JP2013186906A - Method and device for recognizing character string in image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and device for recognizing a character string in an image.SOLUTION: The method includes the steps of: extracting a character string area in an image; performing over-segmentation on the character string area; and recognizing a character string included in the character string area by path searching strategy on the basis of at least one feature of a language type context feature and a character width context feature.

Description

  The present invention relates to the field of character recognition, and more particularly, to a method and apparatus for recognizing a character string in an image.

  With the spread of digital image acquisition devices (for example, mobile phones, cameras, etc.), image search systems based on text information have attracted widespread attention. Of these, the recognition of text in natural scene images is a major component of the overall system and greatly affects system performance. However, under the influence of text size, font and image quality degradation, text recognition is still difficult to achieve highly accurate recognition results. In addition, since a natural scene image usually includes a plurality of kinds of languages, this also greatly affects the accuracy of text recognition.

  Therefore, a technique that can solve the above-described problem is desirable.

  The main object of the present invention is to provide a method and apparatus for recognizing a character string in an image.

  According to one aspect of the present invention, a method for recognizing a character string in an image is provided. The method includes a step of extracting a character string region in an image, a step of dividing the character string region, and a path search based on at least one of a language type context feature and a character width context feature. Recognizing a character string included in the character string area by a strategy (Path Searching Strategy).

4 is a flowchart of a method for recognizing a character string in an image according to an embodiment of the present invention. FIG. 4 is a diagram illustrating an exemplary character string area according to the present invention. It is a figure which shows the character string area | region after the exemplary pre-processing by this invention. FIG. 6 is a diagram illustrating an image of a character string region after an exemplary over-segmentation according to the present invention. FIG. 4 is a diagram illustrating an exemplary optimum route search according to the present invention. 4 is a flowchart for determining a language type context feature based on adjacent characters according to an embodiment of the present invention. 7 is a flowchart for determining language type context features based on adjacent characters according to another embodiment of the present invention. FIG. 4 is a diagram illustrating an exemplary character string area according to the present invention. FIG. 6 is a diagram illustrating a recognition result obtained when an exemplary language type context feature is not introduced according to the present invention. FIG. 6 is a diagram illustrating recognition results obtained when introducing exemplary language type context features according to the present invention. 6 is a flowchart for determining character width context features based on a difference between a character width and a weighted average character width, according to an embodiment of the present invention. FIG. 4 is a diagram illustrating an exemplary character string area according to the present invention. FIG. 6 is a diagram illustrating an exemplary character string recognition result when no character width context feature is introduced according to the present invention. FIG. 6 is a diagram illustrating a character string recognition result when an exemplary character width context feature is introduced according to the present invention. 6 is an exemplary flowchart for recognizing a character string in an image according to the present invention. 1 is a block diagram of an apparatus for recognizing a character string in an image according to an embodiment of the present invention. FIG. 6 is a block diagram of an apparatus for recognizing a character string in an image according to another embodiment of the present invention. FIG. 6 is a block diagram of an apparatus for recognizing a character string in an image according to another embodiment of the present invention. FIG. 2 is an exemplary structural diagram of a computing device that can be used to implement a method and apparatus for recognizing a character string in an image according to an embodiment of the present invention.

  Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

  A process 100 of a method for recognizing a character string in an image according to an embodiment of the present invention will be described with reference to FIGS. 1, 2A to 2C, and FIG.

  As shown in FIG. 1, in step S105, a character string region in the image can be extracted. FIG. 2A is a diagram showing an exemplary character string region of the present invention. As shown in FIG. 2A, a character string region in the image, that is, a character string region including “Fulfillment” can be extracted.

  Optionally, pre-processing can be performed on the character string region prior to over-segmentation of the character string region in step S110 described below. FIG. 2B is a diagram showing an exemplary pre-processed character string area of the present invention. Among them, the pre-processing includes several basic image processing processes such as binarization, image smoothing, inclination removal, and communication area extraction. The main purpose of these preprocessing is to provide reliable image quality for subsequent character segmentation, extraction and recognition. It should be understood that preprocessing is a step that is not necessarily required. For example, if the character string region in the image is very clear, these pre-processing steps need not be performed.

