Apparatus and Method for Recognizing Characters

1. Field of the Invention

The invention relates to the field of character recognition systems, U.S. Class 340/146.3.

2. Prior Art

There is a continuing demand for character recognition systems, particularly those which are able to discern ordinary printed alpha-numeric symbols. The most common approach to character recognition compares, in a two-dimensional array, the viewed characters with predetermined forms. Often this comparison proceeds by examining predetermined areas of the viewed character to determine if they are dark or light. This two-dimensional, sectional analysis is shown graphically in U.S. Pat. No. 3,964,022, FIGS. 9a-9f. This particular system employs a 32.times.32 self-scanned, photodiode array. The video data is digitized and then stored in an appropriate memory. One problem with this type of system is that in addition to requiring large memory storage, it attempts to handle all the information for a particular character at one time. That is, the entire two-dimensional array is considered at one time. This results in poor utilization of circuitry.

In U.S. Pat. No. 3,873,972 another character recognition system is disclosed. In this case, a linear array is employed for scanning the characters as shown in FIG. 2. However, the video information from this array is accumulated in the scratch-pad memory 42. Then both horizontal and vertical features are detected and used to identify characters. Thus, even though this system employs a linear array, it is again a two-dimensional matrix approach which also stores unprocessed video.

As will be described, with the invented apparatus and method the raw video data is not stored, but rather is analyzed as it is read from the linear array. The recognition of characters is done more in a "pipeline" manner which permits better utilization of the circuitry when compared to prior art approaches.

An optical reader for recognizing printed characters employing a hand-held wand, in the presently preferred embodiment, is described. The wand includes a four-diode-wide linear array which electrically scans the characters in vertical slices as the wand is moved across a field of characters. The video signals from these four channels are weighted to provide two channels of video data. Depending upon the horizontal direction of movement of the wanding across the characters, one of these channels is selected as the primary channel. The raw video data is first processed in a video processor which includes a dynamic thresholding means. This thresholding means provides compensation for the wide dynamic range of video signals received from the photodiode array. A feature analyzer is then employed to analyze the video signals in the primary channel for a plurality of predetermined features such as strokes, bars, large bars and gaps. A comparison of the video data in both the channels is made to determine the direction of character development (upper or lower). A digital word is encoded with the detected features for each scan. A state machine is next used to compare each word with the preceeding word to detect new words; each new word is stored in a first-in, first-out register. A logic tree-type analysis is conducted in a digital computer on each word from this register to determine the possible characters possessing the extracted features. As each word from the register is compared with words from the program memory, the possible number of characters having those features is narrowed until a single character is identified.

The apparatus also employs a zone control means for selecting a portion of the video information from each slice for processing. This zone control means compensates for the fact that in handwanding, the characters may not be in the center of the vertical field-of-view of the photodiode array. As the wand is moved from character to character, the zone control means determines where in the field-of-view the character is to appear, and then prevents the processing of data outside this zone.

FIG. 1 is a general block diagram of the invented apparatus.

FIG. 2 is a flow chart illustrating the invented method for recognizing characters, which method is implemented by the apparatus of FIG. 1.

FIG. 3 is a block diagram of the video processor of FIG. 1.

FIG. 4 is an electrical schematic of a portion of the video processor of FIG. 3.

FIG. 5 is a graph of waveforms associated with the video processor of FIGS. 3 and 4.

FIG. 6 is a block diagram of the feature processor of FIG. 1.

FIG. 7 is a block diagram of the feature analyzer portion of the processor of FIG. 6.

FIG. 8 is a flow chart of the logic implemented by the feature analyzer of FIG. 7.

FIG. 9 graphically illustrates the detection of size features of characters.

FIG. 10 graphically illustrates the detection of gap features of characters.

FIG. 11 graphically illustrates the upper/lower feature detection.

FIG. 12 is a diagram illustrating the encoding of digital words with the detected features.

FIG. 13 is a diagram illustrating the changing field of view provided by the automatic line tracking means of the invented apparatus.

FIG. 14 graphically illustrates the functioning of the logic tree means of the invented apparatus in a general form.

FIG. 15 is a portion of the logic tree implemented in the recognition means of FIG. 1.

FIG. 16 is a block diagram of the presently preferred embodiment of the recognition means of FIG. 1.

An apparatus and method are described for recognizing characters such as alpha-numeric characters. As presently implemented, the apparatus and method are employed to recognize the standard OCR-A characters. However, it will be obvious to one skilled in the art that the apparatus and method taught by this application may be employed to recognized characters of other fonts.

In the following descriptions, numerous specific details are provided, such as specific numbers of divides, bits, etc., in order that the invented apparatus and method be thoroughly understood. It will be obvious to one skilled in the art, however, that the invented apparatus and method may be practiced without employing these specific details. In other instances, well-known components, techniques, etc., are not set forth in detail in order not to obscure the invention in unnecessary detail.

In the presently preferred embodiment, the characters such as alpha-numeric characters are scanned by a photodiode array which is contained within a hand-held wand. The wand is moved horizontally across a field of characters (in either direction) while the photodiode array is scanned to provide video signals representative of the viewed characters. In the presently preferred apparatus, the wand includes a light aligning means to assist the operator in aligning the wand with the field of characters. This light alinging means is described in copending application, Ser. No. 918,814, filed June 26, 1978 which is assigned to the assignee of this application. The invented apparatus and method may be employed with other than hand-held optical reading means, particularly those which provide relative motion between a photodiode array and the printed characters.

