JPH0812683B2 - High speed extraction method for specific character strings - Google Patents

Description

The present invention relates to a method for extracting a character string of a document image, and more particularly to efficiently extracting only a specific character string registered in advance in the system from the document image. The present invention relates to a suitable specific character string high-speed extraction method.

[Prior Art] Conventionally, regarding a character recognition processing method for improving reading accuracy by using linguistic information of an input character string, Sugimura: error correction processing method of kanji recognition by candidate character complement and morphological analysis, IEICE Information Page 1-307 from 1-3 of National Convention (1985) 1-3
Discussed on page 08. The character recognition system there has a word dictionary, recognizes the whole sentence, performs morphological analysis on the sequence of candidate characters that is the conventional recognition result for each character, and corrects errors when there is a correct character at the second or lower position. To do.
According to such a method, the recognition accuracy can be improved as compared with the case of individually recognizing each character.

[Problems to be solved by the invention]

The above-mentioned conventional technology uses a priori character string information based on a word dictionary for character recognition.However, since the entire text is recognized, it should be taken into consideration that only a specific character string is efficiently extracted. However, there is a problem that it takes a long processing time.

An object of the present invention is to provide a specific character string high-speed extraction method suitable for efficiently extracting only a specific character string registered in advance in a system from a document image.

[Means for solving problems]

The above-mentioned purpose is to cut out a character from a character image (hereinafter referred to as a document image) on a document, calculate a feature amount of the cut out character, and collate the calculated feature amount with a character feature amount registered in advance. , Quantize the calculated feature quantity, find in advance how the particular character string to be extracted on the document image corresponds to the quantized character feature quantity sequence, create a table, and refer to this table to input character image This is achieved by obtaining a candidate specific character string from the quantized character feature amount sequence of, and performing recognition processing of matching or non-matching of the character pattern for this specific character string.

[Action]

Generally, in pattern classification, when the number of classes to be classified is large, it can be realized by a significantly smaller number of operations to determine whether or not a class belongs to only some specific classes than to completely identify the classes to which the classes belong. This tendency is
It becomes more prominent when determining the arrangement of specific patterns. This is because normally, and have regularity. The extraction of the specific character string from the document image is just this case.

FIG. 2 illustrates the above fact. If the characteristic parameter of the character pattern is represented as x, and the probability distribution of each character i in the feature space is represented as P _i (x); i = 1, ..., Number of character types, the probability distribution P _i (x) is They are distributed with overlap as shown in. In ordinary print character recognition, a character that gives the maximum value of each probability distribution P _i (x) to the input pattern x is set as the recognition result. However, in actual calculation, the character having the smallest distance d ( _i , x) from the average pattern _{i of} each character is often selected. Therefore, recognition of one character requires distance calculation only for several character types. On the other hand, the distance calculation may be performed once or a small number of specific characters to determine whether or not the input pattern x may be the specific X. Moreover, since a character string that is a combination of characters does not allow completely random combinations, even if the probability distributions for one character have an overlap, the probability distribution of the character string in the product space of the character feature space There is less overlap. Therefore, even if a coarse feature amount having a large overlap of probability distributions is used, the character string recognition has a relatively high accuracy.

〔Example〕

An embodiment of the present invention will be described below with reference to FIG. The document image data stored in the optical disc or the like is read from the optical disc device 1 and input line by line to the character frame clipping device 2, and the minimum circumscribing rectangle information of each character, that is, the coordinates of the upper left lower right vertex It is output and stored in the circumscribed rectangle information section 4 of the character information table 3. As the character frame clipping device, the device detailed in Japanese Patent Application No. 60-184242 "Document character clipping image processing method" is used here. When the character circumscribing rectangle information is extracted for the entire document image, the circumscribing rectangle information is read from the character information table 3, and the rough feature amount calculating device 5 calculates a rough feature amount for each character. As the coarse feature amount, the complexity index which is described in detail in p78 to 79 of Literature: Introduction to Character Recognition (1982) is used. The complexity index is the total length of the vertical and horizontal components of the outline of the character pattern. FIG. 3 shows a 2 × 2 mesh element pattern for obtaining the total length of the contour line. The element pattern is the vertical pattern V ((a) in the figure),
There are four types of patterns, a horizontal pattern H (FIG. 2B), a diagonal one side pattern L (FIG. 7C), and a diagonal both side pattern T (FIG. 2D). The thick line in the figure is a polygonal line approximation of the outline of the character pattern. N (V), n (H), n (L) and n (T), the total number of each element pattern in the entire character
Then, the vertical and horizontal components of the contour line are obtained. Therefore, the horizontal and vertical direction complexity indices l _x and l _y are obtained by the following equations.

The horizontal and vertical direction complexity index for each character obtained by the rough feature amount calculation device 5 is stored in the rough feature amount portion 6 of the character information table 3 in association with the character frame information of the circumscribing rectangle information portion 4.

