JP2010039615A - Character recognition method and character recognition apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce processing time and obtain an accurate character recognition result in recognizing characters printed in half-, full-, and double-width characters even if characters are blurred due to contact and joint components are divided. <P>SOLUTION: A half-width likelihood indicating the degree of half-width likeliness of characters in a character line, a double-width likelihood indicating the degree of double-width likeliness, a character blur likelihood indicating the degree of character blur, and a character contact likelihood indicating the degree of character contact are calculated. Pre-segmentation is performed using the character contact likelihood and double-width likelihood, and forced separation is performed using the character blur likelihood and half-width likelihood. Character recognition is performed on character patterns resulting from the pre-segmentation and forced separation. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a character recognition device that recognizes characters from an input image and acquires text information.

  In the character recognition process, first, the character line extraction process that extracts the character line that is the area where the characters are arranged from the input image is executed, and the character extraction process that extracts each character from the character line is executed. Then, the feature of the character pattern that is the extracted character is extracted. Next, characteristics of a plurality of character types are prepared in advance as a dictionary, a character identification process for matching the characteristics of the prepared dictionary and the characteristics of the character pattern to be recognized is executed, and the degree of similarity that is the degree of matching The character type having the highest (identification likelihood) is output as the recognition result.

  In the character extraction process, a connected component that is a series of black pixels is generated, and a character pattern is generated by combining the connected components or an element obtained by combining the connected components in the vertical direction as a unit. Perform pre-segmentation processing. Hereinafter, the unit of the constituent element is referred to as an element, but the unit of the constituent element may be a connected component. When the characters are crushed or the characters are printed close to each other and the characters come into contact with each other, a plurality of characters may be combined into one element. Therefore, the character cutting process includes a forced cutting process that cuts one element and generates a plurality of elements.

  In general, characters to be printed have character widths of half-width, full-width, and double-width. Conventionally, as described in Patent Document 1, when character recognition is performed, since it is not known which character width is described, it is determined which character width the printed character is, The character is cut out with the cut character width, and the character identification process is executed on the cut out character.

  Further, when the character width cannot be specified as in the case of handwritten characters, the character segmentation is not necessarily uniquely determined in the character segmentation process. Therefore, a plurality of character segmentation candidates are generated, a segmentation network that is a network of them is generated, and an optimal path search is performed to select a sequence of character identification results so that the identification likelihood that is the character identification result is high, and recognition is performed. Get results.

  In addition, a word dictionary that is a list of words is prepared in advance, a plurality of character types having a high likelihood of character pattern identification are combined, collated with words in the word dictionary, and combinations of character types collated in the optimum path search are determined. The recognition result can also be obtained with priority (word matching).

Furthermore, it is possible to prepare a word bigram, which is a probability that words A and B appear in the word list in advance, and to consider the probability of consecutive words in the optimum path search. By using word matching or word bigram, even if the first candidate of the character identification result is incorrect, a recognition result obtained by correcting the incorrect candidate can be obtained. This is called knowledge processing. These processes are disclosed in the international workshop on Frontiers in Handwriting Recognition.
Japanese Patent Laid-Open No. 06-150055 Katsumi Marukawa, 2 others, "Online Recognition of Unconstrained Handwritten Japanese Text Using Statistical Information", 2006

  However, when recognizing printed characters, kanji, hiragana, and connected components may be divided into left and right, so even if the method described in Patent Document 1 is used, the discrimination result of half-width, full-width, and double-width is always It is not always correct. Furthermore, when many characters that are crushed and touched are included, and when characters are blurred and the connected components are finely divided, it is more difficult to distinguish between half-width, full-width, and double-width. If the determination of half-width, full-width, or double-width is wrong, there is a high possibility that an erroneous character recognition result will be output. On the other hand, by generating a plurality of character segmentation candidates as in the above-described handwritten character recognition method and selecting an optimum combination of character patterns, a highly accurate recognition result can be obtained.

  However, the above-described handwritten character recognition method generates all character extraction candidates assuming half-width, full-width, and double-width, and increases the number of character patterns to be generated. There arises a problem of erroneously recognizing erroneously cut out character patterns.

  When recognizing characters printed in half-width, full-width, or double-width, the present invention suppresses processing time even when many characters are crushed and touched, and when characters are blurred and connected components are finely divided. In addition, an object is to obtain a highly accurate character recognition result.

  A typical example of the present invention is as follows.

  That is, according to one embodiment of the present invention, there is provided a character recognition method for recognizing characters from an image, and a connected component generation process for generating a connected component that is a series of black pixels constituting a character from an image including characters. , Character line extraction processing to extract the character line that is the area where the characters are lined up, half-width discrimination processing to calculate the half-width likelihood indicating the degree of half-width character of the character in the character line, degree of double-angle likelihood in the character line Double angle discrimination processing to calculate the double angle likelihood indicating the character, character blur detection processing to calculate the character blur likelihood indicating the blurring degree of the character in the character line, character contact likelihood indicating the contact degree of the character in the character line Character contact determination processing, noise determination processing for calculating a noise likelihood indicating the degree of noise mixed in the character line, and character contact likelihood and double-width character likelihood. Select the character pattern that cuts the connected component using the pre-segmentation process that generates the character pattern that is a combination of several connected components and the character blur likelihood and half-width character likelihood, and the position where the connected component is cut And a character identification process for extracting the characteristics of the character pattern generated by the pre-segmentation process and the forced cutting process and calculating the similarity with the characteristics of each character type in the prepared identification dictionary And.

  According to an embodiment of the present invention, it is possible to suppress processing time for recognizing characters printed in half-width, full-width, and double-width, and obtain a highly accurate character recognition result.

  Hereinafter, a character recognition method and a character recognition device to which the present invention is applied will be described in detail with reference to the drawings.

  FIG. 2 is a block diagram showing the configuration of the character recognition apparatus according to the embodiment of the present invention.

  A character recognition device 208 according to an embodiment of the present invention includes a communication device 201, an image acquisition device 202, a display device 203, an external storage device 204, a memory 205, a CPU (Central Processing Unit) 206, and an input device such as a keyboard and a mouse. These components are connected by a communication line such as a PCI bus.

