JPH0225553B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a character extraction method for determining a rectangular character frame circumscribing a character to be recognized in a character recognition device, and in particular, for handwritten characters, characters that exceed a predetermined character frame and invade adjacent character frames. This relates to a character extraction method that distinguishes and recognizes . Optical character reading devices (hereinafter abbreviated as OCR) are used as an input method for information processing systems, but this device can recognize handwritten characters from early devices that recognize characters printed on typewriters or line printers using special type. It is developing into However, in handwritten characters, there are not only individual differences in character shape and size, but also characters that exceed the specified writing frame and enter the writing frames on both sides.
Difficulty arises when cutting out characters. Before describing the problems below, for ease of understanding, a conventional character extraction method will be briefly explained with reference to the drawings. FIG. 1 shows an explanatory diagram of the configuration of a character recognition device. In the figure, characters 2 written on a form 1 enter a sensor 4 through a lens system 3, undergo photoelectric conversion, undergo the necessary processing in a preprocessing circuit 5, are quantized, and are stored in an image memory 6. The processor 7 then performs character recognition processing. The first process performed in this character recognition process is a character extraction process. FIG. 2 shows the memory configuration of the image memory (FIG. 1, 6). In this example, one byte (8 bits) corresponds to one pixel as memory 8, and the configuration of each memory when a 64 Kbyte memory is used is shown by assigning a hexadecimal address to each area. In other words, in the example of this figure, the 1-pixel memory 8 in the upper left corner is address 0000, and as you move further to the right, the lower two digits change sequentially, becoming 00FF in the upper right corner, and when you change vertically, the upper two digits change. It changes sequentially starting from the 1 pixel memory 8 at the upper left corner, and the lower left corner becomes FF00.
The last address will be FFFF in the lower right corner. Next, Figure 3 shows characters 2 on form 1, the extraction base addresses 9-1 to 9-4 when those characters are extracted on the image memory, and the character frame determined by the extraction height and image capture height. 10 relationships are shown. In the figure, character 2 is lined up with the number 534 in this figure, with respect to the image capture height and image capture width (not particularly shown in the horizontal direction of the figure) determined by the character recognition device, and the character frame 10. There is. number 5 now
Considering this, in order to cut out that character, the cutting base points 9-1 and 9-2 are set as the left cutting base point and the right cutting base point, respectively. In the figure, when numerals 3 and 4 are sequentially extracted, the left and right extraction base points are sequentially searched to the left as 9-2 and 9-3 or 9-3 and 9-4. The extraction method will be explained using the number 5 stored as video information on the image memory as an example in FIG. In the figure, an image memory area with a character width of 3 bytes (8 bits/byte) and a cutting height of 40 bits is
A case will be described in which video information forming segments for each vertical scan within the character frame 10 is cut out. (However, in reality, the character width is 5 to 6 bytes.) In the figure, first, a search is performed in the vertical direction from the left extraction base address 9-1 for the number 5 (hereinafter referred to as vertical scan), and the first vertical scan is performed. If it is a white scan, search sequentially until a black scan appears on the right side. If the number 5 is not written, everything up to the right extraction base address 9-2 will be scanned white, and in this case, the character frame 10 will be treated as a blank. In the same figure, the number 5 is clearly marked and its character frame 1
Since it is written within 0, when the vertical scan is searched to the right as described above from the left extraction base point address 9-1, the first black scan appears immediately. In this case, position 11 of the first black scan is set as the left extraction position (address). Also, unlike this figure, the left extraction base address 9-
When the starting scan on 1 becomes a black scan, a vertical scan search is performed alternately on the left and right for the same number of ranges (usually ten or more bits) on the left and right, starting from the starting vertical scan position. If a white scan first occurs on the left vertical scan during this search, use the vertical scan to the right of it as the left cutting position, and if a white scan occurs on the right vertical scan, drive to the right until a black scan is found. The black scan discovery position is set as the left cutting position. If the above-mentioned white scan occurs within the left character frame, the characters written in the character frame are intruding into the left character frame, and if the white scan occurs within the character frame, the characters written in the left character frame are intruded into the left character frame. This is a case where the characters are invading the character frame and there is no overlap between both characters. If a white scan cannot be found within the entire search range, character segmentation is forcibly performed using the left extraction base address position 9-1, which is the left character frame base position, as the left extraction position. This occurs when a character written in the character frame 10 and a character written in the character frame 10 to the left of the character frame 10 mutually penetrate deeply into the other character frame, causing an overlap, or when the characters are connected. In this way, the character extraction position is determined by sequentially searching for the right side, upper side, and lower side using the same procedure, and the character extraction process is completed, and the next pattern recognition process is started. A case will be described in which, when the characters described in the above-mentioned adjacent character frames 10 deeply invade each other's character frames, the characters are forcibly cut out to the left or right from the left or right character frame base position. FIG. 5 shows the case of so-called overflow characters in which the numbers 5 and 3 are written as characters 2 on the form 1 so as to mutually intrude into the other character frame 10. In other words, the horizontal line at the top of the number 5 is the number 3 on the right.