The wand includes incandescent bulbs which illuminate the characters. The reflected light from the characters is focused through a lens system onto the photodiode array. The photodiode array is electrically scanned to provide serial video signals. As best illustrated in FIG. 1, the photodiode array 22 is generally aligned vertically with respect to the characters such as character 20 (letter E). The movement of the wand provides relative motion in either the directions 28a or 28b. As will be explained in greater detail, automatic line tracking electrically locates the characters within the vertical field of view of the linear photodiode array 22. This field of view is shown by the dimension 23 of FIG. 1.

In the presently preferred embodiment, the photodiode array 22 consists of a 65.times.4 photodiode array which is commercially available from Reticon Corporation, Sunnyvale, California. This array is continuously scanned at a rate of approximately four microseconds per diode. Each scan or slice consists of 65 data cycles plus 2 dead time cycles giving a full cycle time of 268 microseconds per slice. The output of each of the sections of the array shown as lines 24, 25, 26 and 27 are coupled through amplifiers (not illustrated) to a video processor 30.

While in the presently preferred embodiment a four-diode-wide array is used, it will be apparent from the following disclosure that a one-dimensional array may be employed where the detection of movement is known. The use of the four-diode-wide array is primarily to improve the recognition reliability by compensating for typical printing imperfections. Also, while an electrically scanned photodiode array is used, other light sensing means may be employed. For example, light sensitive regions coupled to a charge-coupled device (CCD) provide an equivalent video signal by "reading" all the regions at the same time and then serially shifting out the sensed readings. In general, the video processor 30 first weighs the video signals on lines 24, 25, 26 and 27 to provide two channels of video data as will be described in conjunction with FIG. 3. Thresholding means including a dynamic thresholding means (described in detail in conjunction with FIG. 4) are then employed to provide a light/dark indication from the video signals.

The two channels of video data from the processor 30 are coupled to a feature processor 36 by lines 32 and 34. As will be explained in detail in conjunction with FIGS. 6, 7 and 8, and as explained graphically by FIGS. 9, 10, 11 and 12, predetermined features of a character are identified from the video data. These features are: stroke, bar, large bar, gap, large external gap and an upper/lower feature indicator. For each scan or slice of a character taken by the photodiode array, an 8-bit digital word is encoded with the detected features.

The digital words are employed in a logic tree means to recognize the characters. As each new word is received from the processor 36, the recognition means 38 compares the features with known features. The matching of the detected features with the stored features directs the flow in a logic tree so as to lead to the recognition of a single character.

In FIG. 2 the invented method is shown in general form. Block 45 illustrates that the characters are first scanned in the vertical direction. Each slice of each character is then examined to identify predetermined features such as stroke, bar, gap, etc., as shown by block 46. For each slice of a character, the identified features are encoded into an 8-bit digital word (feature word). This is shown by block 47. Each word is then compared with the previous word to identify new words. These new words are accumulated as shown by block 48. Note that in a typical application, as the wand is manually moved over the characters, a number of consecutive, identical feature words results.

The feature words are employed to direct a course in a logic tree as shown by block 49. The specific program for this tree is set forth in Table II.

Referring now to FIG. 3 and the block diagram of the video processor 30 of FIG. 1, the four lines of video signals from the photodiode array are coupled to weighting networks 50 and 51. Specifically, lines 24, 25 and 26 are connected to network 50 and lines 25, 26 and 27 are connected to network 51. The output of the network 50 provides one channel of video data designated as Channel A. A typical waveform of the raw video at the output of network 50 is shown in FIG. 5 as waveform 80. (References to the waveforms of FIG. 5 are shown in circles in FIGS. 3 and 4.) Within each of the networks 50 and 51, three lines of video signals are combined in an ordinary manner to provide a single channel of video. The weighting within these networks is done to provide smoothing and averaging of the signals received from the photodiode array.

The output of the network 51, designated as Channel B, is processed in a similar manner as the output of network 50 and results in an output signal from the processor 30 on line 34. For sake of clarity, only the blocks associated with Channel A are shown. In typical printed characters, numerous defects in terms of the consistency and density of the ink occur. Ink "voids" are detected as white areas by the photodiode array. By weighting the signals from three channels, compensation is provided for these defects. Similarly, excessive ink which appears on intended white areas causes unwanted detections. The weighting networks also compensate for these "spots."

The output of the weighting network 50 is coupled through a lowpass filter 53. The lowpass filter, which may be an ordinary filter, is employed to remove part of the DC component from the raw video signal.

The output of the lowpass filter 53 is coupled to a zone selector 54. The zone selector is used to couple a section of the video signal from each vertical scan or slice to line 57. The zone selector 54 receives a control signal on line 56; a typical waveform for this control signal is shown as waveform 81 of FIG. 5. A typical waveform at the output of selector 54 (line 52) is shown as waveform 82 of FIG. 5. An ordinary switching means may be employed for the zone selector 54. The generation of the zone control signal will be discussed, in detail, in conjunction with the automatic line tracking means.