  A program for executing the character recognition method according to the embodiment of the present invention is stored in a storage device such as the external storage device 204 and / or the memory 205 and executed by the CPU 206. An image including characters input to the CPU 206 may be input from the image acquisition device 202 or the communication device 201 such as a scanner, an OCR, or a digital camera. In addition, an image including characters may be stored in the external storage device 204.

  The result of character recognition is output to the display device 203. Further, the result of character recognition may be output to the outside via the communication device 201 or may be used for another program executed by the character recognition device 208.

  FIG. 1 is a flowchart showing the entire character recognition process according to the embodiment of the present invention.

  First, in the character line extraction process 102, the input image 101 is input, and a character line image that is an area where characters are arranged is extracted.

  FIG. 3 is an explanatory diagram illustrating an example of a result of the character line extraction processing 102 according to the embodiment of this invention.

  In the following description, the position on the character line image is the origin at the upper left corner of the character line image, the right in the left-right direction is the positive direction of the x-axis, and the lower direction in the vertical direction is the positive direction of the y-axis.

  Next, in the connected component / element generation process 103, a connected component that is a series of black pixels is generated from the input character line image, and the connected components that overlap in the vertical direction are collected, for example, an element that is a circumscribed rectangle. Generate.

  Thereafter, a double-angle discrimination process 104, a character contact discrimination process 105, a half-angle discrimination process 106, a character blurring discrimination process 107, and a noise discrimination process 108 are executed. These processes are desirably executed independently in parallel, but may be executed in order.

  Thereafter, a pre-segmentation process 109 is executed. In the pre-segmentation processing 109, half-angle likelihood, double-angle likelihood, character contact likelihood, character blur likelihood, noise likelihood, element, connected component, and character line image are input, a reference element is determined, and a reference element A character pattern is generated so that is at the left end.

  Then, the forced cutting process 110 is executed to separate one character pattern into a number of character patterns. Also in the forced cutting process 110, half-width likelihood, double-angle likelihood, character contact likelihood, character blur likelihood, noise likelihood, element, connected component, and character line image are used.

  Thereafter, character identification 111 is executed. In the character identification 111, the characteristics of the character patterns generated by the pre-segmentation process 109 and the forced cutting process 110 are extracted, and the similarity with the characteristics of each character type in the identification dictionary prepared in advance is calculated.

  Thereafter, knowledge processing 112 is executed. In the knowledge processing 112, the character segmentation network and the character identification result are input, and the character recognition result of the character line image is output.

  Thereafter, a recognition result use character width test 113 is performed to obtain a character recognition result 114.

  Next, details of each process will be described. First, details of the connected component / element generation process 103 will be described.

  FIG. 4 is a diagram illustrating an example of a connected component according to the embodiment of this invention.

  A circumscribed rectangle surrounding the connected pixels of the character line image is generated as one connected component.

  In the example illustrated in FIG. 4, an image indicating “autumn” includes four connected components surrounded by a circumscribed rectangle with a black frame.

  FIG. 5 is a diagram showing an example of an element according to the embodiment of the present invention.

  A circumscribed rectangle in which connected components are collected in the vertical direction is generated as one element.

  In the example shown in FIG. 5, the image indicating “autumn” includes two elements surrounded by a black frame.

  Next, details of the double-angle discrimination process 104, the character contact discrimination process 105, the half-angle discrimination process 106, the character blurring discrimination process 107, and the noise discrimination process 108 will be described. In these processes, the height CLh and the width CLw of the character line are used.

The simplest formula for calculating CLh and CLw is
(Expression 1) CLh = maximum y-coordinate value of black pixel-minimum y-coordinate value + 1
(Expression 2) CLw = black pixel x-coordinate maximum value-x-coordinate minimum value + 1
It is.

  Although the method for calculating the height and width of the character line does not consider the inclination of the character line, CLh and CLw may be calculated in consideration of the inclination of the character line.

  If the width of the element E is Ew and there are n elements {E1, ..., Ei, ..., En} in the character line image, the width of each element is set to {Ew1, ..., Ewi, ..., Ewn}. Furthermore, when the width of the connected component S is Sw and the height is Sh, and m connected components {S1,..., Si,..., Sm} exist in the character line image, each width is set to {Sw1,. , Swi, ..., Swm} and the height is {Sh1, ..., Shi, ..., Shm}.

  In the double angle discrimination process 104, a character line image, a connected component, and an element are input, and a double angle likelihood Ld indicating the degree to which the character in the character line image is a double angle character is calculated. There are various methods for calculating Ld using a character line image, a connected component, and an element.

For example, the following method can be employed.
A set value Vd is determined in advance, and using this set value Vd,
(Condition 1) Ewi> CLh × Vd
Calculate the total Ewt of Ewi of the elements that satisfy. Next, using Elwt, the double-angle likelihood Ld is
(Formula 3) Ld = Elwt / ΣEwi
And can be calculated.

  At this time, by setting Vd to a value larger than 1, Ld becomes a value indicating how much wide elements exist in the character line. Therefore, the larger Ld is, the higher the possibility that the characters in the character line are double angle.

  In the half-width discrimination process 106, a character line image, a connected component, and an element are input, and a half-width likelihood Lh indicating the degree to which the characters in the character line image are half-width characters is calculated. There are various methods for calculating Lh using a character line image, a connected component, and an element.

For example, the following method can be employed.
Using CLh and Ewi and the number of elements n, the half-width likelihood Lh is
(Formula 4) Lh = CLh × n / ΣEwi
And can be calculated.

  Lh is the reciprocal of the calculated average of the ratio of the character width to the height of the character line. Therefore, the greater the Lh, the higher the probability of being a half-width.

  In the character contact determination process 105, a character line image, a connected component, and an element are input, and a character contact likelihood Lc indicating the degree of contact of characters in the character line image is calculated. There are various methods for calculating Lc using a character line image, a connected component, and an element.

For example, the following method can be employed.
A set value Vc is determined in advance, and using this set value Vc,
(Condition 2) Ew / CLh> Vc
The number Nc of elements satisfying the condition is calculated. Next, using Nc, the character contact likelihood Lc is
(Formula 5) Lc = Nc x CLh / ΣEwi
And can be calculated.

  For example, by setting Vc to 2.5, it becomes a value indicating how much a connected component having a width wider than a double-width character is present in the character line. Therefore, the larger the Lc, the higher the degree of contact of characters in the character line.

