JPS6057476A - Key word detecting system - Google Patents

Abstract

PURPOSE:To segment effectively frameless characters by calculating the similarity between a pattern segmented by a window to a character area and a dictionary pattern of a key word to obtain a histogram of similarity and therefore detecting the key word position from the peak position of the histogram. CONSTITUTION:The information written on a mail is supplied to a line detecting/segmenting part 2 after photoelectric conversion through an input part 1 of a key word detection system. Thus the line position of a character train containing a black area. This data is stored to a line memory 3, and each picture element data undergoes D/A conversion through a sampling part 4. The inter-line data adjacent to each other are blurred and converted into (r) units of picture element data. While a key word pattern registered previously to a dictionary pattern memory 5 is recorded. Then the dictionary pattern and the blurred picture element data are applied to a similarity calculating part 7. The window position is linearly shifted to a character train to calculate a histogram of similarity. This histogram is stored to a histogram memory 8. Then the key word position is detected by a peak detecting part 9.

Description

[Detailed description of the invention]

[Technical Field of the Invention] The present invention relates to a keyword detection method that can effectively detect the position of a specific keyword in a character string written on, for example, a frameless sheet of paper. [Technical background of the invention and its problems] Character recognition processing generally involves detecting a character string from a relatively wide observation area (paper) and cutting out each character that makes up this character string. After that, the characteristic pattern of the cut characters is extracted. Specifically, for example, a character string indicating an addressee written on a letter or postal item is detected, and each character forming the addressee is examined to perform character recognition. In order to perform such character verification processing, conventional methods have been used to perform character recognition processing independently of character recognition processing, such as pitch information for character strings, the existence range of white areas (background parts), or artificial detection marks. The inspection process is being carried out using Hiromu as a clue. However, when performing inspection processing in this manner, it is necessary to specify a measurement format consisting of character strings, and for this reason conventionally, a character entry frame is exclusively provided. However, defining all character positions by character entry frames lacks flexibility and lacks versatility. Therefore, it is strongly desired to effectively examine and recognize each character of a character string written on a so-called blank sheet of paper without a frame. [Object of the invention] The present invention was made in consideration of the above-mentioned problem #1,
The purpose is to provide a keyword detection method that makes it possible to effectively check character strings written on unframed paper, for example. [Summary of the Invention] The present invention is based on the position of a specific keyword in a character string written on a frameless sheet of paper, such as "prefecture,""city," or "town" in an addressee.
It is possible to accurately detect a keyword such as H1'5c and use this as a clue to inspect each character of the character string.In particular, while moving the window position relative to the character area while detecting the keyword, It calculates the degree of similarity between /9 turns cut out from the text area using the above window and the dictionary pattern of the keyword, creates a histogram of the similarity to the keyword, and detects the keyword position in the text area from the peak position of this histogram. [Effects of the Invention] Thus, according to the present invention, as the position showing the peak of the histogram of the similarity between the knee turn cut out by the window that is shifted little by little and the row standard pattern of the keyword, It is possible to easily detect the position of a keyword in a character string written on a sheet of paper without a frame with high accuracy. Therefore, it is possible to easily check each character using this keyword position as a clue, and as a result, the efficiency of character recognition processing can be improved. In addition, it is possible to effectively detect keywords in character strings freely written on blank paper without being limited by the text entry frame, so it can be applied to assist in reading addresses on letters and mail, for example. It has great practical effects, such as being able to do things. [Embodiments of the Invention] Hereinafter, with reference to the drawings, 4N, the embodiment of the invention will be explained.

