input script, the spatial relationships between the fore strokes i and i+1 are described by four vectors: ss(i, i+1), se(i, i+1), es(i, i+1), and ee(i, i+1). These vectors are defined as connecting between the starting point ("s") and the end point ("e"), from stroke i to stroke i+1, as illustrated in FIG. 6. In performing the computation, the directions are quantized to accommodate variations in the handwritings. Thus, function quant.sub.-- 8( ) is defined to quantize the directions of the vectors into eight direction codes. All the possible variations of spatial relationships between handwritten strokes i and i+1 are all accommodated in the reference database, and are denoted by four sets of eight-direction codes: SS(i, i+1), SE(i, i+1), ES(i, i+1), EE(i, i+1). For the matched pairs of fore.fwdarw.fore, the cost function ff.sub.-- err.sub.-- ss(i) is defined as: ##EQU1## The other cost functions ff.sub.-- err.sub.-- se(i), ff.sub.-- err.sub.-- es(i), ff.sub.-- err.sub.-- ee(i) are similarly defined (i.e., substituting ss(i, i+1) to appropriate vectors). Assuming that the template character has FM strokes, for the matched pairs of the type fore.fwdarw.fore, the distance function related to the spatial relationship between the fore strokes is defined as: ##EQU2## For the matched pairs of the type fore.fwdarw.null, the distance function is defined as: ##EQU3## where FN is the number of matched pairs of the type fore.fwdarw.null, T.sub.-- STK is the stroke number (i.e., number of strokes) of the template character, l.sub.i is the length of the fore stroke i, and L.sub.T is the summation of stroke lengths in the template. The error cost of one lost stroke is considered as having a weight of 5, which includes the error cost of one fore stroke and four vectors representing the spatial relationships between two contiguous strokes.

For the matched pairs of the type null.fwdarw.fore, the distance function is defined as: ##EQU4## where NF is the number of matched pairs of the type null.fwdarw.fore, I.sub.-- STK is the middle value of the estimated range of input stroke number, l.sub.j is the length of the fore stroke .sub.j, and L.sub.I is the total length of all the strokes in the input character.

The matched pairs of the type back.fwdarw.point occur in the case of connected strokes. The distance function is defined as: ##EQU5## where BP is the number of matched pairs of the type back.fwdarw.point.

The similarity between the template and the input characters is described by the following distance formula:

The candidate character with the minimum distance has the maximum probability of being the correctly recognized character for the input character. When the number of candidate characters with the minimum distance is more than one, then these candidate characters constitute a "similar group". When this occurs, the input character is further recognized by the special structural features of the character in the similar group which are represented by a logical tree. For example, the difference of stroke lengths is used as the discrimination feature in the detailed recognition for characters () and (); whereas the discrimination function for the characters () is whether a "hook" exists in the curved stroke.

The present invention will be described in detail with reference to the drawings showing the preferred embodiment of the present invention, wherein:

FIG. 1 is a schematic diagram showing the recognition architecture disclosed in a preferred embodiment of the present invention.

FIGS. 2a-2c show all the possible basic strokes existing in the "standard pattern", and two different "cursive patterns", respectively, of the character .

FIGS. 3a-3c illustrate the steps of applying the stroke correspondence rules using the examples shown in FIGS. 2a-2c.

FIG. 4 shows the eight reference points of the bounding rectangle of a character in defining geometric feature measures numbered from 11 through 24.

FIGS. 5a and 5b show an example of the application of the stroke correspondence rules for the Chinese character .

FIG. 6 illustrates the spatial relationships between the fore strokes i and i+1 as described by four vectors: ss(i, i+1), se(i, i+1), es(i, i+1), and ee(i, i+1); these vectors are defined as connecting between the starting point ("s") and the end point ("e"), from stroke i to stroke i+1.