  In the character fading determination process 107, the character line image, the connected component, and the element are input, and the character fading likelihood Lb indicating the degree of fading of the character in the character line image is calculated. There are various methods for calculating Lb using a character line image, a connected component, and an element.

For example, the following method can be employed.
Set values Vb1 and Vb2 are determined in advance, and using set values Vb1 and Vb2,
(Condition 3) Vb1 <Swi <Vb2 and Vb1 <Shi <Vb2
The number Nb of connected components that satisfy the above is calculated. Next, using Nb, the character fading likelihood Lb is
(Formula 6) Lb = Nb × CLh / ΣEwi
And can be calculated.

  For example, by adjusting and setting the value of Vb1 to 1/3 and the value of Vb2 to 2/3, the height and width of the connected component is 1/3 to 2/3 with respect to the height of the character line. This is a value indicating how much of the connected component is present in the character line. When a character is faint, the connected component of the original character is interrupted. Therefore, the proportion of the discontinuous connected component is a degree that the character in the character line is faint. For this reason, the greater the Lb, the higher the degree of fading of characters in the character line.

In the noise discrimination processing 108, a character line image, a connected component, and an element are input, and a noise likelihood Ln indicating a degree of noise in the character line is calculated. There are various methods for calculating Ln using character line images, connected components, and elements. For example, the following method can be employed.
Predetermining the set value Vn, using the set value Vn,
(Condition 4) Shi <Vn and Swi <Vn
The number Nn of connected components satisfying the above is calculated. Next, using Nn, the noise likelihood Ln is
(Expression 7) Ln = Nn × CLh / ΣEwi
And can be calculated.

  VLn is a value indicating the proportion of connected components in the character line in which the height and width of the connected component are smaller than Vn with respect to the height of the character line. For example, it is difficult to determine whether or not there is noise in the presence ratio of a small connected component in an area of one character size. On the other hand, it is easy to determine whether or not noise is present by using the proportion of small connected components in the character line. Therefore, the greater the Ln, the higher the degree of noise present in the character line.

  Next, details of the pre-segmentation processing 109 will be described. In the pre-segmentation process 109, a plurality of character patterns including elements within a certain distance Px from the left end of the reference element are generated.

  FIG. 6 is an explanatory diagram illustrating an example of the pre-segmentation process 109 according to the embodiment of this invention. In the example shown in FIG. 6, when there are four elements constituting “vocabulary”, a character pattern generated using element 1 as a reference element is shown. In the pre-segmentation process 109, a character pattern using element 2 as a reference element, a character pattern using element 3 as a reference element, and a character pattern using element 4 as a reference element are also generated. In the example shown in FIG. 6, the character pattern 2 includes only an element for one character, the character pattern 1 is partially missing, and the character pattern 3 and the character pattern 4 include extra elements.

  FIG. 7 is an explanatory diagram illustrating an example of the full-width character pre-segmentation processing 109 according to the embodiment of the present invention. FIG. 8 is an explanatory diagram illustrating an example of the double-byte character pre-segmentation processing 109 according to the embodiment of the present invention. FIG. 9 is an explanatory diagram illustrating an example of the pre-segmentation process 109 for half-width characters according to the embodiment of this invention.

  As shown in FIGS. 7, 8, and 9, the distance between the elements for one character differs between full-width characters, double-width characters, and half-width characters. When the value of Px is set so as to always generate a double-width character, the pre-segmentation processing 109 generates an extra character pattern such as the character pattern 3 and the character pattern 4 in FIG. Conversely, when the value of Px is set so as to generate a full-width character, a character pattern such as the character pattern 1 in FIG. 6 is generated in a character line including a double-width character.

  Thus, if Px is too small, a character pattern including elements constituting one character is not generated, and the character recognition accuracy is lowered. Also, if Px is too large, extra character patterns increase, and processing time increases. Furthermore, since there is a possibility that an extra character pattern may be recognized with high likelihood due to character blurring and character contact, character recognition accuracy may be lowered.

  In the embodiment of the present invention, half-width, full-width, Lh, double-angle likelihood Ld, character blur likelihood Lb, character contact likelihood Lc, and noise likelihood Ln are used to control the value of Px. Even if a character line includes double-width characters, a correct character pattern is generated, and an increase in extra character patterns is suppressed. At this time, the value of Px is determined using a table indicating how Px changes at each likelihood.

  Specifically, in a character line with many character touches, a plurality of touching characters are recognized as one element, the width of the generated element is increased, and the double likelihood is increased. At this time, by using the half-width likelihood Lh and the character contact likelihood Lc, by increasing the value of Px only when the value of (Lh−Lc) is lower than a certain threshold, generation of an extra character pattern It is suppressed. Also, the generation of an extra character pattern is suppressed by increasing the value of Px only when the character blur likelihood Lb is lower than a certain threshold. Also, by increasing the value of Px only when the double-angle likelihood Ld is higher than a predetermined threshold, a correct character pattern is generated in a character line including double-byte characters, and an extra character is included in a character line including double-byte characters. Generation of character patterns is suppressed. Here, if the noise likelihood Ln is higher than a certain threshold value, generating a character pattern excluding elements whose width is smaller than the set value Vp and smaller than the set value Vp generates extra character patterns. Can be suppressed.

  Next, details of the forced disconnection processing 110 will be described.

  FIG. 10 is an explanatory diagram illustrating an example of the full-width forced cutting process according to the embodiment of this invention. In the example shown in FIG. 10, a character pattern in which the characters “open” and “departure” are in contact is input, they are separated, and two character patterns are generated.

  FIG. 12 is a flowchart showing the forced disconnection processing 110 according to the embodiment of this invention. Half-width likelihood 1222, double-angle likelihood 1223, character contact likelihood 1224, character blur likelihood 1225, noise likelihood 1226, and character pattern 1201 are input. The character pattern 1201 is stored in the pattern buffer 1203 and input to the double-angle contact determination 1205, the double-angle eye contact determination 1208, the full-angle contact determination 1211, the full-angle contact determination 1214, the half-angle contact determination 1217, and the half-angle contact determination 1220 one by one. The In the character line information generation 1227, a character pattern 1201 is input, and information such as the height and width of the character line is generated. The generated character line information includes double-angle contact determination 1205, double-angle heel contact determination 1208, and full-width contact. The information is input to the determination 1211, the full-angle heel contact determination 1214, the half-angle heel contact determination 1217, and the half-angle heel contact determination 1220.