  In step S110, over segmentation can be performed on the character string region extracted in step S105. When the difference in the size, font, and arrangement method of the character string in the scene image is relatively large, it is possible to ensure that the characters are not stuck as much as possible by adopting an over-segmentation strategy. Based on the result of the over-segmentation, the division candidate grid can be configured to be convenient for the subsequent optimum route search, that is, the character string recognition. FIG. 2C is a diagram illustrating an image of a string region after exemplary over segmentation according to the present invention. As shown in FIG. 2C, over-segmentation is performed on the character string region including “Fulfillment”.

  Subsequently, in step S115, the character string included in the character string region can be recognized by the route search strategy based on at least one of the language type context feature and the character width context feature. FIG. 3 is a diagram illustrating an exemplary optimum route search according to the present invention. FIG. 3 shows various search paths.

For general character string recognition, the optimum route search is a common method for performing division and character string recognition simultaneously. In the division candidate grid, any one route corresponds to one character string recognition result, and the purpose of the optimum route search is to find a route corresponding to the optimization result of the route target function, and a recognition result corresponding to this route. Is determined to be closest to the true result. Given a character string sequence X, a character type code sequence Y, and a corresponding split path S, the path target function can appear as follows:

は、特徴関数を表し、
（外２）

は、関数の重み（weight）を表す。 here,
(Outside 1)

Represents the feature function,
(Outside 2)

Represents the weight of the function.

  In this embodiment, the feature function includes at least one of a language type context feature and a character width context feature.

は、単字認識器特徴、語義文脈特徴及び幾何学的文脈特徴のうちの１つ又は複数の特徴を更に含んでもよい。これについては、図9を参照して後述する。 Optional feature function (outside 3)

May further include one or more of a single character recognizer feature, a semantic context feature, and a geometric context feature. This will be described later with reference to FIG.

  Usually, it is difficult to obtain a satisfactory recognition result based on single character recognizer features, semantic context features, and geometric context features due to the diversity of characters in natural scene images. In the technical solution according to the present embodiment, the language type context feature and the character width context feature are taken into consideration, so that the recognition accuracy can be improved. Next, the processing for determining the language type context feature will be described with reference to FIGS. 4, 5 and 6A to 6C, and the character width context feature will be determined with reference to FIGS. 7 and 8A to 8C. Processing to be performed will be described.

  First, with reference to FIG. 4, FIG. 5 and FIGS. 6A to 6C, a process for determining a language type context feature based on adjacent characters according to an embodiment of the present invention will be described.

  Natural scene images typically include different types of languages, such as kanji, English, and Arabic numerals. In addition, characters in the same character string usually belong to one type of language. Based on this, according to the present embodiment, the route search function can be calculated based on the language type context feature.

  As shown in FIG. 4, after performing over-segmentation on the character string area in step S110, in step S405, it is determined whether each character in the path and its one adjacent character belong to the same type of language. It can be confirmed. In step S410, the language type context feature can be calculated based on the above-described determination result.

及び符号
（外５）

が与えられるとすると、言語種類文脈特徴関数は、次のように定義され得る。

Specifically, two adjacent characters (outside 4)

And sign (outside 5)

Is given, the language type context feature function can be defined as:

Here, α _{is a} penalty coefficient, and its value may be determined by experience. A language type context feature can achieve the purpose of forcing a character on the same line to belong to the same type of language by penalizing only two adjacent characters belonging to different types of languages. .

  Subsequently, in step S115, the character string included in the character string area can be recognized by the route search strategy based on the language type context feature.

  Next, referring to FIG. 5, a process for determining a language type context feature based on adjacent characters according to another embodiment of the present invention will be described.

  As shown in FIG. 5, after over-segmentation is performed on the character string area in step S110, in step S505, each character in the search path and each adjacent character in the plurality of adjacent characters are of the same type. It can be determined whether or not it belongs to a certain language. In step S510, a language type context feature can be calculated based on the above-described determination result.

であるとすると、言語種類文脈特徴関数は、次のように定義され得る。

Specifically, the number of characters adjacent to the target character may be plural. The set of adjacent characters of target character X ₁ is (Outside 6)

, The language type context feature function can be defined as:

The penalty coefficient _α is preferably obtained by a machine learning algorithm using a training sample. For specific algorithms, for example, “Xiang-Dong Zhou, Jin-Lun Yu, Cheng-Lin Liu, Nagasaki, T., Marukawa, K .: Online Handwritten Japanese Character String Recognition Incorporating Geometric Context. ICDAR 2009.7: 48-52 ”can be referred to.