The present invention discloses a stroke-order-free and stroke-number-free on-line method for recognizing handwritten Chinese characters. The method disclosed in the present invention mainly comprises the following steps:

(a) obtaining a database of template characters, each template character being represented by a set of stroke correspondence rules for describing its constituent basic strokes;

(b) obtaining a database of spatial relationships between strokes of characters and a database of character patterns;

(c) inputting a handwritten input script on an on-line basis;

(d) preprocessing the input script to select candidate template characters for matching against the input script;

(e) performing a basic stroke recognition procedure to identify all basic strokes contained in the input script using a database of basic strokes;

(f) classifying the strokes into fore strokes, back strokes, and points, wherein the fore strokes are strokes that actually appear in the character, the back strokes are fictitious strokes to allow for stroke connections that should not appear, and points are provided to allow for truncated back strokes in the input script;

(g) for each candidate template character, performing a stroke correspondence for each stroke correspondence rule until all the stroke correspondence rules contained in the template character are exhausted, so as to identify all fore stroke.fwdarw.fore stroke correspondence;

(h) performing stroke matchings according to the input stroke order to find other stroke correspondences, i.e., back.fwdarw.point, back.fwdarw.fore, fore.fwdarw.null, null.fwdarw.fore, back.fwdarw.null, and null.fwdarw.back; and

(i) performing computation of discrimination functions to find a template character with a minimum distance.

One of the key elements of the present invention, in addition to the above-listed steps, is that each stroke correspondence rule contains a specific set of stroke information including allowed stroke types, at least one geometric feature measure, and criterion for applying such geometric feature measure; the geometric feature measure is a geometrically related characteristic measure (which can be x or y coordinates, length, distance, etc) associated with a particular stroke, so as to facilitate the stroke correspondence process.

The method disclosed in the present invention may preferably include the step of performing preliminary classification which can be achieved using an estimated range of the stroke number of the input script and a database of statistical features. While this step involves stroke number, this is done only in a very coarse sense, and variations in the stroke number can be accommodated. And the present invention is still classified as a stroke-number-free method.

These steps are described in more detail below.

(a) Obtaining a database of template characters, each template character being represented by a set of stroke correspondence rules for describing its constituent basic strokes:

The method disclosed in the present invention employs a novel rule-based approach for finding stroke correspondence. The stroke correspondence rules employed in the present invention consist of (1) basic stroke types that are allowed in the handwriting, and (2) at least one geometric feature measure associated with the stroke. Each character category (which includes variations of a standard character) has a set of rules, each corresponding to a stroke. The sequence of applying the rules is the same as the sequence of finding the matched pairs in a template. A handwritten input stroke may consist of one or more basic strokes due to stroke connections (i.e., connections between the basic strokes as a result of handwritings). The matching of a template stroke is completed when a possible basic stroke in the input script is selected after it meets one rule. Then, this possible basic stroke is removed from the input pattern, and another application of the stroke correspondence rule is performed onto one of the remaining strokes of the input pattern until all the rules of the template are exhausted. It should be noted that not all of the strokes (expressed as rules) will find their matched fore strokes in the input script. After the rearrangement of the fore strokes in the template character based on the fore stroke matching, other types of matches involving non-fore strokes will be found.

The present invention utilizes one or more of the geometric feature measures, which are geometrically related characteristic lengths of the strokes, to establish the stroke correspondence rules for character recognition. Chinese characters contain two-dimensional arrangements of basic strokes based on predefined and recognizable geometric configuration. By using one or more of such geometrically related characteristic measures, the variations in the stroke order of the handwritten characters can be accommodated in the method disclosed in the present invention.

In a preferred embodiment of the present invention, 27 types of geometric feature measures of strokes are utilized to establish the stroke correspondence rules. The use of the geometric feature measures disclosed in the present invention allows on-line handwritten Chinese characters to be recognized without constraints on either the stroke number or the stroke order, or both. Before applying the stroke correspondence rules, all possible basic strokes are sorted based on the 27 measures in an increasing order. The sorted sequence for each measure is recorded in an array. Each stroke correspondence rule contains the information that the matching stroke must have a maximum or minimum value of a certain type of geometric feature measure. The matching stroke can then be directly acquired from the array storing the sorted results.

In a preferred embodiment of the present invention, after all the possible basic strokes in an input character are identified, all the 27 geometric feature measures for each possible basic stroke are computed. All the possible basic strokes are respectively sorted based on the values of the geometric feature measures in an increasing order. During the stroke correspondence step, the matching stroke is selected from these sorted possible basic strokes. Therefore, the computation time required for performing the stroke correspondence is in the order of O(n), where n is the number of strokes of the matched template.

