CA1118109A - Pattern reading system - Google Patents

Abstract

PATTERN READING SYSTEM

Abstract of the Disclosure A method of recognizing visual patterns of hand printed characters or the like by means of an optical character reader which reads the patterns by a so-called outermost point method. While tracing the contour of the pattern stored in a two-dimensional memory, the distances from the starting point of the tracing and the integrated values of the coordinates of the points traced on the contour from the starting points are simultaneously obtained successively to extract outermost points for the series of contours. According to this outermost points, the contours are segmented into the convex line segments, concavity line segments and hole segments and the corresponding parameters of feature of each segment are detected. Simultaneously, the convex line segment of which is shorter than a predetermined length is rejected and the remaining segments are subjected to matching operation in accordance with preliminarily prepared dictionary, thereby making a decision as to the pattern's identity.

Description

~ 18~109 Background of the Invention This invention relates to a pattern reading system adapted for the recognition of characters, particularly handprinted characters.
In the past, as the systems for recognizing the hand-printed characters having no specified standard forms, the structural analysis approach have been used most widely in which the local characteristic properties of the visual patterns of characters to be recognized are extracted and each of the characters is represented by a construction composed of the extracted features and the construction is recognized. The known such methods may be broadly divided as follows;
a) The methods based on the reco~nition that each chracter is made up of lines and the character is reco~nized on the basis of these line elements.
b) The methods in which each character made up of the black portion is recognized by extracting the geometric features of the character including the which background portion.
The methods of the above a) consist mainly of the type which requires a preliminary processing or thinning pro-cedure whereby the line segments of the resulting thin character are extracted and approximated with straight lines, and these methods have the disad~antage of confusing the features of a character due to the occurrence of "whiskers" and the loss of the delicate shapes of the character caused by the thinning operation. ~nother dis-advantage is the fact that the approxi~ation of curves with broken lines tends to complicate the representation of the ~.

~8~09 curves, and these disadvantages impose so~ne limitations on the methods of the above a). On the other hand, the methods of the above b) are effective in that the features of a character including the white back ground portion can be extracted without any thinning operation and the con-cavity and convexity of the line segments of the character are extracted only by following the contours.
~ lany different techniques have been developed for performing these methods and these techniques have been disadvantages from the hardware point of view. ~or example, in the case of the method in which the concavity and convex se~ments of a character are extracted by a two-dimensional processor in accordance with a two-dimensional parallel processing, the construction of the required apparatus will inevitably be made extremely complicate and the cost will also be increased. In the case of the other method which is designed to overcome these deficiencies and in which a specially designed one-dimensional processor to process data line by line, there is a disadvantage that since the method is based on the line parallel processing, it is necessary to use a special processor for speeding up the processing and moreover a character consisting of discrete line segments cannot~be positively recognized unless a satisfactory continuity of the respective points is main-tained, thus inevitably complicating the apparatus.
Under tAese circumstances, a character recognition system capable of highly accurate character recognition and having a nu~ber of advantages frcm the hardware point of vie-~ has been developed in ~hich 'he contour of a character is traced to extract the individual concavity 1'1~8~L~9 and convex line segmerlts or the contour. This tracing procedure has the advantage of a reduced number of calcu-lation points, as for example, the ratio of tne number na of the intersections on the entire two-dimensional memory to the number nb of the points on the contour of a cnarac-ter or n~/nb would be on the order of 20 according to the actual instances and the calculation points to be processed would be reduced to about one twentieth. Another advan-tage is that no preliminary processing is required for extracting the connected components and that the sequence of all the calculation points or their connections forming the line segments is defined, making it easier to determine the correspondence of the extracted features for a matching operation with a dictionary. ~owever, this tracing system has the disadvantage of being unsuited for use with the two-dimensional parallel processing. On the other hand, the concavity and convexity structure extraction systems which are known to date and employing this tracing technique are also disadvantages in that the recognition, though exact, inevitably tends to become microscopic and that while there will be no problem in the case of a character with a smooth boundary surface, the presence of a large number of small concavity and convex segments on the contour will cause any convex line segment to be mistaken for a concavity line segment and so on.
Summary of the Invention With a vie~ to ove-rcoming the fcregoing deficiencies in the prior art, it is a main ObJ ect o~ the inventiGn to provide a pattern reading system in which without ~eing affected by the local varia~iors, particularly the small concavity and convex segments on the contour of a ch&racter pattern, the contour is segmented on the basis of 2 gloval feature extraction and then the detailed par&meters of feature are extracted so as to effect the pattern rcading.
It is another object of the invention to provide a pattern reading system in which the preparation of a dictionary for effecting the reading is accomplished in such a manner that the detected short convex line segments are rejected and the arrangement of the line segments is controlled for a matchin~ operation, thereby making the dictionary compact without any deterioration in the pattern recognition capacity.
In accordance with the pattern reading system of this invention, a two-dimensional memory stores a two-dimensional character pattern of a character which was scanned by an optical scanner and the stored content is recognized with a high efficiency by a recognition system based on the outermost point method. More specifically, while tracing the contour of the two-dimensional character pattern on the two-dimensional memory, the distances of the traced points on the contour from the starting point of the tracing are obtained and simultaneously the integrated values of the coordinates of the traced points on the contour ~rom the starting point are calculated so as to extract them as outermost points for the series of contours.
'~hen, in accordance with the outermost points, the charac-ter contours are segmented into the convex line se&ments, concavity line segments and hole seg~ments and then the parameters o~ fe2tl1re 2re detected or each of tne segments, whereby the convex line segments which are shorter than a ~ L18~G9 predetermined length are rejected and the remaining segments are subjected to a matching operation with the preliminarily prepared dictionary, thereby recognizing the two~dimensional pattern.
Further, in accordance with the invention, the con-tours are seg~ented into the respective segments in accordance with the outermost points. In extracting the parameters of feature of the segments, on the basis of the concavity line se~ment detected first, the parameters of feature for the other segments are grouped in the o-rder of detection as the first-category features and also the para-meters of feature for the segments ranging from the second concavity line segment in said group up to the segment next to the first concavity line se~ment are sequentially grouped as the second-category features, whereby when there is a specific designation on the dictionary, a ~atch is attempted between the two sets of the parameters of feature and the dictionary, thereby accomplishing the two-dimen-sional pattern recognition.
In accordance with the invention, in particular the recognition of patterns with a high degree of accuracy can still be ensured even if the memory capacity of the dic-tionary is reduced and there is no need to use an additional dictionary for considerably distorted form of the input character. .Ioreover, the input character can be automati-cally made decl3ion by me2ns of the feature parameters representing the partial patterns without using any specially desi~ned processor, and the char2cter-at-a-time processing ensures an effective recognition of the visual pattern to be read which is in the form of a discrete p~ttern.