  In full-width contact determination 1211, it is determined whether or not a full-size character existing on the left in the character pattern is in contact with the right character. Various methods of the full-width contact determination 1211 are conceivable.

For example, the following method can be adopted.
Predetermining the set value Vnc, using the set value Vnc,
(Formula 8) THn = text line height x Vnc
If Thn is obtained by the above and the width of the input character pattern is larger than the obtained THn, it can be determined that the full-width character is in contact. For example, by setting Vnc of (Equation 8) to 1, it can be determined that there is a possibility that full-width characters are in contact with a pattern larger than the character line height.

  FIG. 11 is an explanatory diagram illustrating an example of full-width rivet forced cutting according to the embodiment of this invention. In the full-width heel contact determination 1214, as in the example shown in FIG. 11, it is determined whether or not a full-size character having only a heel on the left in the character pattern is in contact with the right character. Various methods of the full-angle wrinkle contact determination 1214 are conceivable.

  For example, when THn in (Equation 8) is used and the width of the input character pattern is greater than THn, it can be determined that a double-byte character heel is touching. For example, by setting Vnc of (Equation 8) to 1, it can be determined that a pattern larger than the character line height may be in contact with a double-byte character heel.

  In the half-width contact determination 1217, it is determined whether or not the half-width character existing on the left in the character pattern is in contact with the right character. Various methods of the half-angle contact determination 1217 can be considered.

  For example, using THn in (Equation 8) and the width of the input character pattern is greater than THn, it can be determined that full-width characters are in contact. For example, by setting Vnc in (Equation 8) to 1/2, it can be determined that there is a possibility that a half-width character is in contact with a pattern that is larger than half the character line height.

  In the half-width wrinkle contact determination 1220, it is determined whether or not a half-width character having only a wrinkle on the left in the character pattern is in contact with the right character. Various methods of the half-angle wrinkle contact determination 1220 are conceivable.

  For example, when THn in (Equation 8) is used and the width of the input character pattern is greater than THn, it can be determined that a double-width character heel is touching. For example, by setting Vnc in (Equation 8) to 1/2, it can be determined that there is a possibility that a half-width character is in contact with a pattern that is larger than half the character line height.

  In double angle contact determination 1205, it is determined whether or not the double angle character existing on the left in the character pattern is in contact with the right character. Various methods of the double-angle contact determination 1205 can be considered.

  For example, using THn in (Equation 8), if the width of the input character pattern is greater than THn, it can be determined that double-width characters are in contact. For example, by setting Vnc in (Equation 8) to 2, it can be determined that there is a possibility that a double-width character is in contact with a pattern larger than twice the character line height.

  In double-angle heel contact determination 1208, it is determined whether or not a double-angle character that has only a left heel in the character pattern is in contact with the right character. Various methods of double-angle wrinkle contact determination 1208 are conceivable.

  For example, when THn in (Equation 8) is used and the width of the input character pattern is larger than THn, it can be determined that a double-width character heel is touching. For example, by setting Vnc in (Equation 8) to 2, it can be determined that a pattern larger than twice the character line height may be in contact with a double-width character wrinkle.

  FIG. 13 is an explanatory diagram showing an outline of the forced disconnection process according to the embodiment of this invention.

  In full-width position forced cutting 1212, a character pattern determined to have a possibility that full-width characters are in contact in full-width contact determination 1211 is input, a cutting position of the input character pattern is determined, and the input character pattern is determined. Are cut at the determined cutting position, and two new character patterns are generated. There are various methods for determining the cutting position.

  For example, the following method can be employed. First, a character pattern that is determined to be possibly touched by any one of double-angle contact determination 1205, double-angle eye contact determination 1208, full-angle contact determination 1211, full-angle eye contact determination 1214, half-angle contact determination 1217, and half-angle eye contact determination 1220. 1301 is input, and a black pixel histogram 1302 in which black pixels are projected in the y direction is generated. Next, Zx for determining the cutting position search center is calculated, and the frequency of the black pixel histogram 1302 is searched in the section from Zx−h to Zx + h with the distance of Zx in the positive direction of the x axis from the origin as the cutting position search center. The position with the lowest frequency is set as the cutting position.

Specifically, the set value Vpc is determined in advance, and the set value Vpc is used.
(Formula 9) Zx = text line height x Vpc (Vpc is the set value)
To find Zx. For example, by setting Vpc to 1, the character pattern can be cut from the left end of the character pattern near the same x position as the character line height. That is, the input character pattern 1301 is cut from the left end at a full-width character width position to generate two full-width character width character patterns. However, the character pattern including the position to be cut is not always input. Therefore, cutting is not performed when all the frequencies in the zone from Zx−h to Zx + h are higher than a certain value.

  In full-width heel position forced cutting 1215, a character pattern that is determined to have a possibility that a full-width character heel is in contact in full-width heel contact determination 1214 is input, and the cutting position of the input character pattern is determined and input. The determined character pattern is cut at the determined cutting position, and two new character patterns are generated. Various methods for determining the cutting position are conceivable.

  For example, by adopting the same method as the full-width position forced cutting 1212 and adjusting Vpc to 1, for example, it is possible to generate a character pattern having a width corresponding to a full-width character font.

  In the half-width position forced cutting 1218, the character pattern determined as having a possibility that the half-width character is in contact in the half-width contact determination 1217 is input, the cutting position of the input character pattern is determined, and the input character pattern is determined. Are cut at the determined cutting position, and two new character patterns are generated. Various methods for determining the cutting position are conceivable.

  For example, a character pattern having a width corresponding to a half-width character can be generated by adopting the same method as the full-width position forced cutting 1212 and setting Vpc to 0.5, for example.

  In the half-width wrinkle position forced cutting 1221, the character pattern that is determined to be possibly touched by the half-width wrinkle contact determination 1220 is input, the cutting position of the input character pattern is determined, and the input The determined character pattern is cut at the determined cutting position, and two new character patterns are generated. Various methods for determining the cutting position are conceivable.