invades the character frame 10 as an overhang 12,
Further, the lower and rear portions of the number 3 form a protruding portion 12 that intrude into the character frame 10 of the number 5 on the left. The right extraction base point address of number 5 in the same figure (number 3
FIG. 6 shows a situation in which the image is converted into a video on the image memory with 9-2 (same as the left extraction base point address) as the center. In FIG. 6, it is assumed that the right search 13 and the left search 14 are performed alternately 10 times as the extraction search from the right extraction base point address 9-2 of number 5 for convenience of illustration. (Actually, as mentioned above, this is usually done more than ten times at a time). In the same figure, the upper horizontal line of the number 5 becomes the protruding part 12 and protrudes from the area of the left search 14, and the protruding part 12 of the number 3 also intrudes into the left character frame 10, so in any case, Ten
In the right search, no white scan occurs in the vertical scan. The same applies to the right search area 13. Therefore, in the case of FIG. 6, the right extraction position 15 for the character frame 10 in which the number 5 is written is the same as the right extraction base address 9-2, and the character box 10 in which the number 5 is written is entered. A protruding portion 12 with a number 3 is attached thereto. However, if we perform forced extraction as described above,
When distinguishing and recognizing extracted characters, there is a possibility that misrecognition may often occur, so when forced extraction is performed, reject is output as an answer, and in this example, both 5 and 3 are rejected, reducing the operational efficiency of OCR. let The present invention aims to propose a character extraction method that identifies the above-mentioned protruding characters, prevents misrecognition, and improves operational efficiency by reducing the number of rejects. Image information including multiple characters is scanned in the vicinity of the boundary between character frames defined by character position information given in advance so as to cross the characters, and as a result of this scanning, a black scanning group containing black pixels and other When there is a white scan that does not include black pixels between a group of black scans, the scan position where the white scan is detected is regarded as the boundary between characters, and an image of a rectangular area containing information for each character from the above image information is created. In the character extraction method for extracting information,
A rectangular area having a certain size centered on the boundary between the character frames is set, and the continuity of the black area on each scanning line in the rectangular area with the adjacent scanning line is checked, and a rectangular area is set on the starting end side of the rectangular area. a black area continuity inspection means for classifying the black area into a continuous black area and a black area connected to an end side; and a black area image in which the black area connected to the black area on the side opposite to the character side to be processed is removed from the rectangular area, and the black area is removed. a removing/holding means for holding the image in a predetermined holding location, and an increased image in which image information of a rectangular area from which the black area connected to the black area on the opposite side has been removed is connected to the image information within the character frame to be processed. Equipped with an increased image information creation means for creating information;
When a white scan is not detected in all the scans near the boundary between character frames, a test is performed by the black area continuity inspection means, and if the black area cannot be separated as a result, the black area is detected in the vicinity of the boundary between character frames. If the black area image can be cut out and separated, the image information is cut out from the image information created by the increased image information creation means, and the black area image held by the removal/retention means is used by connecting it to the next character image to be processed. This can be achieved using a character extraction method featuring the following. In other words, the numbers 5 and 3 in Figure 5 mentioned above
Taking the overflowing state as an example, the right cutting position of the number 5 is searched to the right by more than ten bits, but since the numbers 5 and 3 overlap each other, no white scan occurs. However, since the numbers 5 and 3 are separated from each other, the protruding part of the number 5, which is a continuous segment from the cutout area, that is, the area of the number 5, is not continuous from the cutout area, If the protruding part of the number 3 that is continuous from the adjacent cutout area, that is, the area of the number 3, is determined to be discontinuous, and the protruding part of the number 3 is removed from the cutout area of the number 5, the number 5 can be recognized correctly. , there is no need to reject the form. Next, when cutting out the number 3, the previously removed protruding portion of the number 3 is returned and the cutting is performed. An embodiment of the present invention will be described below with reference to the drawings. FIG. 7 shows a video 16 when character frame 10 of numeral 5 is to be cut out and its right extraction position is determined with numerals 5 and 3 protruding from each other as shown in FIG. 6.