  Subsequently, in step S115, the character string included in the character string area can be recognized by the route search strategy based on the language type context feature.

  FIG. 6A is a diagram illustrating an exemplary character string region according to the present invention, and FIG. 6B is a diagram illustrating a recognition result obtained when an exemplary language type context feature according to the present invention is not introduced. FIG. 6C is a diagram showing recognition results obtained when introducing exemplary language type context features according to the present invention. As shown in FIG. 6B, the result obtained when the language type context feature is not introduced is “o 1 o −6 7 5 o 2 2 2 9”, of which three “0” are mistakenly “ o ”. As shown in FIG. 6C, the result obtained when the language type context feature is introduced is “0 1 o −6 7 5 0 2 2 2 9”, of which one of the three “0” s. Is mistakenly recognized as “o”. As can be seen, the accuracy of character recognition can be improved by introducing language type context features.

  Next, a process for determining a character width context feature based on a difference between a character width and a weighted average character width according to an embodiment of the present invention will be described with reference to FIGS. 7 and 8A to 8C. . Among them, the character width context feature is determined by a method described later for each search path. In the character string recognition process, misrecognition due to erroneous division is relatively common. In addition, even if the writing method of characters is different, the characters on the same line usually have the same character width in order to be convenient for layout. According to the present embodiment, the above-described error can be corrected based on the character width context feature.

  As shown in FIG. 7, after over-segmentation is performed on the character string region in step S110, initial recognition can be performed on the character string region in step S705.

であり、そのうち、
（外8）

は、初期認識結果中の第i個文字の文字幅であり、
（外9）

は、
（外10）

の文字認識信頼度である。即ち、文字認識信頼度
（外11）

を重み係数の値とする。例えば、目標文字の左側にある第一個の文字の右境界から、該目標文字の右側にある第一個の文字の左境界までの距離（図8Aに示すように）を、該目標文字の文字幅
（外12）

とすることができる。理解すべきは、これは、文字幅の１つの例だけである。例えば、文字幅は、文字自身の幅、文字の両側にある隙間の幅と該文字自身の幅との和、文字の両側にある隙間の幅の半分と該文字自身の幅との和、文字自身の幅と該文字の右側にある隙間の幅との和、及び、文字自身の幅と該文字の左側にある隙間の幅との和、のうちの何れか１つであってもよい。 In step S710, the weighted average character width can be estimated based on the initial recognition result, that is, the weighted average character width (outside 7).

Of which
(Outside 8)

Is the character width of the i-th character in the initial recognition result,
(Outside 9)

Is
(Outside 10)

Is the character recognition reliability. That is, character recognition reliability (Outside 11)

Is the value of the weighting factor. For example, the distance (as shown in FIG. 8A) from the right boundary of the first character on the left side of the target character to the left boundary of the first character on the right side of the target character is Character width (outside 12)

It can be. It should be understood that this is only one example of character width. For example, the character width is the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the sum of the half of the width of the gap on both sides of the character and the width of the character itself, It may be any one of the sum of its own width and the width of the gap on the right side of the character and the sum of the width of the character itself and the width of the gap on the left side of the character.

  In step S715, the difference between the character width of each character in the search path and the weighted average character width can be determined.

For example, the following new features may be defined by combining the character width context feature and the original feature.

  This definition shows that if the character width of the target character and the average character width of the text line are close, a relatively small penalty coefficient can be obtained and, on the other hand, a relatively large penalty coefficient can be obtained.

  In step S720, a character width context feature can be calculated based on the difference described above.

  Subsequently, in step S115, the character string included in the character string region can be recognized by the route search strategy based on the character width context feature.

  FIG. 8A is a diagram illustrating an exemplary character string region according to the present invention, and FIG. 8B is a diagram illustrating an exemplary character string recognition result when no character width context feature is introduced according to the present invention. FIG. 8C is a diagram illustrating a character string recognition result when an exemplary character width context feature according to the present invention is introduced. As shown in FIG. 8B, when the character width context feature is not introduced, the recognition result is “Fulfillment”. As shown in FIG. 8C, when the character width context feature is introduced, the recognition result is “Fulfillment”. As can be seen, the accuracy of character recognition can be improved by introducing character width context features.