  For example, by adopting the same method as full-width position forced cutting 1212 and setting Vpc to 0.5, for example, it is possible to generate a character pattern having a width corresponding to a half-width character font.

  In the double-angle position forced cutting 1206, the character pattern determined to be possibly touched by the double-angle contact determination 1205 is input, the cutting position of the input character pattern is determined, and the input character pattern is determined. Are cut at the determined cutting position, and two new character patterns are generated. Various methods for determining the cutting position are conceivable.

  For example, by adopting the same method as full-width position forced cutting 1212 and setting Vpc to 2, for example, a character pattern having a width corresponding to a double-width character can be generated.

  In the double-angle heel position forced cutting 1209, the character pattern determined to have a possibility that the double-angle character heel is in contact in the double-angle heel contact determination 1208 is input, and the cutting position of the input character pattern is determined and input. The determined character pattern is cut at the determined cutting position, and two new character patterns are generated. Various methods for determining the cutting position are conceivable.

  For example, by adopting the same method as the full-width position forced cutting 1212 and setting Vpc to 2, for example, a character pattern having a width corresponding to a double-width character font can be generated.

  Character patterns generated by full-width position forced cut 1212, full-width saddle position forced cut 1215, half-width position forced cut 1218, half-angle saddle position forced cut 1221, double-angle position forced cut 1206, and double-angle saddle position forced cut 1209 are stored in pattern buffer 1203. It is stored and input again to the double-angle contact determination 1205, the double-angle eyelid contact determination 1208, the full-angle eye contact determination 1211, the full-angle eye contact determination 1214, the half-angle contact determination 1217, and the half-angle eye contact determination 1220.

  This is because full-width position forced cutting 1212, full-width collar position forced cutting 1215, half-angle position forced cutting 1218, half-angle collar position forced cutting 1221, double-angle position forced cutting 1206, and double-angle collar position forced cutting 1209 are input as one character pattern. In order to generate two character patterns, when a character pattern in which three or more characters are in contact with each other is input, one of the two generated character patterns has a plurality of touched character patterns. Because it is included.

  The double-angle contact determination 1205, the double-angle heel contact determination 1208, the full-angle contact determination 1211, the full-angle heel contact determination 1214, the half-angle contact determination 1217, and the half-angle heel contact determination 1220 are performed until all the character patterns in the pattern buffer 1203 are processed.

  In the forced cutting process in the prior art, whether or not the input character pattern is a double-width character, a full-width character, or a half-width character is undefined. Therefore, the double-angle contact determination 1205 and the double-angle heel contact determination 1208 are always performed for all input character patterns. , Full angle contact determination 1211, full angle heel contact determination 1214, half angle contact determination 1217, half angle heel contact determination 1220, full angle position forced cutting 1212, full angle heel position forced cutting 1215, half angle position forced cutting 1218, half angle heel position forced cutting 1221, double angle A position forced cut 1206 and a double angle saddle position forced cut 1209 are executed. For this reason, the processing of cutting character patterns assuming double-width characters, full-width characters, and half-width characters is performed, and many extra character patterns are generated, which takes a lot of processing time.

  In addition, there is a problem that recognition accuracy decreases due to an increase in extra character patterns. For example, when a character pattern including double-width characters is input, the full-width position forced cut 1212, full-width heel position forced cut 1215, half-width position forced cut 1218, and half-width heel position forced cut 1221 are changed to full-width character width and half-width character. By cutting by width, many extra character patterns are generated. The processing time of the character identification 111, the knowledge processing 112, and the recognition result utilization character width test 113 increases due to the increase of the extra character patterns. Furthermore, there is a problem that an extra character pattern is erroneously recognized. In addition, when a character including a half-width character is input, the processing of double-angle contact determination 1205 and double-angle contact determination 1208 is an extra process.

  Therefore, in the embodiment of the present invention, the double-angle contact determination 1205, the double-angle heel contact determination 1208, the full-angle contact determination 1211, the full-angle heel contact determination 1214, the half-angle contact determination 1217, and the half-angle heel contact determination 1220 include the half-angle likelihood 1222, the double-angle likelihood. The degree 1223, the character contact likelihood 1224, the character blur likelihood 1225, and the noise likelihood 1226 are input, and the contact determination is controlled. The forced cutting process can also be controlled by the contact determination control. In this way, by controlling the forced cutting process using the half-angle likelihood 1222, the double-angle likelihood 1223, the character contact likelihood 1224, the character blur likelihood 1225, and the noise likelihood 1226, the generation of extra character patterns is suppressed. Thus, it is possible to suppress the processing time and obtain highly accurate character recognition.

  Specifically, when the double angle likelihood 1223 is higher than the set value Vdc, the half-width contact determination 1217 and the half-angle heel contact determination 1220 are not executed, so that even if a character pattern including double-width characters is input, Cutting the character pattern with the character width can be suppressed.

  In other words, generation of extra character patterns can be suppressed. Further, when the half-angle likelihood 1222 is higher than the set value Vhc1, the processing time can be shortened by not performing the double-angle contact determination 1205 and the double-angle heel contact determination 1208. Further, when the half-width likelihood 1222 is lower than the set value Vhc2 and the character blur likelihood 1225 is higher than the set value Vbc, by preventing the half-width wrinkle contact determination 1220 from being executed, a character pattern including full-width characters can be obtained. Cutting with a half-angle width is suppressed. When the character contact likelihood 1224 is lower than the set value Vcc, the character pattern is cut by executing the double-angle contact determination 1205, the double-angle heel contact determination 1208, the half-angle contact determination 1217, and the half-angle heel contact determination 1220. Can be suppressed, and the processing time can be shortened. When the noise likelihood 1226 is higher than the set value Vnc, the character pattern can be prevented from being cut and the processing time can be shortened by executing the half-angle contact determination 1217 and the half-angle contact determination 1220. .

  Returning to FIG. 12, the description of the flowchart of the forcible disconnection processing 110 will be continued.

  All the character patterns stored in the pattern buffer 1203 are input to the merge pattern 1228.

  The cut character pattern includes a character pattern including only a character deviation or only a character heel. In the merge pattern 1228, a character pattern is generated by combining a character pattern including only the cut bias and a character pattern including only the corresponding heel. In addition, a character pattern is generated by combining a character pattern including only the cut heel and a corresponding biased character pattern. The merge pattern 1228 includes a cut-off pattern merge process and a cut-off pattern merge process, both of which are the same as the pre-segmentation process 109.