This is an edge pattern with edge addresses added to the segment diagram. What is an edge pattern? Taking Figure 7 as an example, if a vertical scan is performed from the bottom to the top, if it changes from white to black, the address of the changed black bit is one edge (in this case, it is called a black edge). If the edge changes from black to white, the address of the changed white bit is used as the address of the other edge (in this case, it is called the white edge), and these addresses are displayed in this figure. The drawing is centered on the right extraction base point address 9-2 of the frame 10. In the figure, a ₁ a ₂ is the address of a pair of black edge and white edge, and if the alpha suffix of a to t shown in the figure is an odd number, it is a black edge, and if it is an even number, it is a white edge. It is. In addition, if there are multiple character areas in the same vertical scan, for example, in the leftmost vertical scan, ie, a scan, each character area (in this case, a ₁ - a ₂ and a ₃ - a ₄₎ 2 sets) are shown. In the two vertical scans of e and s in the figure, there are four character areas, which are the largest portion in this example. From the edge pattern in Figure 7, left and right character frames 10
The number 5 among the characters (5 and 3 in this case) belonging to
9 to 14, the details of the separation/cutting process for the protruding portion on the right side will be explained with reference to the flowchart of FIG. 8. In FIG. 8, when the right extraction process for the protruding portion is started, first search areas of a predetermined number of bits (usually + several bits) are provided on the left and right sides from the right position of the character frame, and vertical scanning is performed from the right side to the left side. If there is a white scan, the position where the first black scan is found after the white scan is set as the right extraction position. If there is no white scan for the entire search area, use the continuous edge detection method using an edge table described below to determine whether the videos within the search area are separated, and if they are not separated, As before, the right position of the character frame is forcibly set as the right cutting position, and if it is separated, the video of the right side character is stored in a specially provided storage area (this is called SAVE in the present invention). This process will be referred to as SAVE from now on), leaving only the characters on the left side in the extraction area. Next, a program switch is provided to remember that the SAVE operation described above has occurred.
Set SW to SAVE = 1. Incidentally, in both the case where the above-mentioned white scan exists and in the case of forced extraction, the SAVE SW is set to 0 before starting the feature extraction process. Returning to the above description, after operating the program switch, the right side of the cut-out character is cut out again, and the entire character including the protruding part is cut out. Next, after performing feature extraction processing, SAVE SW
Check whether = 1, and if it is 1, the character on the right side
Return the saved video to the video memory area where the cutting operation is being performed, and start cutting out the next character. Determination of continuity or separation between segments of adjacent vertical scans of video information that has previously changed will be described below. A top and bottom edge table 17 shown in FIG. 9 is created from the edge pattern shown in FIG. Also, upper and lower edge table final address table 1
8 contains the final pointer ap for each vertical scan.
~tp is stored. From the upper and lower edge table 17 shown in FIG.
We compared the method of detecting continuous edges by comparing the continuity discriminant diagram in Figure 11, which defines continuity or discontinuity depending on the positional relationship between adjacent segments, and the continuity discriminant formula described below, and also compared the edge pattern in Figure 7. Refer to and explain. Let j be the lower edge address and k be the upper edge address of the segment for which continuity determination is to be performed in the segment obtained in the m-1st scan, and continuity determination for the above j and K in the current segment obtained in the m-th scan. Let h be the lower edge address and i be the upper edge address of the one that performs the following. There are eight types of continuity discriminants as mentioned above, and the numbers of the corresponding continuity discriminant diagrams are written on the right side. However, the relationships h<i and j<k always hold true.