For some cases, one geometric feature measure is adequate for stroke correspondence. However, in some other cases, typically those involving relatively complicated characters, it is highly desirable to use two or more geometric feature measures, typically two will suffice. Thus, in a preferred embodiment of the present invention, stroke correspondence rules are classified into two types: type one rules and type two rules, denoted as R1 and R2, respectively. Both types of rules also contain the information of the possible types of basic stroke which are allowed in the handwriting variations. A type one rule utilizes one predefined geometric feature measure of strokes to find the matching stroke. In some cases, more information is required to find stroke matching. A type two rule is required which utilizes two geometric feature measures to find the matching stroke; the first geometric feature measure is used to select candidate strokes for the input stroke, then the second geometric feature measure is further utilized to determine the matching stroke from the candidate strokes.

(b) Obtaining a database of spatial relationships between strokes of characters:

A database of spatial relationships between the fore strokes is provided in which the spatial relationships between fore strokes i and i+1 are described by four vectors: ss(i, i+1), se(i, i+1), es(i, i+1), and ee(i, i+1). These vectors are defined as the connections between the starting point ("s") and the end point ("e"), from stroke i to stroke i+1, as illustrated in FIG. 6. The database of spatial relationships between the fore strokes is particularly useful for computing the distance between a template character and an input script.

(c) Inputting a handwritten input script on an on-line basis:

During the input step, the input script is converted into a collection of sampled points, and each sampled point is described by a vector (x, y, f), wherein x and y are the coordinates of that point and f is a flag, which has a value of "on" or "off", to indicate whether the pen is pressed or lifted. An input stroke is defined to contain sequentially contiguous points all with "on" values and terminates at a point with the "off" value.

(d) Preprocessing the input script to select candidate template characters for matching against the input script:

After the input step, the input script is processed with a preprocessing step, which includes normalization, sampling, and line-segment approximation. Each input stroke is approximated by a set of line segment(s), represented by a string, and is recognized by a string matching algorithm. The string matching algorithm is well-known in the art, and thus will not be described here.

(e) Performing a basic stroke recognition procedure to identify all possible basic strokes contained in the input script using a database of basic strokes:

After the preprocessing step, all possible basic strokes that may exist in the input script are identified. Because the basic strokes in a handwritten Chinese character are often connected, an apparent stroke may actually contain several basic strokes. Thus, all of the possible basic strokes (i.e., collectively, partially collectively, as well as individually) are identified. As described above, the apparent stroke in an input script is considered as consisting of a collection of sequentially connected "on" points). Because of the unique structural complexity of the Chinese characters, the strokes are classified in the present invention into three different types of "primitives", i.e., fore strokes, back strokes, and points. Fore strokes are those strokes that actually appear in a character pattern (which can be either a template character or an input script). In the standard templates, the fore strokes should not be connected. However, in reality, the fore strokes are often connected, and the "pseudo segments" may appear as fore strokes. In the present invention, these pseudo segments that connect two contiguous fore strokes are called "back strokes", meaning that these strokes should not exist in the standard character. In a broader sense, the back strokes are fictitious strokes which are provided to connect between fore strokes (i.e., to allow for the situation that the pen never leaves the paper when it should have). When a connecting stroke exists as a result of hasty handwriting, a fore stroke may appear in the handwritten script which would correspond to a fictitiously created back stroke in template. On the other hand, a back stroke (the imaginary stroke connecting two fore strokes) may also be truncated and thus degenerating into an intersection "point". By these designations, the primitives in the character representation can be classified into the three types of fore strokes, back strokes, and points. In summary, in the present invention, "back strokes", which do not actually appear in a character, are created for a template as well as an input character; and "point" primitives, which do not fit into the definition of a stroke, are identified for the input script, so as to accommodate connected strokes. The designation of these three primitives disclosed in the present invention greatly enhances the efficiency of the character recognition process, and offer a logical explanation of the recognition process.