The use of a multi-element photodiode array has the disadvantage of nonuniform sensitivity from element-to-element, and also from line-to-line. Moreover, such devices are often temperature sensitive, and thus drift. The DC level of the video signal tends to drift also because of the nonuniformity of the whiteness of the paper and the changes in the lighting levels from the incandescent bulbs used to illuminate the characters. The base line tracking circuit which includes amplifiers 58, 59 and 62, constant current source 60 and capacitor 61 substantially reduces the effects of these variations and also enhances the signal-to-noise ratio.

Line 57 is coupled through amplifier 58 to the constant current source 60, capacitor 61, and the input terminal of amplifier 59. The output of amplifier 59 is coupled to one input terminal of the amplifier 62. Line 57 is also coupled directly to the other input terminal of amplifier 62.

The amplifier 58 tracks the DC level on line 57. This amplifier has a relatively slow time constant, and thus does not track the data from the photodiodes contained on the DC level. Capacitor 61 is charged to the DC level tracked by the amplifier 58. The capacitor 61 is slowly discharged by the constant current source 60. The current source 60 enables the capacitor to track a negative signal by slowly discharging the capacitor at a constant rate. In this manner the capacitor may follow either a positive or negative signal. The charge on the capacitor 61 is buffered by the buffer amplifier 59. The DC level at the output of the buffer amplifier 59 is subtracted from the video signal on line 57 within the differencing amplifier 62 to provide the video signal less its DC component on line 64. The waveform of the signal at the output of the differencing amplifier 62 is shown as waveform 83 of FIG. 5.

The remainder of the circuitry of FIG. 3 is used to discern the difference between light or dark, that is, whether the photodiode is observing printing or the background paper. The signal on line 64 is applied to a threshold limiter 65 which establishes a general threshold level at node 75. The dynamic threshold limiter 66 is a faster reacting limiter, as will be described in greater detail in conjunction with FIG. 4, and provides a dynamic threshold level across the resistors 69 and 70. The common junction between these resistors is coupled to an input terminal of amplifier 71. Node 75 is also coupled to this terminal via the diode 68. The other input terminal to the amplifier 71 receives the signal on line 64. The output from the amplifier 71 is coupled to one input terminal of the AND gate 73. The other input terminal of the AND gate 73 is coupled to the output of an enabler 76.

In general, the thresholding circuit operates by comparing the dynamic threshold provided by limiter 66 with the signal on line 64. The limiter 65 provides a clamping voltage to the amplifier 71 to prevent the dynamic threshold from dropping below a predetermined level. The enabler 76 prevents output signals through the gate 73 when there is weak video; this prevents detection of noise.

Referring now to FIG. 4, the presently preferred embodiment of the thresholding means is shown in detail. Line 64 is again shown along with the output line 32. The dynamic threshold limiter 66 of FIG. 3 is realized as a medium speed peak follower. The signal on line 64 is coupled to one input terminal of a voltage comparator 89. The other input terminal of this comparator is coupled to its output, one input terminal of an operational amplifier 91, and to the collector terminal of transistor 95. The emitter terminal of transistor 95 is coupled to ground through a resistor 96 and a capacitor 99. The common junction between this resistor and capacitor is coupled to the base terminal of transistor 95 through a diode 98 and a resistor 97. This transistor terminal is also coupled to ground through a resistor. The output of the operational amplifier 91 is coupled to one of its input terminals and also to the voltage divider comprising resistors 69 and 70 which are also shown in FIG. 3.

The threshold limiter 65 of FIG. 3 is realized as a slow peak follower and includes a voltage comparator 88, one input terminal of which is coupled through a resistor to lines 64. The output terminal of the comparator 88 is coupled to the other input terminal of this comparator and to one input terminal of an operational amplifier 90 through an RC network. The output terminal of the amplifier 90 is coupled to the other of its input terminals, one input terminal of voltage comparator 92 and to a voltage divider network comprising resistors 101 and 102. The common junction between these two resistors is coupled through the diode 68 to the common junction between resistors 69 and 70.

The enabler 76 of FIG. 3 is fabricated from a comparator 92. The output terminal of the comparator is coupled to one input terminal of an NAND gate 106. The output of the amplifier is also fed back to its input terminal through the parallel combination of capacitor 108 and resistor 107. This input terminal is also biased through resistors 111 and 112.

The comparison function performed by the amplifier 71 of FIG. 3, in the presently preferred embodiment, is performed by voltage comparator 93 which has one of its input terminals coupled to receive the divided output from amplifier 91. The other input terminal to this comparator receives the signal from line 64; this input terminal is also coupled to the output of the amplifier through a feedback path consisting of the parallel combination of resistor 103 and capacitor 104. The output of the amplifier 93 is coupled to the other input terminal of the gate 106 through gate 105; gate 105 is connected to operate as an inverter.

In the presently preferred embodiment, the amplifiers 90 and 91 are commercially available operational amplifiers, in particular, Part No. TL 082 available from Texas Instruments. The voltage comparators 88, 92 and 93 are commercially available comparators, Part No. LM 339. The comparator 89 again is a commercially available part, Part No. LM 319. The resistors 69 and 70 have the value 4.32 K and 10 K, respectively.

