JPH061507B2 - Character string recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a character recognition device for optically reading characters, and more particularly to a character string reading device for reading a specific name.

(Prior Art) Conventionally, this type of character recognition device has been developed as an effective data input means. However, under the present circumstances, it is the actual situation that the target is a print character, a handwritten character written in a limited character frame, or that each character has various limitations such as being completely separated. is there.

As an example of the recognition method of the printed character string, RG Casey `` Recursive Segmentation and Clas
sification of Composite Character Pattens) "6th International Conference on Pattern Recognition, Proceeding, pp1023-1026 (19)
82), (Proc.6th ICPR,).

Also, as an example in which each character is completely separated, Yasuaki Nakano et al., “Understanding Tabular Documents In Cooperation With Character Recognition,” IEICE Transactions '86 /3Vo1.J69-DNo.3 (1986) can give.

However, in general, handwritten character strings are often not written in the character frame. To read a string like this,
It is a very effective input method, but in this case, the characters are usually in contact with each other, or they are arranged in an uneven manner, and the kanji contains separated characters. Therefore, under the present circumstances where the character segmentation performance of separating the character string into individual characters before reading the character is not sufficient, the character recognition device described above allows relatively free handwriting such as addressing of mail. Characters are difficult to read.

(Problems to be Solved by the Invention) In the conventional character recognition device described above, in the current situation that the character segmentation performance is not sufficient, it is difficult to accurately read handwritten characters relatively freely written such as the address of a mail. It has the drawback of being difficult. Further, the above word recognition method has a drawback in that it is difficult to obtain sufficient accuracy and processing speed because the order sequence between words is not used.

The present invention extracts a specific character from the beginning of each character line,
By repeating the operation of removing the extracted part from the character line, the character can be stably extracted even when the character comes into contact with or is separated, and the recognition target is composed of a combination of a few words. However, by deciding on a word-by-word basis, regardless of the image size or the number of characters in a word, efficient and accurate character recognition is possible. It is to provide an accurate character string recognition device.

(Means for Solving the Problem) As an input character graphic, an input pattern buffer that holds a binary value by designating the two-dimensional coordinates of the character string, and an input pattern buffer based on the input pattern A feature extraction unit that creates a direction image based on the direction contribution value for each line segment direction in a format corresponding to, a word storage unit that stores words that are known to appear in advance and an ordered sequence between the words, The standard pattern of a character graphic is expressed as a series of line segments, and a standard character pattern storage unit for storing the direction variation of each line segment and the allowable variation of the line segment and the relative position information of the line segment is given. The character recognition unit which extracts a character whose existence position is unknown from the beginning of the character string image by referring to the character standard pattern, obtains the circumscribing frame of the character and the extraction reliability, and the result of the character recognition unit. And above By setting the image to be read next from the input pattern and direction image and specifying the character to be extracted next in the character recognition part, character recognition and character cutting are alternately performed, and recognition reliability and extraction for each word are performed. Based on the results of the word recognizing unit and the word recognizing unit for determining the position of the circumscribing frame of a word, determination is made between the words, and the input pattern and direction image are reset based on the corresponding result. In addition, it has a determination section that alternately performs word recognition to be read and cut out a word image, and designates a word to be extracted next as a word recognition section according to an order sequence between words.

(Operation) A relatively freely written character string (for example, a mail address
The technique of character cutout that determines which area each character is in when optically reading (address, etc.) is an important technology together with the technology of character recognition. To do that, letters as well as words need to be recognized. Therefore, in the present invention, it is assumed that the characters are read in order from the beginning of the character string, and a specific character image is extracted from the beginning of the image of the character string. The extraction method is to extract by tracing the line segment according to a standard pattern, assuming that the character is composed of the line segment. By extracting the character, it is possible to cut out the character from the image of the character string even if the character is in contact. Therefore, in the conventional character recognition, the character image was input and the recognition result was used as the character category name, whereas here the character category name and the character string image were input, and the extracted character image and its extraction reliability were input. It is output.

Furthermore, Japanese characters have similar characters (for example, “question”, “between”).
Etc.) and characters that can be read even if separated (for example, “dai” → katakana “mu”, “b”), it is expected that stable identification can be expected if the identification result is performed in word units.

In this case, if the word appearing in the character string is known, post-processing character string recognition that arranges the identification results for each character and compares the result with the dictionary of the sentence that is a word or a series of words to determine the result, It is more efficient and stable to perform recognition by first selecting a word that can appear in the character string and performing identification.

Further, when there is an ordering rule between words, if the candidates of the word that will appear next can be narrowed down based on the result of the previous word recognition, faster and more stable identification can be performed.