  Next, a case where a character type context feature is introduced and a case where a character width context feature is introduced will be described. Actually, the character type context feature and the character width context feature can be simultaneously introduced by the above-described equation (1). Other features can also be introduced, such as single character recognizer features, semantic context features and geometric context features.

  Next, an exemplary process for recognizing a character string in an image according to the present invention will be described with reference to FIG.

  As shown in FIG. 9, in addition to the language type context feature and the character width context feature described above, a single character recognizer feature, a semantic context feature, and a geometric context feature are simultaneously introduced.

  It should be understood that not all of these features need be introduced at the same time, and one or more of them may be introduced.

  Next, an apparatus 1000 for recognizing a character string in an image according to an embodiment of the present invention will be described with reference to FIG.

  As shown in FIG. 10, an apparatus 1000 that recognizes a character string in an image may include an extraction unit 1005, a division unit 1010, and a recognition unit 1015.

  The extraction unit 1005 can extract a character string region in the image. The dividing unit 1010 can perform over-segmentation on the character string area. The recognition unit 1015 can recognize a character string included in the character string region by a route search strategy based on at least one of the language type context feature and the character width context feature.

  Next, an apparatus 1000 'for recognizing a character string in an image according to another embodiment of the present invention will be described with reference to FIG. The difference between the device 1000 ′ for recognizing the character string in the image as shown in FIG. 11 and the device 1000 for recognizing the character string in the image as shown in FIG. It includes a unit 1015-1 and a first calculation subunit 1015-2.

  The language type determination subunit 1015-1 can determine whether each character in the path and its one adjacent character belong to the same type of language. Alternatively, the language type determination subunit 1015-1 can determine whether each character in the path and each adjacent character in the plurality of adjacent characters belong to the same type of language.

  The first calculation subunit 1015-2 can calculate the language type context feature based on the above-described determination result.

  Next, an apparatus 1000 "for recognizing a character string in an image according to another embodiment of the present invention will be described with reference to FIG. 12. An apparatus 1000 for recognizing a character string in an image as shown in FIG. ”And the apparatus 1000 for recognizing the character string in the image as shown in FIG. 10 are that the recognition unit 1015 has an initial recognition subunit 1015-3, an average character width estimation subunit 1015-4, and a difference confirmation. It includes a subunit 1015-5 and a second calculation subunit 1015-6.

であり、そのうち、
（外14）

は、初期認識結果中の第i個文字の文字幅であり、
（外15）

は、
（外16）

の文字認識信頼度である。差異確定サブユニット1015-5は、経路中の各文字の文字幅と、重み付き平均文字幅との間の差を確定することができる。第二計算サブユニット1015-6は、この差に基づいて文字幅文脈特徴を算出することができる。 The initial recognition subunit 1015-3 can perform initial recognition on the character string area. The average character width estimation subunit 1015-4 can estimate the weighted average character width based on the initial recognition result, that is, the weighted average character width (outside 13).

Of which
(Outside 14)

Is the character width of the i-th character in the initial recognition result,
(Outside 15)

Is
(Outside 16)

Is the character recognition reliability. The difference determination subunit 1015-5 can determine the difference between the character width of each character in the path and the weighted average character width. The second calculation subunit 1015-6 can calculate a character width context feature based on this difference.

  Among them, the character width is the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the sum of the half of the width of the gap on both sides of the character and the width of the character itself, One of the width of the character itself and the width of the gap on the right side of the character, and the sum of the width of the character itself and the width of the gap on the left side of the character.

  The devices 1000, 1000 'and 1000' 'for recognizing character strings in an image may include a preprocessing unit (not shown), and the preprocessing unit can perform preprocessing on a character string region.

  Optionally, the recognition unit 1015 may include at least one of a language type context feature and a character width context feature and at least one of a single character recognizer feature, a semantic context feature, and a geometric context feature. Based on this, the character string included in the character string region can be recognized by the route search strategy.

  According to the embodiment of the present invention, the recognition accuracy of the character string in the image of the natural scene can be improved by using the language type context feature and the character width context feature.

  Among them, the language type context feature reinforces the rational constraint that “characters belong to the same type in the same text line”. The character width context feature reinforces the rational constraint that “character width is close in the same text line”. According to an embodiment of the present invention, two types of new features (ie, a language type context feature and a character width context feature) and a conventional feature are combined into one target optimization under the mechanism of “split-recognition”. It is possible to recognize a character string in an image of a natural scene based on the feature function. Also, according to the experimental results, these two kinds of new features provided by the present invention can be useful for significantly improving recognition accuracy when recognizing a character string in an image of a natural scene.

