JPH01166276A - Image recognizing method - Google Patents

Abstract

PURPOSE:To improve the efficiency of matching with no useless candidate remaining in the result of recognition and to give no load on a sort module for result of recognition by checking the difference of features between an input image and a dictionary for each shape forming characters, etc., and then checking the overall difference of features. CONSTITUTION:The characters, etc., of an original are optically read and stored in an original image area 31 of a RAM 4 as a binary image where white picture elements are shown in '0' with the black picture elements shown in '1' respectively. The white or black picture elements of the contour part are tracked clockwise or counterclockwise by means of a CPU 2 for the original image stored in the area 41. Thus the direction code of the contour part is extracted and added to a 2-dimensional contour image and this contour image is stored in a contour image area 42. At the same time, a 1-dimensional direction code train is stored in a direction code storing area 43. Then the dividing lines which cross vertically and horizontally the contour image stored in the area 42 are decided. The image contour part is divided with said dividing lines for extraction of a type. Then the contour area of each form is equally divided into (n) pieces and a histogram is produced from the direction code of the area 43 and compared with a dictionary memory 5.

Description

[Detailed description of the invention] 〔Technical field〕 .

The present invention relates to an image recognition method in an optical character reading device or the like.

[Prior art]

Conventionally, as a method for recognizing input images such as characters in optical character reading devices, etc., the input image is divided into regions two-dimensionally, a histogram of direction codes is generated in each region, and this is used as the feature quantity of the input image. A known method is to match similar feature amounts registered in advance in a dictionary and use the pattern with the minimum distance value as the recognition result. However, since this method performs two-dimensional region segmentation on the input image, it is sensitive to changes in character shape, and is not effective, especially when dealing with multiple fonts, as it requires a dictionary for each font.

[Objective] An object of the present invention is to provide a simple recognition method that is effective for character shape changes in optical character reading devices and the like.

〔composition〕

The outline of the image is divided by dividing lines set according to the rules on the top and bottom or left and right of the image, and the shape (type)
is extracted, and the region of each shape is divided into n equal parts (n is an integer of 1 or more) to generate a histogram of direction codes. After that, first, the created histogram is compared with the histogram registered in the dictionary in advance for each shape of the outline of one image. When this is completed for all shapes,
Next, the total value of the zero addition value of the distance calculation (the distance addition value of the part of the input image or dictionary where one direction code does not exist) or the zero count value (the distance addition value of the part of the input image or dictionary where one direction code does not exist) is calculated. The total value of the number of appearances) is compared with an appropriate threshold value, and it is checked whether to leave it as a candidate for the recognition result.

In the present invention, the feature difference between the input image and the dictionary is checked for each shape unit that constitutes a character, etc., and the overall feature difference is also checked, so that efficient matching can be performed.
No useless candidates remain in the recognition results.

〔Example〕

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of an example of a hardware configuration for realizing the present invention, and FIG. 2 is a block diagram of the central processing unit i! 2 shows a schematic processing flow of step 2.

The scanner 1 optically reads characters, etc. on a document and converts them into a binary image (for example, a white pixel is ``0'', a black pixel is ``1'').
) in the original image area 41 of the data memory (RAM) 4. First, the central processing unit 2 extracts the direction code of the contour by tracing the white pixels or black pixels of the contour of the original image in the original image area 41 in a counterclockwise or clockwise direction, and then extracts the direction code of the contour. The two-dimensional contour image obtained is stored in the contour image area 42, and the -dimensional direction code string itself is stored in the direction code storage area 43 (step 101). Next, the central processing unit i! Step 2 determines, based on a predetermined rule, dividing lines that intersect the contour image of the contour image area 42 to which the direction code has been added at predetermined positions on the top and bottom or left and right of the image (step 1o2). Next, the central processing unit 2 divides the outline of the image using the dividing line and extracts the shape (type) of the image (step 103). The contour area is divided into n equal parts (n is generally an integer of 1 or more), a direction code histogram is generated for each divided area using the direction code string in the direction code storage area 43, and is stored in the histogram storage area 44 (step 1o).
4). Next, the central processing unit 2 refers to the dictionary memory 5 using the generated histogram to obtain a recognition result (step 1
05).