  In the cut partial pattern merge process, a character pattern is generated that is combined with a character pattern within a distance of a set value Vx from the left end of the cut character pattern to the right in the horizontal direction. As a result, a character pattern is generated by combining the character pattern including only the bias and the corresponding character pattern of the tile.

  In the cut-off pattern merge process, a character pattern is generated that is combined with a character pattern within a distance of the set value Vx from the right end of the cut character pattern to the left in the left-right direction. As a result, when the cut character pattern includes only 旁, a character pattern in which this character pattern is combined with the corresponding biased character pattern is generated.

  In the prior art, when the value of the setting value Vx is set so that a double-width character pattern is always generated, an extra character pattern is generated when a character pattern including full-width characters or a character pattern including half-width characters is input. There was a problem of doing. In addition, when the value of Vx is set so that half-width and full-width character patterns are generated, there is a problem that character patterns of double-width characters may not be generated.

  In the embodiment of the present invention, half-width likelihood 1222, double-angle likelihood 1223, character contact likelihood 1224, character fading likelihood 1225, and noise likelihood 1226 are input to merge pattern 1228. Generation of patterns is suppressed, and character patterns of half-width characters, full-width characters, and double-width characters are correctly generated. For example, the value of Vx is set to a value that generates a character pattern of half-width characters and full-width characters, and the value of Vx is changed to the character pattern of a double-width character only when the value of double-angle likelihood 1223 is higher than the set value Vdm. Change to a value that generates. As a result, it is possible to correctly generate a character pattern including double-width characters and to suppress generation of an extra character pattern even when a character pattern of half-width characters or full-width characters is input.

  As described above, the character patterns generated by the pre-segmentation processing 109 and the forced cutting processing 110 are retained as a character cut-out network having a connection relationship of the generated character patterns as shown in FIG.

  The above-described series of processing for generating a character segmentation network can be applied to characters in languages other than Japanese, for example, Chinese and Korean (Hangul characters).

  FIG. 16 is an explanatory diagram illustrating an example of a character segmentation network (simplified characters) according to the embodiment of this invention. Needless to say, the same processing can be performed for traditional characters.

  FIG. 17 is an explanatory diagram illustrating an example of a character segmentation network (Hangul character) according to the embodiment of this invention.

  At this time, the cut-out network can be changed by adjusting the threshold value used when calculating each likelihood. This makes it possible to adjust the accuracy and processing speed. For example, in a character line including half-width, full-width, and double-width characters, if the half-width likelihood is large, adjusting the threshold value so as not to perform half-width forced cutting results in fewer cutout networks and faster processing time.

  In the character identification 111, the characteristics of the character patterns generated by the pre-segmentation process 109 and the forced cutting process 110 are extracted, and the similarity with the characteristics of each character type in the identification dictionary prepared in advance is calculated. The similarity with the characteristics of each character type is used as a character identification result.

  In the knowledge processing 112, the character segmentation network 1401 and the character identification result are input, and the character recognition result 1 (1501 in FIG. 15) of the character line image is output. There are various methods for outputting the character recognition result 1 (1501 in FIG. 15).

  For example, the following method can be employed. On the character segmentation network, in the one-way route from left to right, calculate the value with the highest similarity in the identification result of each character pattern, and select the route with the highest average value of the calculated values. (Optimum path search). Then, the character type having the highest similarity in the obtained character pattern identification result on the path is output as the character recognition result 1.

  In addition, a combination of character types with high similarity in the identification result of each character pattern is collated with a word in a word dictionary prepared in advance, and the character type of the character pattern included in the collated word is given priority as a recognition result. There is also a method of selecting a path search so that it is output. This method is called word matching.

  Furthermore, a word list is prepared in advance, and the probability that a word registered in the word list and another word registered in the word list appear in the character line in succession is held as a word bigram. There is also a method of outputting the character recognition result 1 with priority given to a sequence of words that have a high probability of appearing continuously in a character line, with reference to a word bigram when searching for an optimal path. These are described in detail in Non-Patent Document 1. By using word collation and word bigram in this way, even if the character type with the highest similarity in the character pattern identification result is incorrect, it is possible to output other correct character types as character recognition results. . That is, a more correct character recognition result can be obtained.

  However, the character recognition result 1 (1501 in FIG. 15) obtained in this way has a high probability of being correct, but may still include a recognition error.

  On the other hand, in the embodiment of the present invention, the recognition result utilization character width test 113 is performed. First, from the character type of the character pattern included in the character recognition result 1 (1501 in FIG. 15), and the width and height of the character pattern, the character width determination indicating whether the character pattern in the character line is half-width, full-width, or double-width. The result is calculated.

  Next, according to the character width determination result, the character recognition result 1 (1501 in FIG. 15) is used by using the character type of the character pattern in the character recognition result 1 and the width and height of the character pattern. It is determined whether or not an inappropriate recognition result is included.

  If an inappropriate recognition result is included, the character pattern generated by the pre-segmentation process and the forced cutting process using the character width determination result, the character pattern character type, and the character pattern width and height are used. From this, an appropriate character pattern and identification result are selected. Then, knowledge processing is performed using only the selected character pattern and identification result, and a corrected character recognition result is output.

  FIG. 15 is an explanatory diagram showing a recognition result utilization character width test according to the embodiment of this invention.

  First, the character recognition result 1 (1501) is input, and the character pattern in the character line is determined from the character type of the character pattern included in the character recognition result 1 and the width and height of the character pattern in the recognition result use character width estimation processing 1502. A character width determination result indicating whether is a half-width, full-width, or double-width is calculated.

  In the character pattern / identification result selection processing 1504, the character pattern of the character pattern in the character recognition result 1 and the width and height of the character pattern are used according to the character pattern / character width determination result. It is determined whether or not an inappropriate recognition result is included in the character recognition result 1.

  If an inappropriate recognition result is included, it is generated by the pre-segmentation process and the forced cutting process using the character width determination result, the character type of the character pattern, and the width and height of the character pattern. From the selected character pattern, an appropriate character pattern and identification result are selected.