The medium speed peak follower follows the peaks of the video signal as is best shown in FIG. 5 by the waveform 84. As may be seen, the follower is fairly responsive to the peaks of the waveform 83; however, after each peak, the output of the amplifier 91 begins to decay. The waveforms 83 and 84 are compared in comparator 93. The waveform 84 is shown as the dotted waveform 85 in FIG. 5, superimposed on the waveform 83 to show the manner in which the light/dark contrasts are detected. Note that the peaks of the video signal for the illustrated example exceed the dynamic threshold sufficiently to cause an output signal to occur. The output signal at gate 106 is shown as waveform 87 and is identified as the detected video signal.

The slow peak follower more slowly follows the peak levels of the video signal and provides a signal through the diode 68 to clamp the dynamic threshold to a minimum level to prevent the detection of noise when larger time intervals occur between peaks. When no signal is present, the enabler 76 through gate 106 prevents any signal from appearing on line 32, thus essentially turning off the channel to prevent the detection of random noise spikes which might otherwise be detected.

It should be mentioned that substantial variations typically occur in the black densities and background white densities. Because of this, the video signal can vary over a considerable dynamic range. The dynamic thresholding means above described compensates for this and permits the reliable detection of the dark/light areas.

The detected video signals for both Channels A and B, lines 32 and 34 respectively, are coupled to the feature processor 36. In the feature processor 36, the video waveforms are examined for each slice for the predetermined features. Before discussing the structure of the feature processor 36, the specific features extracted from the video signal, and the encoding of the feature words, shall be described in conjunction with FIGS. 9 through 12.

One of the features for which an examination is made is referred to in this application as "sizing" (FIG. 9). Sizing consists of an examination of the dark areas, that is, the printing, for its relative size in the vertical direction. Three separate sizes are detectable, in the presently preferred embodiment, and are identified as "stroke," "bar," and "large bar." A stroke is a relatively short, dark area; a bar, a longer dark area; and a large bar, a dark area which extends approximately the entire height of the character.

In examining slice 188, which is drawn through the large numeral 2 in FIG. 9, assume that this is a slice or scan of the numeral 2 observed by the photodiode array. The uppermost extending portion of the numeral 2 is relatively short, and is detected as a stroke. The lower portion is longer and is detected as a bar. For the slice 189, three relatively short dark areas are traversed, and thus a stroke/stroke/stroke is detected. Examining the numeral 9 shown in FIG. 9, the first slice through this numeral would result in the detection of a bar; the second, two strokes; and the last, a large bar. For the numeral 7, the one slice shown through this numeral will be detected as a stroke.

In FIG. 9, the relative threshold values for the stroke, bar and large bar are illustrated as amplitudes. The dimension 190 is the amplitude for a stroke; dimension 191, the amplitude for a bar; and the dimension 192, the amplitude for a large bar. Amplitudes smaller than 190 are discarded as noise. Note that while the threshold values for stroke, bar, and large bar are shown as amplitudes in FIG. 9, they are actually sensed in terms of time. That is, the video signal is examined and the duration of a dark image is sensed to determine if the image should be declared a stroke, bar or large bar. As will be discussed in greater detail, the duration is determined by the running of a digital counter which is stopped at the end of the dark image. The contents of the counter are then compared with the three possible threshold values to determine if the image qualifies as a stroke, bar or large bar.

A second set of featues for which the characters are examined is referred to in this application as "gaps." Two types of gaps are sensed. One is an internal gap, which is shown by slice 194 in the letter O of FIG. 10. The other is a large external gap such as the one shown for the letter U, slice 201.

Examining the slice 193 which passes through the letter P, the relatively short gap between the two strokes is ignored, in the presently preferred embodiment. Referring to the internal gap threshold values shown in FIG. 10, the gap traversed by slice 193 is referred to as a "nominal" gap, dimension 204. The slice 194 through the letter O includes two strokes, however, this time separated by a larger gap. This larger internal gap is detected and categorized as a feature. The threshold for this gap is shown by the dimension 205 in FIG. 10.

In the letter M of FIG. 10, the slice 195 first encounters a stroke at the upper part of the letter, then a relatively long light region which is classified as a large, external gap. The relative threshold for this gap is shown by dimension 206 in FIG. 9. The second slice 196 through the letter M first encounters a short gap at the upper part of the letter, then a stroke, and then another gap. This second gap is not large enough to qualify as a large external gap, but rather is detected as a gap. The threshold level for the external gaps are shown by dimensions 206 and 207. Note that the threshold level for an external gap is lower than the threshold level for an internal gap (compare dimensions 205 and 207.) However, for purposes of feature extraction, these two gaps are considered alike. In the last example shown in FIG. 10, the slice 201 through the letter U first encounters a large gap, followed by a stroke.

As in the case of sizing the dark areas, the gap thresholds represent time. The duration of the light areas is sensed in a digital counter and compared with stored values, as will be explained later.

The third type of feature extracted from the video signal is shown in FIG. 11; this feature is an indication of the direction of character development. It should be noted that for the sizing and gap feature extraction, the information from only a single channel is employed. For the direction of character development, the information from both channels is compared.