1118~0g ~urther, there is no need to increase the memory capacity and even characters whicn are de~ormed more or less can be read effectively, thus making it possible to realize a pattern recognition apparatus which is well suited for automatic reading of handprinted characters.
The invention will be described in greater detail with reference to the accompanying drawings.
Brief Description of the Drawings Fig. la is a block diagram 3howing a hardware s~stem which is useful for explaining the principle of the invention.
Fig. lb is a flowchart diagram useful for explaining the operation of the hardware system shown in Fig. la.
Fig. 2 is a flowchart diagram showing the calculation of comparison values (XY) for the respective points for producing a list of outermost points and the process for storing the outermost points.
Fig. 3 is a diagram showing an outermost point list by way of example.
Fig. 4 is a diagram showing a binary quantized input character and the positions of the outermost points desig-nated by numbers.
Fig. 5 is a dia~ram showing the relation between the contour and the direction ~i of the line element at a traced point on the contour.
~ ig. 6 shows diagram showing the contents of a co-ordinate memory and the pGSitions of a pointer i' with the memory length n=3.
Fig. 7 is a diagram showing an &ngle conversion matrix and its content.

Fig. 8a is a diagram showing by way of example a character having an inner prctrusion.
Fig. 8b is a diagram showing the concavity line segment corresponding to the character of Fi~. &a as well as the features r~il and rLi2 representing the position of the inner outermost point and the inner struc~ure and the direction r ~I of the inner protrusion.
Fig. 9 is a diagram showing that where the same tracing condition is used, the outer contour will be traced counter-clockwise and the hole contour will be traced clockwise, making the sign of pL(9) - p~(l) opposite to each other.
Fig. 10 is a diagram showing the segments derived from the list of outermost points.
Fig. 11 is a diagram showing the detected parameters of feature for the diagram of Fig. 10.
Fig. 12 is a flowchart diagram for a separate matching procedure.
Fig. 13 is a flowchart diagram useful for explaining the selection of the optimum mætching dictionary.
Fig. 14 is a diagram showing by way of example the form and one of content of a mask in a category feature of the dictionary.
Figs. 15a and 15b are diagrams showing exemplary cha-racter contours useful for explaining the effect of reject-ing the shorter convex line segments.
Figs. 16a and 16b are diagrams showing the contours of the same character whose line seg ents will be named differently depending on the deformaticn of the character.
Figs. 17a and 17b are diagrams showing by way of con-tour lines the changes in the grouping of line segments ~8~019 and the corresponding dictionary contents.
~escription of the Preferred Embodiments ~ ig. la is a block diagram showing the overall ccnst-ruction of a hardware system which is useful for explaining the principle of the invention, and ~ig. lb is a flowchart diagram showing the processing procedure for the system of Fig. la. The hardware system shown in ~ig. la comprises a contour tracing unit 2 for receiving binary signals corres-ponding to the white and black portions of a two-dimensional character pattern of the input character from an optical scanner (not shown), a calculating unit 3 and a matching unit 4, and these units are arranged in a three-stage con-figuration. The respective stages are operated for the same processing time through a pipeline controller 5. The operation of this system is as follows. As will be scen from the flowchart of Fig. lb, the input character is detected by a suitable scanning unit which detects for example 64 x 64 bits at 16 levels and the avera~e values of the densities represented by the bits are converted into the corresponding binary signal form. The binary quantized signal is applied to the contour tracing unit 2 in which the input signal is subjected to tracing process-ing along the contour and the extraction of outermost point is effected for each of the traced points of the input character. The calculating unit 3 extracts the con-cavit~- and convex segments of the character from the outer-most points. The calculating unit ~ also se~ment3~he con-tour into the respective convex line segments, concavity line se~ents and hole segments and extracts the parameters of .eature for each of the segments by the method ~hich _ g _ will be described later. In the matching unlt 4, a match is attempted between a set of maskes in each category fe~ture of the dictionary and the input character in accord-ance with the parameters of feature and the name of the best matched category of the dictionary is generated as the result of the decision. More specifically, the binary quantized input signal l is stored in a two-dimentional memory 21 of the contour tracing unit 2, so that the con-tour is traced while controlling the access to the stored content by an address controller 23 through a local controller 22 in response to the control signals from the pipeline controller 5, and simultaneously an extracting circuit 24 extrac-ts the outermost points in a plurality of predetermined directions for each of the traced points. Numeral 25 desig-nates a buffer.
Extraction of outermost point The outermost point is the outermost located point in a predetermined direction from a point on one series of contours forming a character and the extraction of outer most points will now be described with reference to Figs.