  Then, in knowledge processing 1505, using only the selected character pattern and identification result, processing similar to knowledge processing 112 is performed, and character recognition result 2 (1505) is output.

  Various methods are conceivable for the recognition result use character width estimation processing 1502 and the character pattern / identification result selection processing 1504. Examples are described below.

  For example, in the recognition result utilization character width estimation processing 1502, the following processing is performed.

A character pattern of a specific character type is extracted from the character recognition result 1 (1501), and the width and height of the extracted character pattern are used to calculate (Equation 10) Rwh = width / height. Next, the set value Vrwh is determined in advance, and using the set value Vrwh,
(Condition 5) Rwh> Vrwh
The number of character patterns Nrwh that satisfies the above is calculated. Furthermore, using Nrwh and the number of character patterns of the character identification result,
(Formula 11) Pwh = Nrwh / (number of character patterns of character identification result)
Is calculated. A set value THrwh is determined in advance, and the character width is estimated using Pwh and the set value THrwh.

Specifically, for example, a character pattern (kanji result character pattern) whose character type is Kanji is extracted from the character recognition result 1 (1501), Rwh is calculated using (Equation 10), and the set value Vrwh Is 1.1 and the number Nrwh of character patterns (wide kanji result character patterns) satisfying (conditional expression 5) is calculated. Furthermore, Pwh is calculated using (Equation 11). At this time, THrwh is set to 0.6, and the relationship between Pwh and THrwh is
(Condition 6) Pwh <THrwh
If the condition is satisfied, it is determined that the characters in the character line have a half-width or double-width character width.

In the character pattern / identification result selection process 1504, the character recognition result 1 (1501), the character pattern generated in the pre-segmentation process 109, and the character pattern generated in the forced cutting process 110 are input, and the character line height H Is estimated in advance, and the set value Vv is set in advance. Using the set value Vv, the character recognition result 1 (1501) includes:
(Condition 7) (Width of character pattern)> H × Vv
It is determined whether or not there is a character pattern that satisfies. The height of the character line is estimated by the maximum value of the character pattern height.

  When there is a corresponding character pattern, a character pattern (non-wide character pattern) that satisfies the contradiction of (conditional expression 7) from the character pattern generated by the pre-segmentation processing 109 and the character pattern generated by the forced cutting processing 110. Only) is selected. In the knowledge processing 1505, only the non-wide character pattern is input, and knowledge processing is performed to determine the character recognition result 2.

  Thus, not only the distribution of the size of simple components, but also the character width estimation considering the size according to the character type of the character recognition result obtained by character identification and knowledge processing, the character pattern is selected, By performing knowledge processing again using them, a more accurate character recognition result can be obtained.

  The knowledge processing is also possible for characters in other languages, such as Chinese (simplified and traditional) and Korean (Hangul).

It is a flowchart which shows the whole character recognition process of embodiment of this invention. It is a block diagram which shows the structure of the character recognition apparatus of embodiment of this invention. It is explanatory drawing which shows an example of the result of the character line extraction process of embodiment of this invention. It is a figure which shows the example of the connection component of embodiment of this invention. It is a figure which shows the example of the element of embodiment of this invention. It is explanatory drawing which shows an example of the pre-segmentation process of embodiment of this invention. It is explanatory drawing which shows an example of the pre-segmentation process of the full-width character of embodiment of this invention. It is explanatory drawing which shows an example of the pre-segmentation process of the double-width character of embodiment of this invention. It is explanatory drawing which shows an example of the pre-segmentation process of the half-width character of embodiment of this invention. It is explanatory drawing which shows an example of the full-width forced cutting process of embodiment of this invention. It is explanatory drawing which shows an example of the full width wrinkle forced cutting process of embodiment of this invention. It is a flowchart which shows the forced cutting process of embodiment of this invention. It is explanatory drawing which shows the outline | summary of the forced cutting process of embodiment of this invention. It is explanatory drawing which shows an example of the character segmentation network of embodiment of this invention. It is explanatory drawing which shows the recognition result utilization character width test of embodiment of this invention. It is explanatory drawing which shows an example of the character segmentation network (simplified character) of embodiment of this invention. It is explanatory drawing which shows an example of the character segmentation network (Hangul character) of embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Input image 102 Character line extraction process 103 Connected component / element generation process 104 Double angle discrimination process 105 Character contact discrimination process 106 Half angle discrimination process 107 Character blurring discrimination process 108 Noise discrimination process 109 Pre-segmentation process 110 Forced cutting process 111 Character identification process 112 Knowledge processing 113 Recognition result use character width verification process 114 Character recognition result

Claims (13)