The problem with this method is the determination of whether or not a character could be extracted in character recognition.In this method, a line segment is traced based on the character standard pattern, and it is used for line segment tracing. It is not possible to extract if the line segment tracing is not possible by evaluating the line portion on the character string that did not exist and if there are many line segment portions on the character string that were not used for line segment tracing. The above problem is solved by making the determination.

(Example) Next, this invention is demonstrated with reference to drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. In FIG. 1, an input pattern 1 is a two-dimensional, binary character string image. It is described by introducing coordinates in which the line segment part is a black pixel, the background part is a white pixel, and the horizontal axis is x and the vertical axis y. FIG. 2 shows a leading portion of a character string of "Shitaya, Taito-ku ..." As an example of the character string image. In the figure, the hatched portions indicate black pixels and the like.

The input pattern is converted into a direction image 3 by the feature extraction unit 2. The direction image is a binary image formed by checking which direction each black pixel of the character string image belongs to, and dividing each pixel according to that direction.
Four types, such as vertical and upward left, are used. The method for generating the direction image is a conventional technique.

FIG. 3 shows an example of the standard pattern 4 used in the present invention. The standard pattern will be described with reference to FIG. As the standard pattern 4, it is assumed that character graphics are described and held as a series of line segments. Further, since the curved line portion of the character line can be approximated by a piecewise straight line, the standard pattern is represented by a graphic with directed line segments as shown in FIG. Here, the direction of each line segment is quantized into eight directions as shown in FIG. Note that the quantization is not limited to 8 directions, but 4 directions, 1
Although different methods such as 6 directions are conceivable, the following description will be given using an example of 8-direction quantization for simplicity. For characters that cannot be expressed with a single stroke, a virtual line is introduced that means only the connection of line segments. In FIG. 3, the solid line represents a line segment and the dotted line represents a virtual line. Further, in the standard pattern, the character model is such that the line width of the character is constant. Usually, the line segment series is given according to the stroke order. According to the present invention, the standard pattern does not hold the standard pattern as shown in FIG. 3 as a two-dimensional figure as it is, but the allowable variation range (ik) of the information d _k indicating the direction of the Kth line segment and the length of the line segment. , mk) and four types of relative position information between the line segment start point U _k , D _k , L _k , R _k <K (K = 1, ... K; K represents the number of line segments and in the case of FIG. 3, K = 5) and hold it.
The direction d _k is the direction vector (Δ _xk , Δ) in the xy plane.
_yk ) (where Δ _xk = {0, 1, −1}, Δ _yk = {0,
1, -1}) is considered as an example.
(Fig. 4) Now, when the Kth line segment is shown by V _k , the standard pattern B is It is expressed by the equation (1) and is held in a form that can be referred to with respect to K. Note that the standard pattern has one or more for each character class to be recognized.

In the present invention, the shape of the standard pattern is not fixed, but the length of each line segment
It is considered that various standard patterns are generated by changing V _k | within an allowable range. That is, when | V ₃ | of the character “bottom” shown in FIG. 5 (a) is expanded, it becomes as shown in FIG. 5 (b).
When V ₂ | is further shortened, it becomes as shown in FIG. Similarly, if the length of each line segment is changed variously, a variegated font (deformed figure) can be obtained, but these are all the patterns of the character "bottom". However, if the characters are changed too much, different characters may result in different characters. Therefore, the allowable length variation values ik, mk of each line segment and the four types of relative position information U _k , D _k , L _k between the line segment start points are set. , R _k are introduced to control the change of fonts to prevent erroneous recognition due to too many characters, and to reduce the number of generated figures to improve the processing efficiency. Further, mk is also used to classify whether the line segment is a solid line or a virtual line, and mk> 0 represents a virtual line with a solid line mk <0. ik, mk indicates that the length of the line segment changes only within the range of ik ≦ | V _k | ≦ | _mk |, and U _k , D _k , L _k , and R _k satisfy the following conditions, respectively. represents the segment number appearing in the K-th previous segment sequence, the starting point of the K line segment with reference to the starting point of the line segment with a previous segment sequence, below the starting point of the U _k segment Yes, above the starting point of the D _k-th line segment, to the right of the starting point of the L _k-th line segment, and to the left of the starting point of the R _k-th line segment. Therefore, U _k , D _k , L _k , R _k <K, and when there is no line segment to refer to, it is set to 0. (Always U ₁ = D ₁ = L ₁ = R ₁ = 0.) In short, in the present invention, the standard pattern is represented by a series of vectors, and each vector is defined by its direction, length, and starting point position. Of these, the above parameters
dk controls the direction, ik and mk control the length, and U _k , D _k , L _k , and R _k are given to each line segment as parameters for controlling the position.