Next, the character rough feature amounts are read from the character information table 3 in the order of character strings, coded by the quantizer 7 into codes m (4 bits) of 0 to 15, and stored in the shift register 8. The quantizer 7 divides the complexity index in the two directions into four sections with three thresholds, and determines which of the 16 sections is defined. The shift register 8 has three consecutive
The coarse feature amount quantization code for characters is stored, and the address calculator 9 considers the code for 3 characters as 12-bit data, and by referring to the table of about 4 kw, the corresponding address on the character string table 10 Ask for. The specific character string to be extracted is a divided section of about 4K in advance (3 of 16 divided sections).
Which of the product sections) to enter is determined by the rough feature amount, and the divided section is collected as a code string and stored in the character string table 10. However, the divided sections for this determination may overlap each other to allow one character string to correspond to a plurality of sections. The output address of the address calculation unit 9 is the start address of the code string data regarding the division section. However, the number of code strings in the divided section is stored in the head address itself. For the three characters corresponding to the coarse feature quantity quantization code on the shift register 8,
The content of the character string table 10 referred to by the output address of the address calculator 9 indicates the number of candidate specific character strings.
When the number of specific character strings is 0, the determination device 11 determines that there is no specific character string that is a candidate, and the character information table 3
Then, the next character rough feature amount is read out and the character string table 10 is referred to in the same manner.

If the reference result of the character string table 10 is not 0, the character code strings of candidate character strings are successively displayed in the character string table.
It is read from 10, and is converted into the precise feature amount of the character pattern corresponding to the character code by the character precise feature amount table 12. As the precise feature amount of the character pattern, the character pattern K (x,
y) Use yourself. Where x and y are position coordinates in the horizontal and vertical directions, respectively. Is. The fine feature amount register 13 stores the fine feature amount of the candidate character string {▲ K ⁱ ₁ ▼ (x, y), ▲ K ⁱ ₂ ▼ (x, y), ▲ K ⁱ ₃ ▼.
(X, y)} is stored. The subscript i indicates that it is the i-th candidate. On the other hand, when the number of candidate character strings is not 0, the determination device 11 activates the fine feature amount calculation device 14 to calculate the fine feature amount regarding the character string in the shift register 8.
That is, the corresponding character pattern is cut out from the optical disk device 1 and stored in the to-be-determined fine feature quantity sequence register 15. The precise feature amount sequence of both registers 13 and 15 is obtained by the distance calculation device 16 in the following formula, that is, the degree of difference, The threshold determiner 17 determines the degree of difference. If the dissimilarity exceeds the threshold θ, the character string to be judged in the input document image cannot be a candidate character string, and a candidate character string whose dissimilarity does not exceed the threshold θ is selected. Output the character code string with the degree. When there are a plurality of candidate character strings whose dissimilarity does not exceed the threshold value θ, the character code string giving the smallest dissimilarity or all the character code strings are ranked and output. If the differences of the candidate character strings are all threshold θ
If it exceeds, it is determined that the character string to be judged is not one of the specific character strings to be extracted, the character information table 3 is activated, the next character rough feature amount is read, and one character is shifted and three characters are shifted. The above determination process is performed on the character string.

In the embodiment described above, the complexity index is used as the rough feature amount of the character pattern, but it is also possible to use another rough feature amount such as the marginal distribution or the reduced pattern itself.

〔The invention's effect〕

According to the present invention, a specific character string to be extracted on a document image is classified in advance based on the arrangement of coarse character feature amounts. Therefore, a coarse feature amount with a small amount of calculation is used from the input document image character string. Since the registered specific character strings that are candidates are narrowed down and the specific character strings can be extracted at high speed as a result, only the specific character strings registered in the system can be efficiently extracted from the document image.

[Brief description of drawings]

FIG. 1 is an overall system configuration diagram of an embodiment of the present invention, and FIG.
Fig. 3 is an explanatory diagram of the probability distribution of the character pattern in the feature space.
The figure is a diagram showing an example of an element pattern for calculating a complexity index used as a rough feature amount of a character pattern.

Claims (3)

[Claims] 1. A character image is input, a character string is cut out from the input character image, a feature amount column of the cut out character string is calculated, and the calculated feature amount column and a specific character prepared in advance. A specific character characterized by collating the column of the feature amount related to the column while shifting the character position, and outputting only the collation position and the character string in which the collation result of all the characters constituting the character string is equal to or greater than a predetermined threshold value. High-speed column extraction method. 2. The collation process is a process of collating and removing a character string in an input image that cannot be a specific character string to be extracted with a rough feature amount, and a collation failure with a rough feature amount is not unsuccessful. The high-speed extraction method of a specific character string according to claim 1, which comprises a process of matching only a character string in the input image that has not been detected with a fine feature amount. 3. The collating process quantizes a rough feature amount of the input character string to obtain a quantized character feature amount sequence, and a specific character string to be extracted on a document image is the quantized character feature amount. How to correspond to the columns is obtained in advance and made into a table, the specific character strings that are candidates are obtained by referring to the table from the quantized character feature sequence, and the matching of character patterns for the specific character string The specific character string high-speed extraction method according to claim 2, which comprises a process of performing matching with a fine feature amount.