(f) Classifying strokes in the handwritten characters into fore strokes, back strokes, and points, wherein the fore strokes are strokes that actually appear in the character, the back strokes are fictitious strokes to allow for stroke connections that should not appear, and points are provided to allow for truncated back strokes in the input script:

In the present invention, the strokes are classified into three types: fore strokes, back strokes, and points: fore strokes are those strokes that actually appear in a character pattern; back strokes are "pseudo strokes" which are provided in the present invention to allow for imaginary strokes connecting between fore strokes. When a connecting stroke exists as a result of hasty handwriting, a fore stroke may appear in the handwritten script which would correspond to a back stroke in template (i.e., the stroke does not exist in the template). This is shown as stroke "b" shown in FIG. 2b. On the other hand, a back stroke (the imaginary stroke connecting two fore strokes) may also be "simplified" and thus degenerated into an intersection "point", such as the intersection points of segments a' and b' shown in FIG. 2c. By these designations, the primitives in the character representation can be classified into the three types of fore strokes, back strokes, and points. The designation of these three primitives disclosed in the present invention greatly enhances the efficiency of the character recognition process, removes the constraints from the stroke number, makes the explanation of the recognition process easy to understand.

(g) For each candidate template character, performing a stroke correspondence for each stroke correspondence rule until all the stroke correspondence rules contained in the template character are exhausted, so as to identify all fore stroke.fwdarw.fore stroke correspondence:

The step of stroke correspondence performs stroke matching between an input script and a template character; it involves finding a binary relation (designated as "q") between the set of template primitives and the set of input primitives. Because the reference database contains only fore strokes, fore strokes are matched first. As discussed above, a primitive in the present invention is a basic stroke in a broader sense, because not only does it include the conventional fore stroke, it also includes an imaginary back stroke and a "point". The binary relation q for stroke matching from set X to set Y is defined as follows:

where X denotes the set of primitives of a template character, and Y denotes the set of primitives of an input character. For any element x.sub.i .epsilon.X, if the mapped image (i.e., a matched primitive) y.sub.i (y.sub.i =q(x.sub.i), y.sub.i .epsilon.Y) exists, then there exists only one such image y.sub.i.

The process of performing stroke correspondence is to find a possibly corresponding stroke in the input script for each of the rules in the template character. If a possible basic stroke in the input script is found to match with one rule, then this selected basic stroke is removed from the input pattern, and the next stroke correspondence rule is applied onto the remaining strokes of the input pattern until all the rules in the template are exhausted.

(h) Using the above described stroke matching process and the stroke order of the input character to find other strokes correspondences between a fore stroke and a non-fore stroke, a non-fore stroke and a fore stroke, and a non-fore stroke and a non-fore stroke(i.e., back.fwdarw.point, back.fwdarw.fore, fore.fwdarw.null, null.fwdarw.fore, back.fwdarw.null, and null.fwdarw.back):

It is quite likely that some rules in the template character may not find matched strokes in the input character. In a preferred embodiment of the present invention, there are eight possible types of matches, namely: (1) fore.fwdarw.fore, (2) back.fwdarw.back, (3) back.fwdarw.fore, (4) back.fwdarw.point, (5) back.fwdarw.null, (6) null.fwdarw.back, (7) fore.fwdarw.null, and (8) null.fwdarw.fore, where x.fwdarw.y indicates a matched pair, and "fore" represents a fore stroke primitive, "back" represents a back stroke primitive, "point" represents a point primitive, and null indicates no matched primitive. It should be noted that in this preferred embodiment, the point primitive only exists in the input script. Furthermore, a back primitive (i.e., an imaginary stroke) in the template can be corresponded to a fore stroke, a back stroke, a point in the input script, or no match. The ubiquity of back strokes demonstrates its importance in recognizing Chinese handwritings.

The fore strokes of a template character are the essential primitives of a character, and, most of them are expected to exist in the input handwritten patterns, especially for those long or major strokes. A template in the reference database described the standard pattern of a character. When the strokes are connected in the handwritten characters, back strokes in the templates may appear as fore strokes or may even be reduced to "points" in the input patterns (i.e., scripts). These variations are commonly observed in the handwritten Chinese characters, and thus are accommodated in the method of the present invention for character matching. During the matching process, if an input script is matched against a template of the input category (i.e., possible variations of a character), the fore strokes in the template character are first matched with the fore strokes in the input script. The back strokes in the template pattern are obtained by tracing the matched pairs following the input stroke order, which is stored during the input step as part of the on-line recognition process. A back stroke in the template character may be mapped to a back stroke, a point, a fore stroke in the input script, or no match.