Referring first to the numeral 6, two slices 197 and 198 are shown. Assume that the direction of travel of the photodiode array is indicated by arrow 209, and that the primary channel is the trailing channel (slice 197). Slice 197 shows a stroke/stroke/stroke; slice 198, a stroke/stroke. When the data from these two channels are compared on a common time axis, it can be determined that the upper part of the character is not sensed by slice 198. The character must thus have some upwardly extending feature. This is declared an "upper" feature. Examining the letter P and the slices 199 and 200, a stroke/stroke/stroke of the primary channel is compared with a stroke/stroke of the future channel. The common time axis comparison indicates that a feature has been lost in the lower part of the letter. This is declared a "lower" feature. Two examples are shown in FIG. 11 with variations on the backward letter C. Again, assuming that the wand is moving in the direction indicated by arrow 201, the first case results in an upper feature indication and the second, in a lower feature indication. In the case of the letter O of FIG. 11, slice 215 shows a large bar, and slice 216, a stroke/gap/stroke. By comparing these two sets of features, it is apparent that features are developing on both the upper and lower part of the character. This is declared "normal."

The value of the upper/lower feature can be shown by an example. Assume that the wand is traveling in the direction indicated by the arrow 210, and that four slices 211, 200, 199 and 212 are made through the letter P. The feature detection results in: bar; stroke/stroke; stroke/stroke/stroke; and bar. Assume that four slices 213, 198, 197 and 214 occur through the numeral 6. The resultant feature detection at these four slices is: bar; stroke/stroke; stroke/stroke/stroke; and bar. (No gaps are detected for either of the last two examples.) Note that the detected features are the same for both the 6 and P. Thus, the feature detection without the upper/lower determination is ambiguous for some characters. With the upper/lower feature determination, the difference between a 6 or P, in addition to other ambiguous cases, may be discerned.

A single channel of video data may be used for the upper/lower feature determination. This may be done by storing the video signal from slice-to-slice and comparing them with one another. However, more precision is obtained by using two channels since the precise spacing between the diodes is known.

For each slice through a character, the feature processor 36 of FIG. 1 encodes an 8-bit word with the detected features. As shown in FIG. 12 by the 8-bit word 217, the first two bits are employed for the upper/lower/normal designation with the codes indicated; 00 for normal, 01 for lower, and 10 for upper. The second two bits are used to designate the first feature, the next two bits, the second feature, and the remaining two bits, the third feature. The features are coded as indicated; null or void as 00, gap as 01, stroke as 10, and bar as 11. Since three bars cannot occur in a single slice, the 111111 code is used for a large bar. Also, since the two consecutive gaps cannot be detected, the 0101 code is used to designate a large (external) gap. Four examples are shown in FIG. 12. Note in each case, it is assumed that the direction of travel is as indicated by arrow 218, and that the trailing slice is the primary slice.

The features for the above described coding are furnished to the register 146 of FIG. 6 as will be described in greater detail.

The above-described features have been selected as a result of an empirical analysis. However, other features could be substituted for those employed in the presently preferred embodiment, such as the small gaps.

Referring now to FIG. 6, the major portions of the feature processor 36 of FIG. 1 include a direction detector 124, feature analyzer 130, Y-counter 121, zone controller 140, upper/lower feature detector 139, and the circuitry to the right of the dotted line 157, identified as the width and state analyzer.

In the presently preferred embodiment, the wand may be moved in either direction across a field of characters. The direction detector 124 determines the direction in which the wand is moving. The detector 124 receives the video signals from the video processor 30 on lines 32 and 34. This detector also receives a field reset signal on line 136; this signal is generated at the end of an entire field of characters. The signal detector 132, which is a part of the direction detector, receives the data on lines 32 and 34 and determines on which line the video signals first appear. An ordinary bistable circuit may be employed for detector 132. As soon as this determination is made, a flag 124 is set and a signal applied to line 127 to indicate the direction of the wand over the characters. This signal is used within the logic tree means and is crucial in recognizing certain characters. For example, the difference between a "3" and a "E" is determined by the direction of travel of the wand.

The signal detector 132 sets a switch 133 as indicated by the dotted line 134. The switch is set immediately upon the determination of the direction. For the presently preferred embodiment, the leading channel is coupled to line 126 and is identified as the "future" channel and the trailing channel is coupled to line 125 and is identified as the "primary" channel. The video signals in the primary channel (line 125) are coupled to the feature analyzer which detects the bars, strokes and gaps. The primary channel and the future channel are coupled to the upper/lower feature detector 138 which detects the upper/lower feature.

The Y-counter 121 provides a digital count corresponding to the diode in the array which is being interrogated. In the presently preferred embodiment where a four-diode-wide array is used, four horizontally aligned diodes are simultaneously interrogated at each count. The Y-count provides Y-position information since when it is compared with the video signal, the Y-position of the detected light/dark region is known. The Y-count is coupled on line 122 to the feature analyzer 130 and to the comparator 144 which is included within the zone controller 140. An ordinary digital counter may be employed for counter 121.

Before describing the struccture and operation of the zone controller 140, an understanding of its function is helpful; for this purpose, reference is now made to FIG. 13. The zone controller permits the automatic line tracking which is of particular value in the case of a hand-held wand. As the hand-held wand is moved over a field of characters such as the field shown in FIG. 13, the field-of-view of the wand (as defined by the parallel lines 220) may not remain parallel to the characters. In fact, misalignment is more apt to occur than not. The automatic line tracking provided by the zone controller 140 assures the selection of that portion of the video signal which contains the pertinent character data.