2 to 4.
Fig. 2 is a flowchart showing the calculation of the comparison values (XY) made for the respective points to obtain a list of outermost points and the process for storing the outermost points, Fig. 3 is a diagram showing a list of outermost points by way of example, and Fig. 4 is a diagram showing the binary quantized black points (character portion) marked X and the positions of the corresponding outermost points oy means of the two-dimentional coordinates (x, y).
Gererally, where the input character is not a discrete ~ 81(19 one, the outermost points consist of the points at four directions of the top, bottom, right and left ends and are frequently used for the character positioning purposes or as a character size normalizing factor. In accordance with the present invention, tne outermost points of an input character are extracted t for example, for 16 directions instead of the four directions ænd they are used to deter-mine the concavity and convex structure of the character.
The number of directions which is considered suitable is 16 from the practical point of view, although the invention is not intended to be limited to this number.
rl'he outermost points may be extracted in the following way.
(a) In the course of scanning the bits in the two-dimensional memory 21 in a raster scan mode so as to discover any untraced contour (where the intended object of the scannin~ is the white ground and there is a black character at an adjacent point in any one of the four sides but there is no line segment name near any of the 8 directions).
(b) In the course of a scanning for tracing, in a contour tracing mode, the contours of the character pattern in the two-dimensional memory 21 fro~ the starting point back to the starting point successively.
~ y storing the outermost points for the points traced on the contours, the outermost points on the series of con-tours will be extracted. ~Iore specifically, considering a certain point i on the contour, i- the point is outside the conve~ hu~l formed by the traced contour in accordance with the evalution function in a direction r, the length along the contour from the starting point ~ OI the tracing to the present point (~X(i), ~Y(i)) (in fact, the length is increased by 1 if the direction is vertical or horizon-tal, and the length is increased by ~ if the direction is diagonal) or ~Li is stored in PX(r), ~Y(r) and ~(r), respectively, and at the same time the integrated values of the respective X and Y coor~inates for the contour from the st&rting point ~ to the point i (however, the value is multiplied by ~ if the direction is diagonal) or X~(i) and Y~(i) are respectively stored in GX(r) and GY(r). ~hen the contour has been traced completely, the coordinates (PX(r), PY(r)) of the outermost point in each direction r, the length P~r along the contour from the starting point S
tc the outermost point in the direction _ and the sums GX(r) and Gy(r) of the coordinate values along the contour from the starting point S to the outermost point in the direction r are obtained for the outermost points in the 16 directions.
In this way, a list of outermost points given by r=U1~16 (PX(r), PY(r), P~(r), GX~r), GY(r)) is obtained. One example of such list is shown in Fig. 3 and the list represents the entire data on the outermost points which is applied to the calculating unit 3 for the line segment structure extracting purposes.
Extraction of line segment (concavity, convex, hole) The calculating ~lit 3 comprises a calculator 31 for performing the extraction of line segments from the outer-most points, the segmentation of contour into line segments and the extraction of parameters of feature for tne line segments as will be described later, a local controller 33 responsive to the control sign~l3 from the pipeline cont-roller 5 to cGntrol the operation of the individual circuits and input and output buffers 33 and 34.
Fig. 5 shows the relation be~ween the direction ~ of the line element at a certain traced oint on the contour.
~ig. 6 shows the contents of a coordinate memory having a memory length n=3 ~nd the positisns of a pointer i', and the directional accuracy 2nd the resistance to noise may be improved by increasing the memory length n but with the corresponding complication in the construction of the memory.
Fig. 7 shows an angle conversion m2trix and its content.
In order to extract the concavity and convex se ments of the character from the outermost points, the contour of the character is segmented into a plurality ol line segments according to predetermined conditions. This process is called as an e~traction of line se&~ents. In accordance with the invention, each contour is segmented according to the following convex and concavity discrimination condi-tions.
More specifically, noting the (r, r+l) mod 16 segments according to the list of outermost points, the following values are calculted rLl=PL(r+l)-PL(r) rL2= ~(PX(r+l)-PX(r))2+(PY(r+l)-PY(r))2 rL12=rLl-rL2 and also XpG=(Gx(r+l)-Gx(r))/
YPG=(GY(r+l)-GY(r))/r XPC=(PX(r+l)+PX(r))/'2 YPC=(PY(r+l)+PY(r))/2 8~J~