When the character recognition unit 5 of FIG. 1 receives a character name, a character string image, and a character string direction image as inputs, it uses the standard pattern corresponding to the character name from the standard pattern 4 to start from the beginning of the character string image. Matching is performed to obtain the best matching character graphic (hereinafter referred to as a matching character graphic) and its extraction reliability. Here, obtaining the matching character figure by the standard pattern is equivalent to uniquely determining the length | V _k | of each line segment V _k of the standard pattern so as to best match the input image.

The matching character graphic is obtained as a line graphic shown by a solid line in FIG. 7 by collating the character string image as shown in FIG. 6 using the standard pattern of FIG. If the coordinates x ₁ , x ₂ , y ₁ , y ₂ of the outermost contour are obtained using this matching character figure, the area of the character figure can be easily determined and the character can be cut out.

The output of the character recognition unit 5 is the area of this character cutout and the above-mentioned extraction reliability. However, if the matching character graphic cannot be extracted, a reject code is output instead of the reliability.

A specific example of detecting a matching figure and its extraction reliability from a character string is described in detail in the applicant's application (Japanese Patent Application No. 61-307052 (Japanese Patent Laid-Open No. 63-157290), character recognition device). Therefore, details are omitted.

In the word storage unit 6, the word name, the level to which it belongs, and the level of the word that appears next are registered. The word level defines the order relation between words.
Words that can appear in the same position are preliminarily classified as one level.

Hereinafter, the present invention will be described by taking the address reading in Taito-ku, Tokyo as an example. FIG. 8 shows a part of the word dictionary of Taito-ku, Tokyo. In FIG. 8, level 1 is composed of words that may come at the beginning of the character string, level 2 is the ward name, and level 3
Is composed of the words that may occur after the ward name.

The determination unit 7 receives the character string image 1 and the direction image 3,
The words of the level to be read are fetched one by one from the word storage unit 6 and sent to the word recognition unit 9, and the word recognition unit receives the word extraction reliability of the word from the character string image and the position of the extraction circumscribing frame. The determination unit compares the word extraction reliability for each word, determines the word that has the highest extraction reliability at the same level, outputs the result as the determination result 13, and should be read next according to the result. The word level is set, and the character string image and its direction image are reset using the corresponding extracted circumscribing frame positions. This process is repeated until there is no more reading level or word extraction cannot be performed.

The case where the input pattern of FIG. 2 and the word structure of FIG. 8 are given is illustrated. It is difficult for the determination unit to retrieve the first level 1 word into the word storage unit. First, the character string image of FIG. 2, the direction image corresponding to it, and the word name “Tokyo” are sent to the word recognition unit, and the image of the word specified by the word recognition unit is extracted from the character string. Receive degree C (Tokyo) and position P (Tokyo) of the circumscribing frame. The same processing is repeated for the other words of level 1, and the extraction reliability levels C (Tokyo), C (Taito Ward) and the position P (Tokyo) P (Taito Ward) of the circumscribing frame are obtained. If C (Tokyo) <C (Tokyo) <C (Taito Ward), the judgment result is set as "Taito Ward" and this result is output to 13 and the input image is based on the frame size P (Taito Ward). Then, the frame size portion is removed from the direction image and the direction image, and the image is reset so that the next portion of the input image and the direction image comes first. FIG. 9 shows a matching character figure of each character of “Taito Ward” in FIG. 2 by a solid line, and shows a character string image reset based on the word circumscribing frame size P (Taito Ward). Then, the level 3 to be read next is set, the character string image is set to FIG. 10, and the direction image is set to correspond to this, and the above processing is repeated.

When the word recognition unit 9 receives the direction image corresponding to the character string image and the word name to be extracted from the determination unit 7, the characters that form the word name are designated for character recognition from the beginning. When designating the character name at the beginning of a word, the character string image and the direction image received from the determination unit 7 are sent to the character recognition unit 5 as they are, and the character recognition unit 5
The image of the designated character is extracted, and the extraction reliability c and the extracted character circumscribing frame position p are received from the character recognition unit. To specify the second and subsequent characters of a word for character recognition,
First, the character image and the direction image in which the extracted portion of the image for the previous character in the same word is removed are set as in the determination unit, and the extraction reliability c and the circumscribed frame position are set in the same manner as the first character. Find p.

If a reject code is returned from the character recognition part,
Immediately, the processing in the word recognition unit is stopped and the reject code is passed to the determination unit. Otherwise, the extraction reliability c and the character cutout position p are sequentially obtained for all the characters included in the word. The average value of the extraction reliability of each character is used as the word extraction reliability C, and the combination of the circumscribing frames of each character is output as the word circumscribing frame position P to the determination unit.