A program memory (ROM) 3 includes the central processing unit 2
Stores the programs necessary for processing. Each process will be explained in detail below.

In the image contour section, a direction code (this is called a movement direction code) indicating the direction of movement to the next tracking point is assigned to the white pixel or black pixel of each tracking point. shall be described as follows.

FIG. 3 shows the relationship between the moving direction of the tracking point and the direction code. In other words, if the position (coordinates) of the pixel of the point of interest is regarded as the position vector IP, movement to the next tracking point is determined by the four unit vectors (movement direction vectors) shown in Fig. 3(a).
The coordinates of the movement destination (next tracking point) are expressed as IP+ΔIP, which is the sum of the unit vector ΔIP and the position vector IP of the point of interest. The moving direction codes shown in FIG. 3(b) are assigned to these four unit vectors, and the image outline is described by the direction codes. This is detailed in Japanese Patent Application No. 62-20289.

Min1lu1wa[1 Here, horizontal dividing lines are set at the top and bottom of the image. This uses a horizontal projection for each movement direction code, #1. #3 Movement direction code (vertical direction code)
Find the position where the projection of matches from both the upper and lower ends,
This is achieved by drawing a line (region dividing line) laterally from the position.

A concrete example of this is shown in FIG. Figure 4(a) shows the letter “S”
'' shows a contour image in which the white pixels of the contour part are traced counterclockwise and a movement direction code is assigned.

If we calculate the horizontal projection for each movement direction code #1 to #4, we will get something like Figure 4(b).The horizontal projection in Figure 4(b) was calculated as follows. It is. For example, in the case of one line, there is one code #1, one code #2.
#3 is none, code #4 is 5, and the projection is code #
1. #3. In order of $2゜#4, 1. It becomes O, 0, 5. The same applies to the projections of lines 2 to 22. Fourth
The upper and lower area dividing lines from which CUT-Yl and CUT-Y2 in the figure were determined are shown, respectively.

FIG. 5 is a flowchart of the area dividing line determination algorithm, in which figure (a) is the algorithm for determining the upper area dividing line CIT-Y1, and figure (b) is the lower area dividing line CUT-Y.
2 is shown. Since the algorithm for determining the upper area dividing line and the algorithm for determining the lower area dividing line are symmetrical, only the determination of the upper area dividing line CUT-Yl will be described in Section 2.

The lower region dividing line CUT-Y2 can be easily determined from this.

In FIG. 5(a), a process 201 is performed using the area dividing line CI.
This determines the end point of the range for searching for TY1. This corresponds to detecting the thickness of characters. In the case of the example in FIG. 4(a), the point at which the horizontal projection of movement direction code #2 is no longer zero is line 4, which is defined as (LIM
+1), and the previous position (line 3) is set as the end point LIMI of the range in which the area dividing line CUT-Y1 is searched.

Processing 202 is to determine the starting point of the range for searching for the area dividing line CUT-Y1. This means finding a position where the horizontal projections of movement direction codes #1 and #3 are both non-zero, and corresponds to finding the upper end of the character area. In the example of FIG. 4(a), line 2 is the starting point.

Processes 203 and 204 are conditions for determining area dividing lines. That is, when searching the search range from top to bottom and detecting the position where the horizontal projections of movement direction codes #1 and #3 first match, the horizontal straight line at that position is defined as the area dividing line CUT-Y1. shall be.

In the example of FIG. 4(a), line 3 is the relevant position, and C
UT-Y 1 is drawn. If a position that satisfies the conditions is not found, move the end point LIM1 of the search range to the area dividing line C.
The position is UT-Y1.

The above-mentioned conditions for determining the area dividing line have the effect of noise processing. Figure 6 shows a concrete example of this, and Figure 1 (
Figure a) shows the case where there is no noise, and Figures (b) and (Q) show the case where there is noise. That is, there is no noise on the area dividing line, and the most stable position of the image contour can always be set as the area dividing line.