PC=((PY(r)-~Y(r+l))-(XPG-XPC) + (~X(r+l)-PX(r))-(YkG+~-Pc))/rL2 In accordance with the above r~l2 and PG, if rLl2~-lc (specified value) and PC ~PC (specified value) the segment is considered a concavity segment and the other segments are considered convex segments. The contour portion consisting of such concavity segment is extracted as a con-cavity line segment and the similar process is accomplished for each of the 1~ directions so as to extract consecutive convex segment as a unified convex line segment.
For instance, if a point r is represented by (PX(r), PY(r)), then assuming that the segments (r, r+l) and (r+l, r+l) represent convex segments, the locus PL along the convex line segments on the contour is the line segment from the point r to the point r~2 and pLl=rLl+rL2. On the other hand, the straight line segment from the point r to the point r~2 represents a straight line segment PL2 inter-connecting the starting and ending points of the convex line segment. In Fig. 10 the straight line segment PL2 represents the length of L2 in Fig. 4 and pLl represents the length of Ll in Fig. 4. In this case, p represents the name of line segments and it is given to the line segments in the order of their detection as shown in Fig.
10 by the numbers 1, 2, ~ and 4 which are stored as infor-mation for the respective traced points.
In addition to the above pLl and PL2 and the index of concavity and convex, the following parameters of feature are extracted pG : the center of gravity (Xpg, Ypg) o~ pLl p~ : the right-h2nd direction perpendicular to the straigh' line segment from the starting point to the ending point of PL2 pa: the length of a perpendicular drawn from pG
to P~2 which is given as follows, if the coordi-nates of the starting and ending points Of P~2 are respectively represented by (Xps, Yps) and (Xpe, Ype) pO (Yps-Ype)-(Xp~-Xps) - (Xpe-Xps) (Yps-Yp~) ~1 (xpe-xps)2 + (ype-yps)2 H(16): the angular distribution,quantized in 16 directions, of the right-hand direction ~i per-pendicular to the line segment at a traced point i on the contour.
Although the direction ei may be obtained mathemætical-ly from the coordinates of the preceding traced points up to the point (i - n) and the coordinates of the present point with the use of a sine function as shown in Fig. 5, engineeringly the direction ~i can be obtained simply at a high speed in the following way.
In other words, the direction ~i cf a given traced point i can be given by referring to a memory LG storing therein the coordinates of the preceding traced points up to the point (i - n) as shown in Fig. ~, a pointer _' indicating the coordinates of the point (i - n) in the memory LG and a separate angle conversion matrix TMX of (2n~1)x(2n+1) having the direction values stored therein as shown in Fig. 7, as follows K=(ex-exi')+(2n+1)(ey-eyi')+2n(n+1)+1 =(ex-exi')+2n(ey-eyi')+ey-eyi'+2n(n+1)+1 =(ex-exi')+4(ey-eyi )+(ey-eyi ) 3 ( (2-bit shi~t and four additions) =(ex-exi)+5(ey-eyi')+13 =T.~(K) (K=lr- 25) In other words, ~ is obtained from the difference between the present coordinates (ex, ey) and the coordinates (exi', eyi') in the memory LG ~Ihich are indicated by the pointer i and the direction ~i is directly derived from this K &nd the matrix TMX. After the direction ~i has been obtained, the coordinates (ex, ey) are substituted in the area ~G(i~) indicated by the pointer i' and the pointer i' is increased by 1. (i' is the mod of n). Thereafter, this process is performed repeatedly to obtain the directions ~i for the other traced points.
Extraction of inner structure (partial outermost point and its function) Referring now to Figs. 8a and 8b, the extraction of an inner structure will be described. Fig. 8a shows by way of example an input character including an inner protrusion, and Fig. 8b shows the corresponding concavity line segment, the positions of the inner outermost points for the charac-ter, the features rLil and r~i2 representing the inner structure and the direction r ~ of the inner protrusion.
The extraction of the outermost points for the entire contours has been described so far, and if the extraction of outermost points is confined to a single concavity line se~ment, the differcnce between the length rLil along the CGntour between the successive outermost points and the euclidian distance rLi2 between the points or rLI=rLil-r~i2 will be increased similarly as in the case of the previously described conca-vity detection conditions for the inner protrusion projected internally of the concavity segment and this may be utilized to extract the inner structure so ~s to represcnt the outermost point of the concavit~
line segment by the open direction I for rLil and r~i2.
Extraction of hole.
The method of extracting concavity and convex line segments has been described so far and the extraction of hole segments is accomplished by the calculator ~1 in the following manner.