(i) Performing computation of discrimination functions to find a template character with a minimum distance:

After the step of stroke correspondence, which finds the matched strokes, a set of discrimination functions using a database of character patterns and a database of spatial relationships between strokes of characters are computed to find the best matching character(s). In the method disclosed in the present invention, the spatial relationships between strokes, as well as the information of lost strokes, superfluous strokes, and connected strokes, are utilized to compute the character distance between a template character and an input script, for character discrimination.

After the stroke correspondence between a template character and an input character script is accomplished, the matching relation between template strokes and input strokes is now considered to have been obtained. Based on the matching relation, input strokes are re-ordered in accordance with the stroke order of the template character. For an input script, the spatial relationships between the strokes based on the template stroke order are used in character discrimination.

As discussed above, the matching relation between the template and input primitives is a binary relation q: X.fwdarw.Y, where X denotes the set of template primitives and Y denotes the set of primitives of the input script. The matching types of fore stroke.fwdarw.fore stroke and back stroke.fwdarw.back stroke are considered correct matchings for both fore strokes and back strokes. The matching types of back.fwdarw.point and back.fwdarw.fore also exist in the case of connected strokes. The matching types of back.fwdarw.null and fore.fwdarw.null occur in the case of lost stroke(s). And the matching types of null.fwdarw.back and null.fwdarw.fore can occur in the case involving superfluous stroke(s).

The distance functions used for character discrimination are defined as the sum of the distances calculated in the various matching type(s). For the matching type of fore.fwdarw.fore, the matched pairs of fore strokes are re-ordered according to the stroke sequence in the template character. In the input script, the spatial relationships between the fore strokes i and i+1 are described by four vectors: ss(i, i+1), se(i, i+1), es(i, i+1), and ee(i, i+1). These vectors are defined as the connections between the starting point ("s") and the end point ("e"), from stroke i to stroke i+1, as illustrated in FIG. 6. In performing the computation, the directions are quantized to accommodate variations in the handwritings. Thus, function quant.sub.-- 8() is defined to quantize the directions of the vectors into eight direction codes. All the possible variations of spatial relationships between handwritten strokes i and i+1 are all accommodated in the reference database, and are denoted by four sets of eight-direction codes: SS(i, i+1), SE(i, i+1), ES(i, i+1), EE(i, i+1). For the matched pairs of fore.fwdarw.fore, the cost function ff.sub.-- err.sub.-- ss(i) is defined as: ##EQU6##

The other cost functions ff.sub.-- err.sub.-- se(i), ff.sub.-- err.sub.-- es(i), ff.sub.-- err.sub.-- ee(i) are similarly defined (i.e., substituting ss(i, i+1) with appropriate vectors). Assuming that the template character has FM strokes, for the matched pairs of the type fore.fwdarw.fore, the distance function related to the spatial relationship between the fore strokes is defined as: ##EQU7##

For the matched pairs of the type fore.fwdarw.null, the distance function is defined as: ##EQU8## where FN is the number of matched pairs of the type fore.fwdarw.null, T.sub.-- STK is the number of strokes of the template character, l.sub.i is the length of the fore stroke i, and L.sub.T is the total lengths of all the strokes in the template character. The error cost of one lost stroke is considered as having a weight of 5, which includes the error cost of one fore stroke and four vectors representing the spatial relationships between two contiguous strokes.

For the matched pairs of the type null.fwdarw.fore, the distance function is defined as: ##EQU9## where NF is the number of matched pairs of the type null.fwdarw.fore, I.sub.-- STK is the middle value of the estimated range of input stroke number, l.sub.j is the length of the input fore stroke j, and L.sub.i is the total length of all the strokes in the input character.

The matched pairs of the type back.fwdarw.point occur in the case of connected strokes. The distance function is defined as: ##EQU10## where BP is the number of matched pairs of the type back.fwdarw.point.