For the initial character in the field-of-view, such as the letter A shown in FIG. 13, it is assumed that the character is centered within the field-of-view as shown. At the end of the first character, a determination is made on the location of the character within the field of view. An adjustment is then made by the zone controller to select a different portion of the video signal for processing. This is done by establishing a new distance between the upper and lower edge of the field-of-view and the upper and lower zone limits shown as lines 221. As the field of characters moves either up or down within the field-of-view, the zone limits automatically track that movement.

The zone controller 140 includes a load control means 141 which loads digital numbers into either a register 142 or a register 143. The contents of registers 142 and 143 are compared in a digital comparator 144 with the Y-count provided on line 122. When the contents of register 142 matches the Y-count, the top of the zone has been reached, and a signal on line 56 so indicates. When the contents of register 143 matches the Y-count, the comparator 144 terminates the zone by removing the signal from line 56. The load control means 141 receives the field reset signal on line 136. When the signal is originally received, the load control means 141 loads the numbers into registers 142 and 143 which assume that the first character will be centered within the field-of-view. The maximum height of the characters is known; the height of the zone is this known height, plus an additional margin (6 counts on top and bottom as implemented). Thus as is best seen in FIG. 13, the letter A is within the initial zone, which zone includes a space above and below the letter A. For each character, the feature analyzer 130 provides a count (on line 222) indicating the upper and lower limits of the character plus the set margins. These numbers are loaded into registers 142 and 143 at the end of each character. Thus for each slice, the comparator 144 is able to provide a zone control signal. While in the presently preferred embodiment the zone limits are, in effect, updated for each character, they may be updated more frequently or less frequently.

The circuit for the feature analyzer 130 of FIG. 6 is described in conjunction with FIGS. 7 and 8 and Table 1. The upper/lower feature detector 138 may be fabricated employing the same general logic which will be discussed for analyzer 130 or by employing well-known circuit designs. Two of the possible conditions of the primary and future channels are shown within the block representing detector 138 along with the resultant coding. The other possible conditions are apparent from considering FIG. 11.

The 8-bit digital words (feature words) are coupled from the feature analyzer 130 and the upper/lower feature detector 138 into the feature register (F register) 146. The 2 bits for the upper/lower designation are coupled on line 150 to the first two stages of this register; the remaining 6 bits are coupled on line 131 to the other six stages. The contents of the register 146 are shifted (on command from comparator 148) to an 8-bit register 147. The contents of the registers 146 and 147 are compared by the comparator 148. The contents of the register 147 are shifted (on command from comparator 148) into a first-in/first-out (FIFO) register 154. The register 154 is able to store 8-bit words from the register 147 in addition to the other information as will be discussed. (In the presently preferred embodiment, register 147 is 12 bits wide and 64 bits deep.) The comparator 148 is coupled to a digital counter 152 which counts at the rate of the scans. Each time a word is shifted by the comparator, the counter 152 is reset.

In operation, assume that a digital word encoded with the features detected from the first slice of a character is in register 146. The contents of this word are compared with the contents of register 147. Since the contents of this register are cleared at the beginning of each character, the comparator indicates a difference between the word contained within register 146 and the "blank" contained within register 147. The word in register 146 is shifted into register 147. When this shift occurs, the comparator 148 resets counter 152. The feature word from the second slice is next placed within the register 146. This word is compared with the first digital word now stored in register 147. If the words are the same, no shifting occurs. In practice, the same feature word occurs a number of times and this number is determined by counter 152. When the contents of register 146 are different from those in register 147, the comparator 148 detects this difference and shifts the word from register 147 (along with the contents of counter 152) into the FIFO register 154 and also shifts the contents of register 146 into register 147. Thus register 154 receives each new feature word along with a count which indicates the relative width of the encoded feature.

The contents of the register 156 represent the video data in a condensed, digital form. Note, no storage of video data occurs. In fact, the video data is analyzed for features in a serial or peripheral manner as it occurs. In the presently preferred embodiment, the feature words are removed in a first-in/first-out basis on line 156 as fast as they may be processed by the recognition means of FIG. 1. This digital information, however, may be stored for other uses. For example, it may be stored or transmitted with the recognition analysis performed at a later time or at a different location. The data from the register may be employed to provide a visual display of the characters without recognition of the characters. This may be done by using the information from the words to display strokes, gaps, bars, etc.

Referring to FIG. 7, the block diagram of the feature analyzer, the upper part of the diagram includes a pair of registers 164 and 165. These registers are coupled to the comparators 166 and 167, respectively. The Y-count on line 122 is also coupled to these comparators. At the beginning of each character, register 164 is loaded with the digital number representing the anticipated top of the character and register 165, with the number representing the anticipated bottom. For the first character of the frame as shown in FIG. 13, it is assumed that the character will be centered within the field-of-view. When the Y-count is greater than the number stored within register 164, a signal is present on line 168. When the Y-count is greater than the number stored within register 165, a signal appears on line 169. The use of these signals is discussed in conjunction with FIG. 8.