~ ig. 9 shows that if the same tracing conditions are used, the sign of p~(9)-p~(l) will become opposite to eæch other, since the outer contour is traced countercloc~wise and the hole is traced clockwise. In Fig. 10, the segments derived from tne outermost points are indicated as concavity, convex and hole, the encircled numbers indicate the segment-atiOn points of the contours and each group of the same numbered contour points indicates the locus of one line segment. ~ig. 10 also shows by way of example the positions of p11, P~2, p~, p~ and pG and the relation therebetween for the concavity line segment represented by the number 1.
In the case of an input character including a hole as will be the case with G; when the e~traction o outer-most points for the inner contour of the hole is effected by the contour tracing unit 2, if, in the tracing mode, the tracing is effected so as to trace the just preceding white portion while looking the black or print portion on the left side to the direc~ion cf tracing, the contour will be trac~d clockwise resulting in p~(9)-pL(l)> C. In ot~er words, the si~r is opposite to that obtained ~,ihen the outer .G9 contour OI the input charac-ter is traced. ~s a result, in the calculator 3i the cloc~wise tracing locus is not segmented but considered as a single hole on the basis of the opposite sign and the hole is handled in the same manner as each of the previously described concavity and convex line se~ments. In this case, the starting point S(p) of the hole is the same with its ending point E(p) so that p12=0 and P~l= the length OI the contour.
In the case of a discrete character composed of a plurality of discrete constituent elements resulting in a plurality of contours, the contours which have already been detected and traced by the starting conditions of the scanning in the contour tracing, have their own segment names so that when any contour having no segment name is detected, it is guaranteed that the detected contour is a new contour. In this way, when all the bits of the input character in the two-dimensional memory 21 have been scanned, all OI the contours forming the input character hæve been each given its own segment na~e as shown in Fig. 10 and the corresponding parameters of feature (feature axis) have been obtained as shown by way of example in ~ig. 11.
The parameters of feature for the se~ments represented by the encircled numbers in ~ig. 10 are respectively shown in the columns designated by the corresponding parenthesized numbers in ~ig. 11.
I~atching Next, the matching unit 4 will be described. The uni~
4 comprises a dicticnary memo-ry 41, a calculator 42 for performing the computational operations which will be des-cribed later, an inpu~ buffer 4~, a matching iogic circuit -- 1~ --l~ls~als 44 and a local controller 45 responsive to the control signals from thc pipeline controller 5 to control the individual circuits, and the matching unit 4 operates according to the flowpaths of action shown in Figs. 12 and 13.
More specifically, of the input parameters of feature (feature table) the s~ort con~ex line segments shorter than a predetermined length are rejected and comparison with the up~er and iower limit values of all the parameters of feature is accomplished for the feature aY~es of each of the other seg~ents. Fig. 12 shows the method of detecting the degree of matching Dj with each category feature of the dictionary, in which q designates a corresponding feature name (segment name), i the designated content of a corres-ponding feature axis, Iiq the input value, Uiqj the upper limit value of the mask, Wiq the weight of an axis i, ~Ui the upper tolerance, ~Li the lower tolerance, Dmin and Ddif the rejection constants and i the name of the category feature of the dictionary.
Fig. 13 shows the selection of a subset of the category in the dictionary according to the candidate selection and the rejection of input ~hen there is no subset of the category feature and;~ig. 13 also shows the process by which the matching degrees Dj obtaincd for the selected category of the dictionary are compared each other to obtain the result o~ recognition of the input character, whereby the input character ~Jill be rejected if there is no Dj which meets thc requirements.
'l'he ~atcr.ing is accomplished by performing the candidate selection, the calculation of matching degrees LO~
and the selection of the category feature having a minimum value.
~ he candidate selection is effected by subjecting the input parameters of feature to a candidate selection in accordance with the nu~ber of discretions of an input character, the total number of segments ~, the number of hole and the open direction of concavity line segments and selecting the subject of the category in the dictionary.
No-.~ referring to the types of the dictionary, the con-struction of the dictionary corrrsponding to the respective subset of the category will now be descrlbed.
The dictionary comprises a plurality of category features each having some maskes. Each mask comprising the total number of line segments~ is generally comprised of feature vectors of respective segment and each of the feature vector is comprised of feature axes of a fixed number K. Each of the feature axes comprises an upper limit value Uiqj and a lower limit value liqj.