A character recognition method for recognizing characters from an image,
A connected component generation process for generating a connected component that is a series of black pixels constituting a character from an image including the character;
A character line extraction process for extracting a character line that is an area where characters are arranged;
Half-width discrimination processing for calculating a half-width likelihood indicating the degree of half-width character of a character in a character line;
A double angle discrimination process for calculating a double angle likelihood indicating the degree of doubleness of the characters in the character line;
Character blur determination processing for calculating a character blur likelihood indicating the blurring degree of characters in a character line,
A character contact determination process for calculating a character contact likelihood indicating a contact degree of a character in a character line;
A noise discrimination process for calculating a noise likelihood indicating the degree of noise mixed in the character line;
Pre-segmentation for generating a character pattern that is information combining one or a plurality of connected components using the character contact likelihood calculated by the character contact determination processing and the double-angle character likelihood calculated by the double-angle determination processing Processing,
Using the character blur likelihood calculated by the character blur determination process and the half-width character likelihood calculated by the half-width discrimination process, a character pattern for cutting a connected component is selected, and a position at which the connected component is cut is determined. Forced disconnection processing,
A character identification process for extracting the characteristics of the character pattern generated by the pre-segmentation process and the forced cutting process and calculating the similarity to the characteristics of each character type existing in the identification dictionary prepared in advance. Character recognition method. The character recognition method generates an element in which connected components overlapping in the vertical direction are combined,
The character recognition method according to claim 1, wherein the half-width discrimination processing calculates a half-width likelihood using the number of elements, the width of the elements, and the height of the character line. The character recognition method generates an element in which connected components overlapping in the vertical direction are combined,
2. The character recognition according to claim 1, wherein the double-angle discrimination processing calculates a double-angle likelihood using a width of an element satisfying a predetermined condition of a width of the element and a height of the character line. Method. The character recognition method generates an element in which connected components overlapping in the vertical direction are combined,
In the character blur determination process, the character blur likelihood is calculated using the number of connected components in which the width and height of the connected component satisfy a predetermined condition, the height of the character line, and the width of the element. The character recognition method according to claim 1. The character recognition method generates an element in which connected components overlapping in the vertical direction are combined,
In the character contact determination process, the character contact likelihood is calculated using the number of elements satisfying a predetermined width of the element, the height of the character line, and the width of the element. The character recognition method according to claim 1. The character recognition method generates an element in which connected components overlapping in the vertical direction are combined,
The noise determination process calculates a noise likelihood using the number of connected components in which the width and height of the connected components satisfy a predetermined condition, the height of the character line, and the width of the element. The character recognition method according to claim 1, wherein:   The pre-segmentation process is performed by calculating the calculated half-angle likelihood, the calculated double-angle likelihood, the calculated character blur likelihood, the calculated character contact likelihood, and the calculated noise likelihood. The character pattern having one or a plurality of widths is generated by adjusting a width of a character pattern to be generated using one or a plurality of likelihoods among the two likelihoods. Character recognition method.   The forced cutting process is performed by calculating the calculated half angle likelihood, the calculated double angle likelihood, the calculated character fading likelihood, the calculated character contact likelihood, and the calculated noise likelihood. The character recognition method according to claim 1, wherein the width of the element to be cut and the position for forced cutting are adjusted using one or a plurality of likelihoods among the two likelihoods. A character recognition device comprising a processor, a memory connected to the processor, and an interface connected to the processor, for recognizing characters from an input image,
A connected component generation unit that generates a connected component that is a series of black pixels that form a character from the input image;
A character line extraction unit that extracts a character line that is an area in which characters are arranged;
A half-width discriminator that calculates a half-width likelihood indicating the degree of half-width character of a character in a character line;
A double angle discriminating unit that calculates a double angle likelihood indicating the degree of doubleness likelihood of the characters in the character line;
A character blur determining unit for calculating a likelihood of blur of a character indicating a blurring degree of a character in a character line;
A character contact determination unit for calculating a character contact likelihood indicating a contact degree of a character in a character line;
A noise discriminating unit for calculating a noise likelihood indicating the degree of noise mixing in the character line;
Pre-segmentation for generating a character pattern that is information combining one or a plurality of connected components using the character contact likelihood calculated by the character contact determination unit and the double-angle character likelihood calculated by the double-angle determination unit Part,
Using the character blur likelihood calculated by the character blur determination process and the half-width character likelihood calculated by the half-width discrimination process, a character pattern for cutting a connected component is selected, and a position at which the connected component is cut is determined. Forced cutting part,
A character identification unit that extracts features of the character pattern generated by the pre-segmentation unit and the forced cutting unit and calculates a similarity with the feature of each character type existing in an identification dictionary prepared in advance. Character recognition device. The character identification unit
A word collation unit that collates with a combination of character types with high similarity and a word list in the word dictionary prepared in advance;
A knowledge processing unit that determines a character recognition result 1 in a character line by using the similarity for each character type as a result of the character identification processing and the word matching result;
Character width determination result indicating whether the character pattern in the character line is half-width, full-width, or double-width from the character type of the character pattern, the width and height of the character pattern, and the number of character patterns included in the determined character recognition result 1 And an inappropriate recognition result is included in the character recognition result 1 using the character type of the character pattern in the character recognition result 1 and the width and height of the character pattern according to the character width determination result. If an inappropriate recognition result is included, the pre-segmentation process is performed using the character width determination result, the character type of the character pattern, and the width and height of the character pattern. A recognition result-use character width verification unit that selects an appropriate character pattern and identification result from the character pattern generated by the forced cutting process and performs knowledge processing using only the selected character pattern and identification result; With features Character recognition apparatus according to claim 9 that.   The pre-segmentation unit is configured to generate the calculated half-angle likelihood, the calculated double-angle likelihood, the calculated character blur likelihood, the calculated character contact likelihood, and the calculated noise likelihood generation. The character pattern having one or a plurality of widths is generated by adjusting a width of a character pattern to be generated using one or a plurality of likelihoods among the five likelihoods. The character recognition device described.   The forced cutting unit is configured to generate the calculated half-angle likelihood, the calculated double-angle likelihood, the calculated character blur likelihood, the calculated character contact likelihood, and the calculated noise likelihood generation. The character recognition device according to claim 9, wherein the width of the element to be cut and the position for forced cutting are adjusted using one or a plurality of likelihoods among the five likelihoods. A character recognition device comprising a processor, a memory connected to the processor, and an interface connected to the processor, for recognizing characters from an input image,
A connected component generation unit that generates a connected component that is a series of black pixels that form a character from the input image;
A character line extraction process for extracting a character line that is an area where characters are arranged;
A pre-segmentation unit that generates a character pattern that is information obtained by combining one or a plurality of connected components;
Select a character pattern that cuts the connected component, and a forced cutting unit that determines the cutting position of the connected component;
A character identification unit that extracts the characteristics of the character pattern generated by the pre-segmentation unit and the forced cutting unit, and calculates a similarity to the characteristics of each character type existing in an identification dictionary prepared in advance;
A word collation processing unit that collates with a combination of character types having high similarity in the character identification process and a word list in the word dictionary prepared in advance;
A knowledge processing unit for determining a character recognition result 1 in a character line using a similarity for each character type and a word matching result as a result of the character identification process;
From the character type of the character pattern included in the character recognition result 1, the width and height of the character pattern, and the number of character patterns, a character width determination result representing whether the character pattern in the character line is half-width, full-width, or double-width is calculated. Whether or not an inappropriate recognition result is included in the character recognition result 1 using the character type of the character pattern in the character recognition result 1 and the width and height of the character pattern according to the character width determination result. If an inappropriate recognition result is included, the character segment determination result, the character type of the character pattern, and the width and height of the character pattern are used. A recognition result-use character width verification process for selecting an appropriate character pattern and an identification result from the generated character pattern and performing knowledge processing using only the selected character pattern and the identification result. Character recognition Apparatus.

Images