  Each step, component unit, and the like in the method and apparatus for recognizing a character string in an image according to the above-described embodiment may be realized by a method of software, firmware, hardware, or any combination thereof. When realized by software or firmware, a program constituting the software or firmware can be installed from a storage medium or a network to a device having a dedicated hardware structure (for example, the general-purpose device 1300 shown in FIG. 13). This apparatus can perform various functions of the above-described constituent units and steps when various programs are installed.

  FIG. 13 is an exemplary structural diagram of a computing device that can be used to implement a method and apparatus for recognizing character strings in an image according to an embodiment of the present invention.

  In FIG. 13, the central processing unit (CPU) 1301 performs various processes based on a program stored in the ROM 1302 or a program loaded from the storage unit 1308 to the RAM 1303. The RAM 1303 stores data necessary when the CPU 1301 executes various processes as necessary. The CPU 1301, ROM 1302, and RAM 1303 are connected to each other via a bus 1304. An input / output interface 1305 is also connected to the bus 1304.

  Also connected to the input / output interface 1305 are an input unit 1306 (including a keyboard and a mouse), an output unit 1307 (including a display such as a CRT and an LCD, a speaker, etc.), and a storage unit 1308 (including And a communication unit 1309 (including a network access card such as a LAN card and a modem). The communication unit 1309 performs communication processing via a network, for example, the Internet. If desired, the drive 1310 can also be connected to the input / output interface 1305. A removable medium 1311, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor storage device, etc., may also be attached to the drive 1310 as necessary, and the computer program read out from it may be It can be installed in the storage unit 1308.

  When the above-described series of processing is realized by software, a program constituting the software may be installed from a network, for example, the Internet, or a storage medium, for example, a removable medium medium 1311.

  It should be understood by those skilled in the art that such a storage medium has a program stored therein and is a removable medium as shown in FIG. 13 that is distributed independently of the apparatus so as to provide the program to the user. It is not limited to 1311. Examples of the removable medium 1311 include a magnetic disk (including a floppy (registered trademark) disk), an optical disk (including a CD-ROM and a DVD), a magneto-optical disk (including MD (registered trademark)), and a semiconductor memory. including. Alternatively, the storage medium may be a ROM 1302, a hard disk included in the storage unit 1308, or the like, in which a program is stored and distributed to a user together with a device including them.

  The present disclosure also relates to a program product comprising machine (eg, computer) readable instruction code. When this instruction code is read and executed by a machine, the method according to the above-described embodiment can be executed. Accordingly, the present disclosure also includes a storage medium that stores a program product comprising the machine-readable instruction code described above. Such storage media include, but are not limited to, magnetic disks (floppy disks), optical disks, magneto-optical disks, memory cards, memory memory sticks, and the like.

  Also, elements and features described in one drawing or embodiment of the present disclosure may be combined with elements and features shown in one or more other drawings or embodiments.

  In addition, the steps of performing the above-described series of processing may be performed in time order according to the order described above, but are not necessarily performed in time order. Some steps may be performed in parallel or independently of each other.

  It is also apparent that each processing process of the above-described method according to the present disclosure can be realized by a computer-executable program stored in various machine-readable storage media.

  The object of the present disclosure may be realized by the following method. That is, a storage medium storing the above-described executable program code is provided directly or indirectly to the system or apparatus, and the computer or CPU in the system or apparatus reads out and executes the above-described program code To do.

  At this time, as long as the system or apparatus has a function of executing the program, the embodiment of the present invention is not limited to the program, and the program may be in an arbitrary format, for example, an object program, an interpreter. It may be an executable program or a script program for the operating system.

  The machine-readable storage medium described above includes various storage devices and storage units, semiconductor devices, magnetic disk units such as optical, magnetic and magneto-optical disks, and other media suitable for information storage. It is not limited.