This method divides the image outline using area dividing lines to capture the rough shape of the image.

Figure 7(a) is an example of dividing the outline of the character "0" using upper and lower dividing lines, and shows the shape (type) of the character "0".
is expressed as '1, 2, 4.5'.

Similarly, FIG. 7(b) is an example of the character "I(", and the character rHJ
This shows that the shape is expressed as "1, 2, 3° 2.4, 5, 6.5". 2N, the numbers attached to each divided area have the meanings shown in FIG. 7(c).

Histogram Generation J Process This is done by dividing the outline of an image using vertical or horizontal dividing lines, dividing each shape (type) area into n equal parts, and then creating a direction code for each of the n equal parts. This generates a histogram of . Figure 8 shows a specific example when n = 3. Figure 8 (a) shows each shape (type) 1, 2゜4 extracted by dividing the outline of the character "0" by upper and lower dividing lines. Figure 8 (b) shows the case where .5 is divided into three equal parts.
) is a histogram of direction codes for each region for each shape 1, 2, and 4.5. 2, codes 1 to 4 correspond to movement direction codes #1 to #4 explained in FIG. FIG. 8(c) is a dictionary histogram for the input histogram of FIG. 8(b).

μmJ! LJ! L-Star This is a process that recognizes the input image by comparing a histogram generated for each shape of the outline of the input image with a dictionary histogram that has been generated in advance in the same way and registered in the dictionary memory. be. FIG. 9 shows the processing flow.

In process 301, among the histograms obtained for each shape of the outline of the input image, first, for example, a type 1 histogram is compared with similar histograms registered in the dictionary, and distance values and zero addition values are calculated. Or find the zero count value. In the example of Figure 8, for type 1, the distance value is 15+51+14-31+15-51+1l-2
1+11-11++5-51+14-31=3, zero addition value is 10-01+l0-01+l0-01+10-0
1+l0-01+l0-01=O, similarly determined as zero count value=0.

In process 302, the distance value and zero addition value or zero count value of the type determined in process 301 are compared with an appropriate predetermined threshold, and if the determined value is greater than the threshold, subsequent processes are performed. This is omitted and the process immediately proceeds to the next dictionary histogram comparison, and only when the obtained value is equal to but smaller than the threshold value, the process proceeds to the next process 303.

In process 303, the values for each type found so far are added for each of the distance value and the zero addition value or zero count value. For the example in Figure 8, the first type 1
At the stage where the histograms have been compared, processing 303
The addition result (total value) is the same as the calculation result in process 301.

That is, the total value of distance values=3, the total value of zero addition values=0, and the total value of zero count values=0.

In process 304, it is determined whether the processes have been completed for all types, and if not, the process returns to process 301.

In process 301, next, for example, a distance value and a zero addition value or a zero count value are determined for type 2. In the case of Figure 8, for type 2, the distance value is +2-31++2-2
1+11-01+14-41+11-21+12-21
+12-21=3, zero addition value is 10-01+l0-0
1+11-01+1o-01+l0-Of-110-0
1=1, zero count value=1. If it is determined in process 302 that these values are equal to or smaller than the threshold value, the process proceeds to process 303, where the total value of distance values is 3+3=6, and the total value of zero addition values is Of1=1. The total value of the zero count values is updated to Of1=1.

If the distance value and zero addition value or zero count value of each type unit is equal to or smaller than a predetermined threshold value, process 30
1,303 is repeated. Figure 10 shows the results when matching types 1, 2, 4, and 5 in the order of type 1, 2, 4, and 5 for the example in Figure 8.
The distance values, zero addition values, and zero count values for each type and total obtained in processes 301 and 303 are summarized in a table.

If the result of the process 304 is YES, it means that the item was not rejected in the verification on a per-type basis. In this case, in process 305.