A case will be described as an example where the determination unit gives the word recognition unit the character string image of FIG. 2 and the corresponding direction image and the word name “Taito-ku”. (Refer to FIG. 9) First, the character string image of FIG. 2, the corresponding direction image and the character name “stand” are output to the character recognition unit, and the character recognition unit extracts the extraction reliability c (stand) and the character cut-out position. Enter p (unit). If the result is not reject, the portion corresponding to p (stand) in the table 9 in FIG. 9 is removed from the character string image and the direction image, and the remaining portion below p (stand) is newly added to the binary character. Set as a row image and a direction image. And
Next, the newly set image and the character name "East" are output, and the extraction reliability c (East) and the character cut-out position p (East) are obtained as inputs from the character recognition unit. After that, p
The character string image and direction image are newly set based on (East), and the character name "ku" is output together with these and the extraction reliability c
Input (district) and character cutout position p (district).

In this way, the extraction reliability of each character c (unit) c (east) c (ward)
If is found Then, the word extraction reliability C (Taito-ku) is obtained, and the word cut-out position P (Taito-ku) is set to P (Taito-ku) = p (ku) and these are output to the determination unit.

By repeating the processes of the determination unit, the word recognition unit, and the character recognition unit as described above, the character string is sequentially recognized for each level registered in the word recognition unit.

(Effects of the Invention) As described above, the present invention enables recognition of a character string including a connected character without a character frame by fusing character recognition and character cutout, and character recognition and character The processing efficiency is higher than that of cutting out separately, and since the judgment is performed on a word-by-word basis and the word name is first determined and character recognition is performed continuously, the recognition rate is improved, and in addition, an order sequence between words is used. By providing the above, the processing efficiency and the recognition rate are improved, and there is an effect that a handwritten character string relatively freely written such as a mail address can be read.

[Brief description of drawings]

FIG. 1 is a functional block diagram showing an example of the configuration of the present invention. FIG. 2 is an example of an input image applied to the present invention. FIG. 3 is an example diagram for explaining a standard pattern.
The figure shows the directions that each line segment of the standard pattern can take. FIG. 5 shows that various character graphics can be expressed by the standard pattern of FIG. FIG. 6 shows an example of a binary character string image input to the character recognition unit. FIG. 7 shows the character string image of FIG. 6 extracted by the character recognition unit using the standard pattern of FIG. The matching figure and its character cutout position are shown. FIG. 8 is an example showing the configuration of the word storage unit. FIG. 9 shows the word “Taito-ku” for the input pattern shown in FIG.
It is a figure showing a matching figure of each character at the time of extracting, a cut-out position of each character, and a cut-out position of a word. FIG. 10 is newly set after the word is cut out by the judgment unit when the judgment result is “Taito-ku” at level 1 for the input pattern of FIG. 2 and the word storage unit of FIG. It is a binary image showing a character string. In the figure, 1 ... Input pattern, 2 ... Feature extraction unit, 3 ... Direction image unit, 4 ... Character standard pattern, 5 ... Character recognition unit, 6 ... Word storage unit, 7 ... Judgment unit, 9 ... Word recognition unit, respectively. Show.

Claims (1)

[Claims] 1. An input pattern buffer for holding a character string as a input character graphic in a format in which a binary value can be referred to by designating a two-dimensional coordinate, and a format corresponding to the input pattern buffer from the input pattern. The feature extraction unit that creates multiple direction images based on the direction contribution value for each line segment direction and the standard pattern of each character graphic are expressed as a series of line segments, and the variation of each line segment direction and the allowable line segment length A character standard pattern storage unit for storing the amount and the relative position information of the line segment, and a word storage for storing a word known to appear in a character string and an order sequence between the words in advance. Part and the character to be read from the beginning of the given character string image are collated by referring to the character standard pattern, and the character recognition part that obtains the character cut-out position and extraction reliability of the character and each character that constitutes the word to be read Name and By giving a partial image to be collated in the force pattern and the direction image to the character recognition unit, and by character cutting position as a result of the character recognition unit, character cutting is performed to create a partial image to be collated next, Character recognition and character segmentation are performed alternately, and determination is made between words based on the results of the word recognition unit and the word recognition unit that obtain the extraction reliability of the word and the position of the circumscribed frame of the extracted word, that is, the word segmentation position. By making a partial image of the input pattern to be read next and the direction image based on the word cut-out position of the corresponding result, and giving the word to be read next to the word recognition unit according to the order sequence between words. A character string recognition device having a determination unit that performs identification determination on a word-by-word basis while alternately performing word recognition and cutting out a word image.