[Explain each row. FIG. 1 is a schematic 114 diagram showing a keyword detection section of a character recognition device constructed by applying the method of the embodiment. This device is used, for example, to detect the position of the keyword "town" from among the addresses written on letters and mail. The input unit 1 inputs 1. [] The information posted,
The cutting section 2 and 11 take a density projection in the direction in which the characters of the document image are written, and detect the line position indicating the character string by detecting the portion where the black area exists. . The image information (character string 1# information) of the line detected in this manner is stored in the line memory 3 line by line. Note that the above characteristics do not indicate the character string, but rather the pixels of the image that constitute the character string. For example, as shown in FIG. 2, the sandal section 4 is composed of a plurality of D/A converters 4 and a plurality of operational amplifiers 4bK, and after D/A converting each data of the pixel row, A so-called scarecrow process is performed by increasing the addition lit between data in a plurality of rows, and the data is converted into r columns of pixel data. This process results in one line of ii! Is the character string indicated by the +i element line? It will be converted as r-string character string data that has been subjected to oak processing. As explained below, for the image information of the r string indicating the character string obtained in this way,
Yu? 4: ``-''-2-nl ^, 1 no L medium I
ff +' - all b The degree of similarity between the 9-turn image and the dictionary pattern of the keyword registered in the dictionary memory 5 in advance is divided. In addition, -1,: notation 91・
? 1 in turn, the above-mentioned character (edge part 2
After being normalized according to information about the size of the drawn characters, that is, the number of rows of pixels that make up the characters, the characters are subjected to the similarity calculation. Moreover, class Q+i, degree it *, j
When using the composite similarity method with l
Furthermore, if the above-mentioned correct cutting and conversion are %V/ll, the size of the turn may be made to match the width of the character image after the sampling processing, and this is the most effective
This is considered to be a habit. Also, in this case, since only the width of the above line can be obtained as normalization information, it is possible to specify only the horizontal or vertical size of the dictionary) or turn, but the curved blade can be normalized at the same ratio. Just do it like this. Therefore, the similarity calculation unit 1F, record the similarity between the dictionary pattern of the keyword 2 and the pattern cut out by window processing of the image information of the character string obtained from the sampling unit 4. The window position is moved little by little in sequence, and is configured as shown in FIG. 3, for example. In the first part 7 of the configuration shown in FIG.
l SAW (5urf
ace Acou@tic Wave) device 7a
is used to propagate this device 7a (window processing is performed using the if signal, and the intervening signal (up to the window and inspected) turns), which dictionary pattern, which correlation calculation is performed, and @device The correlation calculation results determined by 7a are combined via an adder 7b, 2 multiplier 7C, etc., and a composite similarity value is obtained corresponding to the window position.A method of correlation calculation η using this SAW device For example, η1.A Journal of Communication Society Vol. 1.65
, Ifz ], 2 pp. 1267-1273, well-known techniques can be used as they are. According to the similarity calculation unit 7 having such a configuration, while changing the window position for the character string little by little in a direct kJI manner, at the same time, the similarity between the window-processed pattern and the dictionary/arm turn is divided. It's possible, so it's very
It is 611 go. Thus, in this way, the one degree of similarity determined corresponding to the window position 1h'' is sequentially stored in the histogram memory 8, and the histogram of the *i (H degree) is formed in this memory 8. The peak detecting unit 9 determines the window position indicating the peak value from the similarity histogram, and the determining unit 10 discriminates the peak value using a jet constant threshold and determines the window position indicating the peak lll'I greater than or equal to a predetermined value. is detected as the position of the keyword with respect to the input character string. Figure 4 shows the alprism for detecting the peak ffL in the histogram.
The histograms are read out in sequence corresponding to the window position, and it is determined whether the histogram changes are upward or downward. Then, the window position immediately before the time when the histogram changes from upward to downward is detected as the peak position. As described above, through a series of processes in each section, a keyword position is detected as a window position indicating a peak from a class 1 or higher histogram of a window-controlled character pattern and a keyword dictionary pattern. In addition, 11 in the figure is a series of operations of each part described in 2.
This is a control unit that performs t'J. In this way, according to this method, it is possible to very effectively, easily and reliably detect the keyword position in a character string that is freely marked with W, without specifying the character position @ using a character entry frame. can. In other words, after window processing, iτ1;
A publicly available dictionary of character string patterns and keywords
A histogram of the similarity with the e-turn is created by changing the window little by little, and the keyword position is determined from the peak value of this histogram, so the detection accuracy is very high. Therefore, great effects can be achieved as pre-processing for address recognition of letters and mail. Note that the present invention is not limited to the embodiments described above. For example, as shown in FIG. 5, the processing 't' RUM 4 a f: may be performed in the Pungle unit 4, and the related performance may be performed using a digital LSI (for example, TRW tllllJ TDC-100RJ) 7
It may also be done using d! Although the case where the window is moved linearly in the direction of the character string has been described, the moving speed may be determined according to the processing speed of the correlation calculation. Furthermore, if the first peak in the histogram occurs only once, that is, if it is proven that only one keyword exists in the character string, only a simple peak detector as shown in Figure 6 can be used. It is possible to configure the device by detecting peaks using the method and omitting the histogram memory. In addition, the present invention may be applied to llRh without departing from the gist thereof.
It can be modified and implemented.

[Brief explanation of the drawing]

The figure shows a keyword detection device Kf constructed by employing method f according to an embodiment of the present invention, and FIG. FIG. 3 is a diagram showing the configuration f11 of the similarity calculation unit, the 41st bar is a diagram showing an example of the algorithm of peak 1 of the histogram, and FIG. 5 is a diagram showing another configuration example of the wrangler/similarity calculation unit. , FIG. 6 is a diagram showing another configuration example of the t-beak detection section. 2... Discontinuation part, 3... Line memory, 4...-Linol part, 5... Dictionary memory, 6... Normalization part, 7.
...Similarity calculation unit, 8...Histogram memory, 9.
...Peak detection section. The applicant's representative, patent attorney Suzue Ni, is the representative of Figure 3 and Figure 4.

Claims (3)

[Claims] (1) While moving the window position relative to the text area, calculate the degree of similarity between the pattern cut out from the text area using the window and the keyword pattern.
A keyword detection method comprising: calculating a histogram of similarity to two keywords, and detecting the position of the keyword in the previous 11 character area from the peak position of this histogram. (2) The keyword 〃・i′ written pattern is normalized according to the character size estimated from the character width of the character area and is then used for the similarity R1′R. Section 6 [Keyword detection method described in '2. (3) The similarity is calculated by 1 according to the composite similarity fll19 method. Claim 1: Ne tliZ
(4) The movement of the window position is performed linearly along the character string in the character area.
Keyword detection method described in section.