The similarity between the template and the input characters is described by the following distance formula:

The candidate character with the minimum distance has the maximum probability of being the correctly recognized character for the input character. When the number of candidate characters with the minimum distance is more than one, then these candidate characters constitute a "similar group". When this occurs, the input character is further recognized by the special structural features of the character in the similar group which are represented by a logical tree. Many methods disclosed in the prior art can be used for this detailed recognition and thus will not be repeated here.

The present invention will now be described more specifically with reference to the following examples. It is to be noted that the following descriptions of examples, including the preferred embodiment of this invention, are presented herein for purposes of illustration and description, and are not intended to be exhaustive or to limit the invention to the precise form disclosed.

In this Example, the strokes of the Chinese characters are classified into 14 basic stroke types, as shown in Table 1. These data are useful in the stroke correspondence rules disclosed in the present invention.

FIG. 2a illustrates the "standard pattern", and FIGS. 2b and 2c show two different "cursive patterns", respectively, of the character . The two cursive patterns shown in FIGS. 2b and 2c vary in both the stroke number and stroke connections relative to the standard pattern shown in FIG. 2a. FIG. 2a also shows the basic strokes of the standard pattern for the character . All the possible basic strokes that are contained in the cursive patterns are also shown in FIGS. 2b and 2c. The input cursive character shown in FIG. 2b contains three strokes, which consist, collectively, of eight possible basic strokes in a variety of combinations, and the input cursive character shown in FIG. 2c contains only one stroke, from which 11 possible basic strokes can be independently extracted. When a connecting stroke exists as a result of hasty handwriting, a fore stroke may appear in the handwritten script which would correspond to a back stroke in template. This is shown as stroke "b" in FIG. 2b. On the other hand, a back stroke may also be "simplified" and thus degenerating into an intersection "point", such as the intersection points of a' and b' as shown in FIG. 2c.

FIGS. 3a-3c illustrate the process of applying the stroke correspondence rule using the examples shown in FIGS. 2a-2c. FIG. 3a shows the stroke correspondence rules for the characters . The stroke correspondence rules consist of (1) basic stroke types that are "allowed" (i.e., commonly encountered) in the handwriting, and (2) the geometric feature measure(s) of the strokes used in the stroke matching process. Each character category (which includes variations of a standard character) has a set of rules, which accommodate a range of possible variations. The sequence of applying the rules is the same as the sequence of finding the matched pairs for the template character. A handwritten input stroke may consist of one or more basic strokes due to stroke connections (i.e., connections between the basic strokes as a result of handwriting). In matching against a template stroke, a possible basic stroke in the input script is first selected based on one rule. Then, this selected basic stroke is removed from the input pattern, and another stroke correspondence rule is applied onto one of the remaining strokes of the input pattern until all the rules in the template character are applied. FIGS. 3b and 3c show the process of applying the stroke correspondence rules for the two cursive patterns in FIGS. 2b and 2c, respectively.

27 types of geometric feature measures of strokes are utilized to establish the stroke correspondence rules. These geometric feature measures are listed in Table 2. Before applying the stroke correspondence rules, all possible basic strokes are sorted based on the 27 measures in an increasing order. The sorted sequence for each measure is recorded in an array. Each stroke correspondence rule would contain the information that the matching stroke must have a maximum or minimum value of certain type of geometric feature measures. The matching stroke can then be directly acquired from the array storing the sorted results.

Each geometric feature measure is considered a special geometrically related characteristic length associated with a stroke to aid stroke recognition. First, each stroke is considered to be bounded by a minimum bounding rectangle (MBR). The x and y coordinates of the four boundaries and the center point of the MBR of a stroke are designated as geometric feature measures numbered 1 through 6, respectively, i.e., MBR.sub.-- Xmin, MBR.sub.-- Ymin, MBR.sub.-- Xmax, MBR.sub.-- Ymax, MBR.sub.-- Xcenter, and MBR.sub.-- Ycenter. The x and y coordinates of the starting point and end point of the strokes are designated as geometric feature measures numbered 7 through 10, i.e. Sx, Sy, Ex, Ey, respectively. These geometric feature measures are often adequate for simple strokes. For complicated strokes, additional or auxiliary geometric feature measures are provided to facilitate the stroke correspondence.