【table】

[Table] Condition (1) is when the left segment is above and the right segment is below, and there is no continuity between them, and the opposite is the condition (8). There are 10 types of continuity conditions (2) to (7) based on the size and connection state of the left and right segments. These discriminants will now be explained by applying them to FIGS. 7 and 9. First, when we compare a ₁ a ₂ and b ₁ b ₂ in Fig. 7, (3) holds and the form shown in Fig. 11 3-4 is continuous. Next, when a ₁ a ₂ and b ₃ b ₄ are compared, (8) holds true and the state shown in FIG. 11 is discontinuous. Next, a ₃ a ₄ and b ₁ b ₂ are discontinuous at (1), and a ₃ a ₄ and b ₃ b ₄ are (4)
is established, and the state shown in FIG. 11 4-2 continues. By comparing the edge addresses of adjacent segments in this way, it is found that d ₁ d ₂ and e ₁ e ₂ are discontinuous, and d ₁ d ₂ and e ₁ e ₂ are discontinuous, From now on, the edge series is continuous to a ₁ a ₂ and a ₃ a ₄ , which are continuous from the left region, and the edge series is continuous to e ₁ e ₂ , q ₃ q ₄ , and s ₅ s ₆ , which are discontinuous from the left region. It can be determined that the series is discontinuous. When the result is stored in the continuous edge registration table 21, it becomes as shown in FIG. 10, and the area is monitored by the final pointers ap to tp corresponding to each scan of the table final address table 18. The edge addresses that were not transferred from Fig. 9 to Fig. 10 are stored separately by the SAVE processing described above, and are retrieved and used again when cutting out the number 3 in this example, but that processing will be done later. 13th
This will be explained when explaining the circuit block diagram in the figure. Next, if we look at the part e ₃ e ₄ −f ₃ f ₄ −e ₅ e ₆ in Figure 7,
For a segment of f ₃ f ₄ , an edge series continuous to e ₃ e ₄ and an edge series continuous to e ₅ e ₆ are continuous. The same situation occurs in e ₅ e ₆ −d ₃ d ₄ −e ₇ e ₈ . This shape is called a "curve", a discriminant condition diagram is shown in FIG. 12, and a discriminant is described below. The symbols h, i, j, and k are the same as in the 11th, and the lower and upper edges above k are n and o, respectively, and the lower and upper edges above i are l and m, respectively.
In addition, the numbers for the discriminants are given following the above. First, in the case of a left turn, (9) when h<k<n<i, 9 in Fig. 12.
This is the case of the related positions shown in -1 to 9-5. Next, in the case of a right turn, since it is symmetrical with a left turn, (10) j<i<l<k, and in the case of the related positions shown in 10-1 to 10-5 in Fig. 12. be. Through this "curvature" determination, it is determined that all the continuous edges in FIG. 10 are continuous. The determination of bending in the vertical direction is also performed in the same manner, so the explanation will be omitted. Although right-hand cutting has been described above, separation cutting is also performed for the left, upper, and lower protruding portions using similar processing procedures and operations. The circuit operation of the separation process will be explained with reference to a circuit block diagram of an embodiment of the present invention shown in FIG. In the same figure, the circuit operation from the beginning will be explained although it overlaps with the above explanation to some extent. First, an extraction circuit start signal is issued from the conventional device, the extraction circuit 22 is activated, and at the same time, the video storage 2
The video data 〓〓 is input from 9 and the cutting operation is started. Continuing the explanation up to now, the case of right extraction will be explained. In the extraction circuit 22, a vertical scan is started to the left from the right extraction base point address 9-2 in FIG. 7, as described above. In this example, since the characters 5 and 3 protrude as described above, black scans are performed from the first vertical scan. In this case, according to the sequence, the search area of more than ten bits on each side of the right extraction base point address 9-2 is scanned alternately left and right with the base address 9-2 as the starting point, and the presence or absence of a white scan is tested. However, as mentioned earlier, in this explanatory diagram, the range is 10 bits each on the left and right sides. When a white scan is not found even after searching this range, the extraction circuit 22 inputs the signal to the edge table creation circuit 23 as a signal. The edge table creation circuit 23 is the cutout circuit 22
The cutting area data indicating the edge table creation area is received as a signal, and a write signal 〓〓 is sent to the edge table storage circuit 24 to store the upper and lower edge addresses as edge table data. As mentioned earlier, the extraction circuit 22 scans the search area twice, in the case of right extraction, rightward and leftward, and the results are collated in the edge table creation circuit 23 to find a match. If they do not match, an edge table creation circuit end signal is sent to the continuous segment determination circuit 25 to activate it, and if they do not match, an extraction failure status signal is sent to the extraction circuit 22 and the abnormality processing sequence circuit. The cutout circuit 22 that received the above signal is SAVE.