The actual detection of the strokes, bars, large bars, gaps and large gaps are made by counting the number of "light" or "dark" diodes and then comparing this number with the threshold values shown graphically in FIGS. 9 and 10. In FIG. 7, counter 170 is the black feature size counter which is referred to in FIG. 8 and Table 1 as "BCTR". This ordinary digital counter counts upon receiving a signal on line 177, and is reset when receiving a signal on 176. The contents of the counter are coupled to a comparator 173. The predetermined threshold values for stroke, bar and large bar are also coupled to this comparator. The controller 162 senses the presence of either a dark or light region from the primary channel video signals. If a dark region is encountered, counter 170 is caused to count. When the dark region ceases to be detected, the count is compared with the various threshold levels by comparator 173. An appropriate signal is then coupled to the feature encoder 175 indicating a stroke, bar, etc.

The gap detection employs a digital counter 171 which is referred to in FIG. 8 and Table I as "WCTR". The contents of this counter are coupled to a comparator 174 as are the threshold values shown in FIG. 10 for gaps. When a light region is sensed, the counter 171 begins to count under the control of a signal received on line 179 from the controller 162. When the light region is no longer sensed, the count in the counter 171 is compared with the threshold values to determine whether a null, gap or large gap has been detected. The counter 171 is then reset by a signal provided on line 178. An appropriate signal is coupled to the encoder 175 to indicate whether a null, gap or large gap has been detected. In the case of gap detection, it is necessary to also know whether strokes or bars have previously been detected in order to declare the proper gap (because of the different thresholds for internal and external gaps). The logic to accomplish this is described in conjunction with FIG. 8.

The feature encoder 175 may be an ordinary encoding means for implementing the coding shown in FIG. 12.

The logic implemented by the feature analyzer controller 162 of FIG. 7 is best understood from the flow-chart of FIG. 8. The feature control state (table of FIG. 8) is set forth below in the Specification as Table I. A one-to-one layout exists between the table of FIG. 8 and Table I. For example, block 184 is the upper lefthand block of Table I. The possible branches in FIG. 8 (A through H) are also shown in Table I. In implementing the logic set forth in Table I, each of the blocks, such as blocks 184, 185, 186 and 181, controls the counters 170 (BCTR) and 171 (WCTR) of FIG. 7. Furthermore, each block is capable of performing the comparison test indicated in that block and is able to provide direction to the next appropriate block. Note that the first two bits of the feature word in Table I use the code 11 to indicate an idle state or an error.

Assume that a scan is initiated (YC=0). With YC=0 and commencing on line 182, the first count occurs and the information from the first diode is obtained. Following along line 183, in every case for the first diode, branching occurs on Branch A which is designated "Before Top of Character". If the first diode senses a light region, then the information in block 184 applies. With YC>Y TOP, which is indicated by the signal on line 168 of FIG. 7, branching occurs to Branch B and a WCTR reset occurs (see Table I). Then, if the second diode senses a light region, the WCTR counter is advanced (block 185). On the other hand, if the second diode or a subsequent diode senses a dark region (block 186), the BCTR counter is set and the contents of the WCTR counter are examined. If the contents of that counter are less than 2, it is assumed that no gap occurred at the upper part of the character. If the contents of that counter are between 2 and 5, it is assumed that a gap occurred. Referring to FIG. 12, the second and third bits are set to 01. By continuing in this manner for each slice, the first, second and third features shown in FIG. 12 are determined. Block 181, as shown in Table I, implements the setting of the zone top and zone bottom. The appropriate counts are loaded into registers 142 and 143 of FIG. 6 by line 222.

The presently preferred embodiment for the feature processor described above is a hardware embodiment; however, the processor may be, and in fact has been, realized with a digital computer.

The recognition means which implements the logic tree, in the presently preferred embodiment, is implemented with a digital computer, specifically, a commercially available microcomputer sold under the Part No. 8080. The programming of this computer shall be described in detail.

It is possible for each of the new feature words stored within the FIFO register to be employed in the logic tree analysis as it is generated. However, in the presently preferred embodiment, the width count for each word is used to compute a normalized width which is independent of the rate at which the wand is moved over the characters. The feature words, along with the normalized width, are stored in a feature buffer. The logic tree is generally implemented by taking feature words from the feature buffer (in the same order that they are formed) and then comparing them with stored words. As a match occurs with a stored word, it determines the next subtree or branch in most areas. The feature words, which are removed one by one from the feature buffer, provide one input in the comparison process used with a selected tree operating code taken from a tree network or a table of tree commands. The current tree address is used to select the output that is employed in the tree analysis. More specifically the tree analysis consists of selecting a new tree command which is combined with a conditional advance to the next feature word in the feature buffer. At the completion of the recognition analysis, an output code is provided that indicates the character that has been selected.

An example of this process is shown in FIG. 14. The numeral 5 is shown with slices 224 through 229. Only those slices that provided a new word are shown in FIG. 14. Slice 224 results in the detection of the features upper/stroke/void (null)/void. This word has a relative width of 2 on a normalized basis. The next feature word consists of bar/stroke/void (width of 3). These features are extracted at slice 225. At slice 226, the predominant feature of this character which has a width of 14 occurs (stroke/stroke/stroke). At slice 227, the features lower/stroke/bar/void with a width of 3 is detected; and, finally at slice 228, stroke/gap/gap with a width of 1 is detected. At slice 229, the end of the character has been reached and no features are detected. Note that the slices 224 through 229 are not parallel; they more closely resemble the slices which actually occur with a hand-held wand.