FIG. 4 shows the eight reference points of the bounding rectangle of a character. These eight reference points are: LT (left-top corner point of the rectangle), TC (top-center point of the rectangle), RT (right-top corner point of the rectangle), LC (left-center point of the rectangle), RC (right-center point of the rectangle), LB (left-bottom corner point of the rectangle), BC (bottom-center point of the rectangle), and RB (right-bottom comer point of the rectangle). From these eight reference points, 14 auxiliary geometric feature measures, numbered from 11 through 24, can be derived. They are defined as the Euclidean distances from the starting point, end point, and center point of a stroke's MBR to these eight reference points, respectively. Additionally, the length of a stroke is considered as a geometric feature measure, numbered 25. Finally, the geometric feature measure numbered 26 is defined as the "city block" (or block distance) from the left-bottom corner point of a character to the left-bottom comer point of an individual stroke's MBR, and the geometric feature measure numbered 27 is defined as the city block from the left-bottom corner point of a character to the right-top corner point of that same stroke's MBR. Some, if not all, of these additional measures are useful for ensuring stable stroke correspondence for some complicated cases. Most of the strokes correspondence rules developed in the present invention require only one geometric feature measure. In other rules, two geometric feature measures are utilized.

The application of the stroke correspondence rules can be illustrated by FIGS. 5a and 5b. FIG. 5a shows a standard pattern (template) of character , with its strokes identified and labeled. FIG. 5b lists the stroke correspondence rules. Each stroke correspondence rule contains the rule type, the basic stroke types allowed in considering handwriting variations, and the geometric feature measure(s) of the strokes to be used to find the matching stroke. In FIG. 5b, the geometric feature measure is denoted as Fi (1.ltoreq.i.ltoreq.27), where i is the geometric feature number (i.e., the type of geometric feature measure). For a template stroke, the matching stroke must have the minimum (denoted as S0) or maximum (denoted as S1) value of certain geometric feature measure Fi in the possible basic strokes of the in script. A type one rule utilizes one geometric feature measure denoted by (Fi, Sj) (1.ltoreq.i.ltoreq.27 and j=0 or 1). A type two rule, on the other hand, utilizes two geometric feature measures denoted by (Fi, Sj, Ck; Fi', Sj'), where the meanings of i, i', j, j' are similar to those defined for type one rules, and the value Ck indicates that k candidates strokes, of the basic stroke types allowed in the rule, have been selected from the remaining possible basic strokes of the input script based on the first geometric feature measure. The second geometric feature measure is used to determine the matching strokes from the k candidates strokes. The order of applying the stroke correspondence is based on a predefined rule sequence. The rules of strokes .alpha. and .beta. in FIG. 5b are taken as examples for explanation. FIG. 5a shows that the rule for stroke a belongs to type two, and stroke type 6 is allowed. Based on the first geometric feature measure, (which, in this case, is the stroke length), strokes .alpha. and .delta. are selected as the candidate strokes from FIG. 5a. Next, the matched stroke, with start point of minimum x coordinate (F7), is selected from the two candidate strokes.

The handwritings of strokes .beta., .gamma., .epsilon., and .eta. in the character .eta. are very close to "written dots", or short strokes. A handwritten short stroke is usually unstable and the stroke type recognized often is not the correct (i.e., original) stroke. Therefore, a number of possible variations (in this case, stroke types 0, 1, 2, and 3) should be taken into consideration and are accommodated in the rule. The rule for stroke .beta. belongs to type one. The only geometric feature measure is the Euclidean distance from the center point of the stroke's MBR to the left-top corner point of the character. From the remaining strokes with the allowed stroke types in the rule, the matched stroke is the stroke with the least value of the geometric feature measure. A source code describing the above examples is provided in the Appendix.

Two studies were performed to verify the effectiveness of the method disclosed in the present invention. One study verified the performance of the structural analysis in the recognition process, including the recognition accuracy and the speed. The other study verified the effectiveness of the preliminary classification included in the process of the present invention.