A SW=0 signal is sent to the AND gate 32, and a start signal is sent to the recognition circuit 31 along with a feature extraction circuit end signal of the feature extraction circuit 30 to terminate the character processing. The continuous segment determination circuit 25 that received the above signal sends a read signal 〓〓 to the edge table storage circuit 24, receives the edge table data, and receives the above-mentioned continuity discriminants (1) to 8 and the bend discriminant.
Using (9) and (10), write the edge address of the continuous segment into the continuous edge registration table of the continuous segment storage circuit 26, write the number of bends into the bend register of the address of the bend segment, and write the number of bends into the bend counter. Store it as data. If a segment continuous from the left region and a segment continuous from the right region cannot be separated in the entire search area, a separation failure status signal is transmitted to the extraction circuit 22 and the abnormality processing sequence circuit, and the above-mentioned extraction failure status signal is transmitted. Perform the same processing as when If there is a separate segment, the SAVE data is transferred to the next continuous segment and the video data comparison circuit 2.
7, it is stored separately. When the processing of the continuous segment judgment circuit 25 is completed, a continuous segment judgment circuit end signal is sent to the continuous segment and video data comparison circuit 27 to start its circuit operation, and a read signal is sent to the continuous segment storage circuit 26 to Receives segment data, plays it into a video pattern, and simultaneously sends read/write signals to video storage 2.
9 and the address data sent from the extraction circuit 22, the video data is extracted and the two video patterns are compared. If both video patterns match, the remaining video patterns in the video data are treated as discontinuous segments, and a write signal 〓〓 is sent to the discontinuous segment storage circuit 28 to store the discontinuous segment data, while the video storage 29 The video storage erase data is sent to remove the portion of the discontinuous segment. Summarizing the operations to remove this discontinuous part,
I mentioned earlier that it is called SAVE, but this SAVE
When this is done, a signal of SAVESW=1 is generated and sent to the cutout circuit 22 and the AND gate 34. When the operation described above is completed, the cutout circuit 2
A cutout circuit end signal is transmitted from the cutout circuit 22, and the feature extraction circuit 30 is activated upon receiving left and right, upper and lower cutout address data and a cutout circuit start signal from the cutout circuit 22. When the operation of the feature extraction circuit 30 is completed, the feature extraction circuit end signal is sent to the AND gates 32, 33,
Send to 34. First, the AND gate 34 is connected to the above signal.
Upon receiving the SAVESW=1 signal, a load start signal of the discontinuous segment to the video storage is sent, which is a command to transfer the video data of the discontinuous segment previously stored in the discontinuous segment storage circuit 28 to the video storage 29. . Upon receiving this signal, the discontinuous segment storage circuit 28 transfers the discontinuous segment data to the video storage 29, and adds the SAVE data to the video pattern of the next character to be extracted. The AND gate 33 receives the load end signal of the discontinuous segment from the discontinuous segment storage circuit 28 to the video storage, and sends an activation signal to the recognition circuit 31 via the OR gate 35, and the feature extraction circuit 30 Receive feature data 〓〓. The AND gate 32 receiving the signal activates the recognition circuit 31 via the OR gate 35 when the signal SAVESW=0 is being transmitted from the extraction circuit 22. As described above, according to the present invention, characters that overlap each other, which could not be separated in the past, can be separated and recognized, thereby reducing rejects in OCR reading operations and improving the success rate of character recognition. , has a great effect on improving OCR operating efficiency and reliability.