The total zero addition value or total zero count value is compared with a predetermined appropriate threshold value. And if,
If the total zero addition value or total zero count value is greater than the threshold value, it is not retained as a candidate, and at that point the process moves to the next dictionary histogram comparison, and if it is equal to or smaller than the threshold value, it is determined as a final candidate. For example, in the case of the example shown in FIG. 8, the total zero addition value for types 1, 2, and 4.5 is 4, and the total zero count value is 3, as shown in FIG. In process 305, this total zero addition value = 4
Alternatively, the total zero count value = 3 is compared with an appropriate threshold value to decide whether to make it a final candidate.

This process 305 is carried out in such a way that the distance value and zero addition value (or zero count value) for each type of outline of the input image do not exceed the threshold determined for each type, and can be viewed for each type. However, when all types are accumulated, it becomes a large value, and it is effective when it is necessary to reject the whole type.

FIG. 11 shows the relationship between the value of each type unit, the total value, and the threshold value in the case of a zero-added value.

In this example, in the matching for each type, each zero addition value is smaller than the threshold and is not subject to rejection, but
When viewed as a whole, it becomes a large value like 12. Therefore, by setting the total threshold value TH to an appropriate value (for example, 10), rejection can be achieved.

〔effect〕

As explained above, according to the present invention, the feature difference between the input image and the dictionary is checked for each shape unit constituting a character, etc., and the overall feature difference is also checked. Matching can be performed and no useless candidates are left in the recognition results, which has the effect of not placing a burden on the recognition result sorting module.

[Brief explanation of the drawing]

Figure 1 is a block diagram of the hardware configuration of an embodiment of the present invention, Figure 2 is a diagram showing an example of a processing flow according to the present invention, Figure 3 is a diagram explaining movement direction codes, and Figure 4 is a dividing line. A diagram showing a specific example of determination, FIG. 5 is a flowchart of the dividing line determination algorithm, FIG. 6 is a diagram showing a specific example of determining the dividing line when noise is included, and FIG. 7 is a specific example of shape extraction. 8 is a diagram showing a specific example of histogram generation,
FIG. 9 is a diagram showing a detailed flow of the recognition process, and FIGS. 10 and 11 are diagrams showing specific examples of the recognition process. 1...Scanner, 2...Central processing unit, 3...
Program memory, 31... Contour tracking program, 32... Area dividing line determination program. 33... Shape extraction program, 34... Histogram generation program. 35... Recognition program, 4... Data memory. 41...Original image area, 42...Contour image area, 4
3...Direction code storage area. 44...Histogram storage area. Fig. 3 (b) (g) Fig. 5 Fig. 7 (C 7 7' + 234567 Fig. 8 (α) ri 1 a 5 q)
(C) Handwritten Shimata Door" Ram Figure 9 LtNio Sort △ Figure 10 Figure 11

Claims (3)

[Claims] (1) Trace the contour of the image to extract the direction code of the image contour, set dividing lines that intersect the image according to a predetermined rule at the top and bottom or left and right positions of the image, and use the set dividing lines to divide the image. Divide the outline and extract the shape, divide the region of each shape into n equal parts, generate a histogram for each region, and recognize the image by comparing it with the histogram registered in the dictionary in advance. In the image recognition method, in the recognition process, candidates are first obtained by comparing the generated histogram with the histogram registered in the dictionary for each shape unit of the outline of the image, and then candidates are obtained for each shape unit. An image recognition method characterized in that a value obtained by integrating matching results is compared with a predetermined threshold value to ultimately decide whether to keep it as a candidate or not. (2) For each shape, add the distance between the generated histogram and the part of the histogram registered in the dictionary where one direction code does not exist, and then integrate that value for all shapes and use it as a predetermined threshold. 2. The image recognition method according to claim 1, wherein the comparison is made to decide whether or not to keep the image as a final candidate. (3) For each shape, count the number of occurrences of the part where one direction code does not exist in the generated histogram and the histogram registered in the dictionary, and set the value obtained by integrating this value for all shapes as a predetermined threshold. compared to
2. The image recognition method according to claim 1, further comprising the step of deciding whether or not to leave the image as a final candidate.