In the course of a learning for the preparation of the feature vector corresponding to each segment, the detection of short convex line segments is effected by the calculating unit 3 and the feature vector for the thus detected short convex line segments are not supplied to the matching unit 4. For example, when the input character is the sign ~ , in the case of the prior art system the number of the feature vectors in each of the masks must be equal to the number ~ of the line seg ents ~1) to (~) as shown in Fig. 15a and consequently the eight feature vectors must be prepared. In accordarlcc ~ith the invention, ho~-ever, i' is so designed that as many line se~ents as there c9 are the concavity line segments in the input cnaracter, that is, only the four ie2turc vectors corresponding to the concavity line segments (1), (3), (5) and (7) are required as shown in Fig. 15b and in this way the memory capacity for the dictionary can be reduced to one hal~
the capacity required previously.
Fig. 14 is a diagram showing the types and contents of one of the ~sk in the category '~eature of the dictio-nary corresponding to the exemplary character and each of the masks stores the upper and lower li~it values of the parameter of feature for the feature vector of the corres-ponding segment.
logics of matcning The logics for matching will now be described.
To effect the matching, the categories of the dictionary selected by the candidate selection are each subjected to the logical operations shown in Fig. 12 and the degrees of matching Dj between the input feature table and the feature vectors of the mask are obtained. In other words, if the content Iiq of the input feature axis i is between (the upper limit value + the upper tolerance3 and (the lower limit value - the lower tolerance), a distance is obtained which is determined by multiplying the largest one of the values of (Iiq - the upper limit value), (the lower limit value - Iiq) and zero by the weight which is dependent on the a~is i. If the content Iiq is greater tnan (the up~er li~it value + the upper tolerance) or smaller than (the lower limit value - the lower tolerance), a very l~rge corst~nt is selecJed -~or the distance n a sense of rejecting it logically. This dist&nce on the a;~
axis i is me~s~red for all thc axes OI the same feature vector and then added to cb~ain -the matching degree Dj of the mask, whereby the degree o~ deiormation on each of the axes is checked to see whether it is within a predetermined tolerance and simultaneously the degree of defcrmation of the whole input character as compared with the masks is obtained as a measured quantity. Then, as shown in Fig.
13, the degrees of matching of all the prospective category are compared with each other to obtain the lowest ~alue Dj ~nd the next lowest value Dj', whereby when the lowest value Dj is smaller than the allowable maximum distance Dmin and also the ~-alue of Dj'-Dj is reater than the tole-rance Ddif for the next lowest value Dj', the value Dj of this category is accepted as the result of the decision.
If there is no Dj which satisfies the requirements, the input character is rejected.
With the matching operation described above, if an input character is for example deformed as shown in ~igs.
16a and 16b so that the starting points of tracing differ from each other and the resulting line segments are desig-nated differently, it will be necessary tc increase the number of masks for each of categories. In accordance with the invention, this will be prevented in the following manner.
More specifical y, when the line segments are processed by the calculating unit 3, as shown in ~ig. 17a, two feature groups, i.e., a first-category group (the unparenthesized numbers) for the parameters of feature of the line segments designated in the order of their detection and a second-catego-ry group (the parenthesized numbers) for the ~ara-meters of feature of thc Line segments which were desig-nated by noting the concavity line se~ments in the first cetegory group so as to se ect the second concavity line segment as the first concavily line se~ment, arc supplied i to the mr~tching unit 4. Cn the otner hand, as shown n ~i~. 17b, the category feature including the corresponding deformation is provided ~ith a specific mark and the rnas~
for the iine segments corresponding to the secsnd-category groups are stored in the category feature of the dictionary.
When a match ls attempted between the input parameters of feature and the dictionary in the matching unit 4, in the absence of the specific mark, a match i5 attempted between the first-category group and the dictionary, and the deci-sion is made in the similar manner as mentioned previously.
'when there is the specific mark, the control is effected in accordance with the presence or absence of the specific mark so that a match is attempted between the parameters of feature on the second-category group and the dictionary in addition to the specilic mark. In this way, no addition-al dictionary will be required by a deformed charactcr and the memory capacity will be reduced.