[Brief explanation of drawings]

Figure 1 is an explanatory diagram of the configuration of the character recognition device, Figure 2 is a diagram of the image memory configuration, Figure 3 is a diagram of the positional relationship between images on a form and format information, and Figure 4 is a diagram of video and cutout information on the image memory. , FIG. 5 is a diagram of the positional relationship of protruding characters on a form, FIG. 6 is a video diagram, and FIG. 7 is a segment diagram of an embodiment of the present invention.
FIG. 8 is the same right cutting flowchart, FIG. 9 is the same upper and lower edge table, FIG. 10 is the same continuous edge registration table, FIG. 11 is the same continuous discrimination condition diagram, and FIG. 12 is the same curved edge discrimination condition diagram. FIG. 13 is a circuit block diagram as well. In the figure, 1 is a form, 2 is a character, 3 is a lens system, 4 is a sensor, 5 is a preprocessing circuit, 6 is an image memory, 7 is a processor, 8 is a 1-pixel memory, 9-1
9 to 9-4 are the extraction base address, 10 is the character frame, 11 is the left extraction position, 12 is the protruding part, 1
3 is the left search, 14 is the right search, 15 is the right cutting position, 16 is the video, 17 is the upper and lower edge table,
18 is a table final address table, 19 and 20 are final pointers, 21 is a continuous edge registration table, 22 is an extraction circuit, 23 is an edge table creation circuit, 24 is an edge table storage circuit, 25
26 is a continuous segment storage circuit; 27 is a continuous segment and video data comparison circuit; 28 is a discontinuous segment storage circuit;
29 is a video storage, 30 is a feature extraction circuit,
31 is a recognition circuit, 32 to 34 are AND gates,
35 is the OR gate, is the cutout circuit start signal, is the SAVESW=1 signal, is the edge information signal, is the SAVESW=0 signal, is the cutout circuit end signal, is the cutout failure status signal, is the edge table data, is the edge table creation circuit end signal, is edge table data,
and 〓〓 is a read signal, 〓〓 is a separation failure status signal, and 〓 is a continuous segment data, is a continuous segment judgment circuit end signal, and 〓〓 is a video data, is a read/write signal, is a video storage erase data, and is an invalid data. Continuous segment data, is the load start signal of the discontinuous segment to the video storage, is the load end signal of the discontinuous segment to the video storage, is the feature extraction circuit end signal, 〓〓,, 〓〓 is the write signal, 〓 〓 and 〓〓 are address data, and 〓〓 is characteristic data.

Claims (1)

[Claims] 1. Scanning image information including a plurality of characters stored in an image memory near the boundary between character frames determined by character position information given in advance so as to traverse the characters. However, as a result of this scanning, if there is a white scan that does not include black pixels between a black scan group that includes black pixels and another black scan group, the scan position where the white scan is detected is regarded as the boundary between characters. In a character extraction method that extracts image information of a rectangular area containing information for each character from the above image information, a rectangular area having a constant size is set centered on the boundary between the character frames, and a black area continuity testing means 1 for testing the continuity of black areas on each scanning line with adjacent scanning lines and classifying the rectangular area into a black area connected to a starting end side and a black area connected to a terminal end side of the rectangular area; a removing/holding means 2 for removing a black area connected to a black area on the opposite side of the character side to be processed from the area and retaining the removed black area image in a predetermined holding location; and a character frame to be processed. an additional image information creating means 3 for creating increased image information in which image information of a rectangular area from which a black area connected to the black area on the opposite side has been removed is connected to the inner image information; When white scanning is not detected in , a test is performed by the black area continuity testing means 1, and if the black area cannot be separated as a result, it is cut out along the character frame boundary, and if separation is possible, Cutting out the image information created by the increased image information creation means 3,
A character cutting method characterized in that the black area image held by the removing/holding means 2 is connected to and used for the next character image to be processed.