| 摘要 |
916,408. Automatic reading of cursive script. WESTERN ELECTRIC CO. Inc. May 18, 1961 [May 31, 1960], No. 18156/61. Class 106 (1). Apparatus for automatically reading cursive script comprises means for generating electric signals representative of the positions of the writing instrument, means for detecting from the signals shape-elements of characters and for storing them in horizontal order, means for separating the stored elements into characters and for comparing those relating to each character with stored standard element signals to obtain the identities of the characters. As shown in Fig. 1, the writing is made upon a surface with parallel guide lines, an upper line A, a base line B and upper and lower limit lines U and L. The writing instrument is connected to a telewriter 31, Fig. 3, which produces, at every instant, voltages Ex(t), Ey(t) representing by their amplitude the horizontal and vertical co-ordinates of the writing instrument. These voltages are converted in quantizer 32 into binary numbers, 8-bits for the X co-ordinate and 5 bits for the Y. Instead of digitising the co-ordinate signals they may remain as analogue signals and all subsequent circuitry will be designed accordingly. The signals, discrete or analogue as the case may be, are applied to a buffer store 33. The signals are stored in a random manner with respect to time but are sampled at regular time intervals. Sequences of signals which are identical, as would result from the writing instrument dwelling at the same place for a period of time, are ignored. Identical following signals are detected by a delay unit and a comparator in each of the X and Y lines. The signals are then applied to pattern analysis circuits 34 which produce signals representing particular shape elements. These are as follows : (1) Vertical extent. A character may lie wholly between lines A and B, Fig. 1, go into the upper or lower space or both. (2) Auxiliary marks. Characters " i," "j," " t " and " x " have dots, a cross and a diagonal slash respectively. (3) Retrograde motions. Some characters, e.g. " a," " b," " g " and " f " have points at which the writing instrument turns back in the X direction. (4) Cusps. These are sharp changes of direction as in the character " i." Cusps are further classified with respect to their height in the character and whether it occurs before or after a highest point (Y max.) of the character and how many cusps have been detected. (5) Closure. A closure always occurs in some characters such as " 0," sometimes in others such as " b " and never in the rest, as " C." These are also classified with respect to their position in the character and whether or not other features exist. The expected distribution of the various features for the alphabetic characters is shown in Fig. 6. The writing surface is divided with 32 units in the Y direction and 256 in the X direction as shown in Fig. 4. A storage matrix 35, Fig. 7, is provided having a flip-flop for each position of this array, the rows being allotted each to a different feature and the columns correspond to the 256 X positions of the writing instrument as it writes a line of characters. When a feature is detected it enters a " 1 " in the appropriate row in the column corresponding to the position in the X direction where the feature was detected. The incoming signals to feature block 34 are continuously scanned for upper and lower extremes as at a and d, Fig. 12. Changes in the direction of Y are found by delaying the signal successively in four delay units and subtracting the input and output signals of each. The subtractors are arranged so that when rising Y values are in two delay lines and falling values in the other two an And gate emits an extreme signal a. An oppositely poled set of subtractors detects point d. It is desired to find the points b and c where the line making the extreme crosses the middle line (Y=16) after or before the extreme. This is useful in separating the characters one from another. The extreme signal is therefore stored and when next Y=16, a " 1 " is written into the corresponding X column in row 4. The slope of the character is therefore of no consequence for the determination of point b. For the lower extreme d which comes after the centre line crossing it is necessary to have the crossing point stored (row 16, Fig. 7). If a lower extreme occurs after a centre crossing, the position of the crossing is recorded in row 6, Fig. 7. Other features are extracted by similar methods. Separating characters.-A. number of different ways are used in blocks 36-40, Fig. 3, to determine the separating points which are illustrated in Fig. 21. In line C marks are shown at the centre of each vertically extended retrograde stroke or special mark (dot). These approximate in most letters to the centre of the letter and serve as a suitable landmark when finding the character dividing points. The average width of a letter is estimated and half this placed on each side of the marks to indicate separation points. The width is obtained by dividing the word length by the number of letters. The number of letters is roughly determined by counting the number of centre line crossings and dividing by three since, on an average, each letter cuts it three times as shown in line D. This separates nearly all the letters. In line F the " a " and " e " both have retrograde points at the top, i.e. where the line turns back and the " t " has an extreme (+Y max.). Corrected for slope. By placing a ¢w. on each side of these points accurate separation lines are obtained. The word " tug " in line 9 is cut up accurately in a similar way although the letter " U " has no landmark. Where words have two or more adjacent letters without landmarks the space between separation points is measured and divided into two, three, four, &c. according to its size. The length of the word is estimated by scanning out row 23, Fig. 7, of the store in which a " 1 " was sorted for each position having a line trace. Where a succession of five or more " 0's " are found a gap between words is assumed. Gaps of less than five spaces are ignored as shown in Fig. 22. The length 2A of the word is divided by the number centre line crossings divided by three and half this number gives the half letter width ¢w. which is stored in rows 20-22. Separation marks are written into row 27, Fig. 7,. by the store distribution which scans out rows 4, 6, 10 which give character points. When such a point is reached the distributer is reversed for a number of steps in the X direction equal to ¢w and a separation mark is recorded. The distributer then advances a distance w and makes another mark. This is repeated for each letter, where there are separation marks close together the space between is ignored. Other landmarks in letters are derived from the dot, and slash marks. Recognition.-The features recorded in store 35 between each pair of separation marks should belong to a single letter. They are read out and compared in block 24, Fig. 3, with stored shape element data for each possible character based on the table, Fig. 6. The character showing the highest correlation is indicated as being the one read out from the store and it passes to an output device 26, e.g. a printer. U.S.A. Specification 2,925,467 describes a suitable telewriter 31. |