Claims (2)

What is claimed is:

1. In a pattern recognition system comprising a first step whereby while tracing the contour of a two-dimensional character pattern of a character scanned by an optical scanner and stored in the form of a binary quantized signal in a two-dimensional memory, the extraction of outermost points in a plurality of predetermined directions is accom-plished for each of points traced on said contour from a starting point of said tracing; a second step of extracting the information obtained by said first step as an outer-most point information associated with a series of contour lines forming said contour; a third step whereby in accord-ance with said outermost point information obtained by said second step said contour lines are each segmented as a segment forming a convex line segment, concavity line segment or hole segment in the order of tracing thereof so as to extract parameters of feature of each said segment;
and a fourth step whereby a match is attempted between the parameters of feature of said line segments extracted by said third step and a plurality of predetermined feature vectors of a dictionary so as to decide said two-dimensional character pattern, the improvement wherein said first step comprises measuring the distance of each of said traced points on said contour from said starting point and simul-taneously producing an integrated value of the coordinates on said contour of each of said traced points from said starting point, and wherein said fourth step comprises attempting said match between said dictionary and only those of said line segments excluding the convex line segments which are smaller than a predetermined length.

2. A system according to claim 1, wherein said third step comprises extracting a first feature group comprising the parameters of features of line segments having line segment names given in the order of detection thereof starting from the first-detected concavity line segment, and a second feature group comprising the parameters of feature of line segments having line segment names given sequentially starting from the second one of the concavity line segments in said first feature group up to the line segment next to said first-detected concavity line segment, and preliminarily giving a special designation to each of the informations stored in said dictionary and associated with characters each of which anticipates a deformation tending to change the starting point of tracing thereof, and wherein said fourth step comprises attempting a match between said parameters of feature and the content of said dictionary having said special designation with respect to each of said first and second feature groups.