Referring now to the partial logic tree shown in FIG. 14, as the analysis begins, commands are accessed at the starting point shown as node A. The first feature word in the feature buffer, specifically U/S/V/V, is compared with each of the stored words leading from node A, such as those shown in blocks 231 and 232. When this first feature word is compared with the word in block 231, a match occurs. No comparison is then made with the word shown in Block 232 or any of the other words leading directly from node A. Since a match occurred at block 231, the tree address is now changed to node B. At node B, the first comparison is made to determine if the second feature word from the feature buffer, (B/S/V), compares favorably with S/S/V. Since the comparison fails, the second feature word is next compared with S/V/G, and again, there is no match. In each case, the next operating code is selected, but the feature word does not change because a match has not occurred. When the next match occurs, the new address for node C is selected and placed in the current tree address register. This analysis continues from node C to node D with a match of the word S/S/S. When the tree address is changed to node D, a test is made to determine the width of the S/S/S feature. The "JW" is a code meaning "jump if the width is greater than shown." Specifically, the width 14 is examined to determine if the width is at least 6/16's of the total width of the character. Since it is, the address for node E is selected. Following this, two additional tests are made to examine the end of the character. At the end of these tests, the character is declared a "5". These end tests are generally used to confirm a particular character, rather than to recognize it. For example, if the photodiode array was truly parallel with the numeral "5," the last feature word S/G/G may not have occurred. In this case, the character would still have been declared a "5," however, without proceeding to node G, but rather from node F.

In FIG. 15 a portion of the actual logic tree is shown, specifically the subtree used to discern the difference between a "7" and a "fill" symbol (the greater than symbol). At node 235, the stored words with which the next feature word are compared are shown. (The notation "JR" indicates "jump reverse.") If, by way of example, a match occurs with S/V/V, the address for node 236 is selected. Note that at that node, if the normalized width is greater than 5, the character is declared a "7."

in the presently preferred implementation, the analysis does not always proceed forward, but rather in some instances a jump occurs back to a higher subtree, such as indicated by lines 237 and 238. In some instances, a match occurs when the feature word is any one of a plurality of predetermined words. For example, referring to block 239, a match occurs when the feature word is either "FGG" or "GFF." (F indicates a stroke or bar.) The "Z" notation indicates blanking. It will be obvious to one skilled in the art that the subtree of FIG. 15 may be implemented into a digital computer program or into hardware logic circuits.

The specific program employed in the 8080 computer for the basic logic tree used in the presently preferred embodiment is shown in standard 8080 language in Table II.

If a relatively inexpensive microprocessor, such as the 8080, is employed for the logic tree analysis, its speed in ordinary use is not fast enough to keep pace with a hand-held wand. To provide enhanced performance from such a commercially available microprocessor, a special tree access apparatus and method was developed.

In FIG. 16, a microprocessor is shown within the dotted line 242 which includes a CPU 245, a random-access memory 244, a memory access chip 246 and a timing chip 247. The RAM, CPU and memory access chip are coupled to the bidirectional data bus 248.

In normal computer operation, the program stored in the program memory 249, is addressed by a program counter or similar means which is included within the computer. However, in the described system, the computer itself does not directly address the memory 249, but rather memory addressing occurs through registers 252, 253 and 254. Each of these registers may be loaded with an address from either the computer or the program memory itself. Also, register 253 increments after it is read.

The A register 252 along with the B register 253 provides a 13-bit address for the memory 249 when selected by the multiplexer 251. The C register 254 supplies the 6 least significant bits of an address with the most significant bits being either 0000001 or 0000000 when selected by the multiplexer 251. Multiplexer 251 is controlled by the computer as indicated by the line 256. The loading of addresses into the A, B and C registers is also controlled by the computer as indicated by line 257. Thus the memory 249, which in the presently preferred embodiment provides 8K bytes of storage, may be accessed by (1) A and B registers, (2) C register, with the most significant bits equal to 000001, or (3) C register, with the most significant bits equal to 000000.

With the system of FIG. 16, only those addresses which need to be changed are processed by the computer. Most of the addresses are generated within the loop consisting of the memory 249, the portion of bus 248 leading to the A, B and C registers, and the register and multiplexer 251. Very highspeed scan rates of the memory are achieved by repeatedly accessing an address from the A and B registers. Each time an address is removed from the B register, it increments. In this manner, the normal protocol associated with incrementing an address through the computer is eliminated. This addressing scheme is particularly useful when the CPU is searching for a match. Each of the stored words leading from a node such as node 235 of FIG. 15, is sequentially accessed for a comparison by the computer with a feature word without the need of incrementing and address by the computer. Generally, the computer is not required for controlling the selection of an address until a match occurs. Also, memory pair look-ups require the loading of only a single address into the C register with the toggling of the two possible most significant bit patterns by the computer. It has been found that with the use of the registers shown in FIG. 16, an improvement by a factor of 4 to 5 in the recognition processing is achieved.

Thus an apparatus and method has been disclosed for the recognition of characters such as ordinary alpha-numeric characters. The raw video data which results from a plurality of vertical scans is immediately processed without intermediate storage. Predetermined features are extracted from this video data in a serial manner and these features are encoded into digital words. Each of these new digital words is used to direct the analysis in a logic tree. In the presently preferred embodiment, this logic tree is implemented in a digital computer which employs extra registers to enhance computer processing. ##SPC1## ##SPC2## ##SPC3##