Two databases of handwritten samples were established. One database contained on-line handwritten samples of Chinese characters, designated as ITRI-OLCCR database. The ITRI-OLCCR database contained data which are pen tracks of handwritten Chinese characters; it contained 5,401 character categories (a character category contains the standard character as well as a variety of "allowed" handwritten variations) of most frequent used everyday Chinese characters, written by 230 persons via a tablet. For each character category, there are 23 samples written by different persons. The other database contained off-line handwritten Chinese characters, designated as ITRI-HCCR database. The ITRI-HCCR database contained handwriting samples from more than 2,600 persons. For each character category, the number of character samples ranged from 214 to 284. During the collection of the character samples, each character was written within a square, sized 11 mm by 12 mm, on an A4-sized sheet using a pencil or a ball-point pen. The written samples were then transformed into bit-patterns using a 300 dpi scanner. The ITRI-HCCR database was used as the training set for the database of statistical features in the recognition system.

In Example 6, 1,225 character categories were selected from the 5,401 categories of the ITRI-OLCCR database as the testing sample set. In these 1,225 character categories, their stroke numbers ranged from 1 to 21. Eleven testing samples per character category were taken from the ITRI-OLCCR database. In this study, only the estimated range of the input stroke number was used in the preliminary classification. The average number of candidate characters after the preliminary classification was 274. Table 3 shows the overall recognition rate of this study, in which the input character was recognized within ranks 1 through 5. The input was said to be recognized within rank k, if the corresponding reference character was among the top k most similar characters to the input. Consequently, an Lth rank cumulative recognition rate was defined as the recognition rate at which the correct reference character of an input sample appeared with the first L ranked candidates. The Lth column indicated that the Lth rank cumulative recognition rate. Altogether, the average first rank recognition rate was 93.51%, and the average 5th rank cumulative recognition rate reached 98.16%. In a separate study, the recognition rate was improved when more handwriting variations were included in the reference database. The simulation was implemented using a program written in C language on a PC/AT-486 personal computer. The average recognition speed was about 0.8 seconds for each character. When all the 5,401 character categories were used in the reference database of a recognition system, the recognition speed could become an important factor to be concerned. The number of candidate characters selected during the preliminary classification could become an important factor affecting the recognition speed

The second study of Example 7 involved performing a preliminary classification for a reference database containing all the 5,401 character categories. The estimated range of possible input stroke number and the statistical features were utilized in the preliminary classification to reduce the number of candidate characters to be used in the later stages of structural analysis. Six statistical features were used in this study including: stroke density function, stroke density distribution, contour line length, peripheral background area, background area, and gradient feature. The database of the statistical features was trained by 100 samples per character category taken from the ITRI-HCCR database. A K-means algorithm was utilized to cluster the 5,401 categories into 80 clusters. A character category may appear in more than one cluster, resulting in overlapped clusters. The statistical feature vector of an input character was first computed. Then, the 80 distances, measured between the vector and the mean feature vectors of the 80 clusters, respectively, were computed. From these 80 clusters, the five clusters with the shortest distances (i.e., within the first five ranks) were selected and combined into a set of candidate characters. The templates of the candidate characters were processed in the later stages of structural analysis. In this study, 23 testing samples per character category were also taken from ITRI-OLCCR database. The recognition rate of preliminary classification were defined as the "hit rate". On the average, when the statistical features alone were used, the hit rate was 99.61%, and the number of candidate characters was 1,480. On the other hand, when the estimated range of input stroke number alone was used, the average hit rate was 98.41%, and the number of candidate characters was 1,040. When both the statistical features and the estimated range of input stroke number were used, the number of candidate characters was reduced to only 353, and the hit rate was 98.05%. The main cause of the hit errors was attributed to the erroneous testing samples. Thus, the present invention represents an excellent character recognition method for Chinese characters which can be accomplished on a real-time basis even for a very large reference character set. The recognition speed can be further improved by utilizing additional features in the preliminary classification step.

The foregoing description of the preferred embodiments of this invention has been presented for purposes of illustration and description. Obvious modifications or variations are possible in light of the above teaching. The embodiments were chosen and described to provide the best illustration of the principles of this invention and its practical application to thereby enable those skilled in the art to utilize the invention in various embodiments and with various modifications as are suited to the particular use contemplated. All such modifications and variations are within the scope of the present invention as determined by the appended claims when interpreted in accordance with the breadth to which they are fairly, legally, and equitably entitled.