  In addition, the client computer connects to a corresponding server via the Internet, downloads and installs the computer program code according to the present invention to the computer, and then executes the program to realize the present invention. You can also

  Lastly, it should be explained that in the text, for example, terms such as “first” and “second” are only used to distinguish one entity or operation from another. And has no meaning or suggestion that there is such an actual relationship or order between these entities or operations. Also, the word “comprising”, “having” or other variations are used to encompass non-exclusive “comprising”, whereby a process, method, article or device comprising a series of elements May include other elements not specified, or may include unique elements possessed by the process, method, article or device. In the absence of more limitations, an element defined by the phrase “including” does not exclude other identical elements present in the process, method, article, or apparatus containing the element.

  Moreover, the following additional remarks are disclosed regarding the embodiment including each of the above-described examples.

(Appendix 1)
A method for recognizing a character string in an image,
Extracting a character string region in the image;
Over-segmenting the string region;
Recognizing a character string included in the character string region by a route search strategy based on at least one of a language type context feature and a character width context feature.

(Appendix 2)
The method according to appendix 1, wherein
The language type context feature is:
Determining whether each character in the search path and its one adjacent character belong to the same type of language; and
The method is determined by calculating the language type context feature based on the determination result.

(Appendix 3)
The method according to appendix 1, wherein
The language type context feature is:
Determining whether each character in the search path and each adjacent character in the plurality of adjacent characters belong to the same type of language; and
The method is determined by calculating the language type context feature based on the determination result.

ここで、
（外17）

は、前記初期認識結果中の第i個文字の文字幅であり、
（外18）

は、
（外19）

の信頼度である、方法。 (Appendix 4)
The method according to appendix 1, wherein
The character width context feature is:
Initial recognition is performed on the character string area,
Based on the initial recognition result, the weighted average character width is estimated by the following formula 5,
Determining the difference between the character width of each character in the search path and the weighted average character width; and
Determined by calculating the character width context feature based on the difference;

here,
(Outside 17)

Is the character width of the i-th character in the initial recognition result,
(Outside 18)

Is
(Outside 19)

Is the confidence of the method.

(Appendix 5)
The method according to appendix 4, wherein
The character width is the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the character itself And the width of the gap on the right side of the character and the sum of the width of the character itself and the width of the gap on the left side of the character.

(Appendix 6)
The method according to any one of appendices 1 to 5, wherein
A method further comprising pre-processing the character string area before over-segmenting the character string area.

(Appendix 7)
The method according to any one of appendices 1 to 5, wherein
Based on at least one of the language type context feature and the character width context feature, the step of recognizing a character string included in the character string region by a route search strategy includes:
The path search based on at least one of the language type context feature and the character width context feature and at least one feature of a single character recognizer feature, semantic context feature, and geometric context feature. Recognizing a character string included in the character string region by a strategy.

(Appendix 8)
A device for recognizing a character string in an image,
An extraction unit for extracting a character string region in the image;
A division unit that performs over-segmentation on the character string region;
A recognition unit for recognizing a character string included in the character string region by a route search strategy based on at least one of a language type context feature and a character width context feature.

(Appendix 9)
The apparatus according to appendix 8, wherein
The recognition unit is
A language type determination subunit for determining whether each character in the search path and its one adjacent character belong to the same type of language;
A first calculation subunit that calculates the language type context feature based on the determination result.

(Appendix 10)
The apparatus according to appendix 8, wherein
The recognition unit is
A language type determination subunit for determining whether each character in the search path and each adjacent character in the plurality of adjacent characters belong to the same type of language;
A first calculation subunit that calculates the language type context feature based on the determination result.

ここで、
（外20）

は、前記初期認識結果中の第i個文字の文字幅であり、
（外21）

は、
（外22）

の信頼度である、装置。 (Appendix 11)
The apparatus according to appendix 8, wherein
The recognition unit is
An initial recognition subunit for performing initial recognition on the character string area;
Based on the initial recognition result, an average character width estimation subunit that estimates a weighted average character width according to the following equation 6,
A difference determination subunit for determining a difference between a character width of each character in the search path and the weighted average character width;
A second calculation subunit that calculates the character width context feature based on the difference, and

here,
(Outside 20)

Is the character width of the i-th character in the initial recognition result,
(Outside 21)

Is
(Outside 22)

The reliability of the device.

(Appendix 12)
The apparatus according to appendix 11, wherein
The character width is the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the character itself And the width of the gap on the right side of the character, and the sum of the width of the character itself and the width of the gap on the left side of the character.

(Appendix 13)
The device according to any one of appendices 8 to 12,
The apparatus further includes a preprocessing unit for performing preprocessing on the character string area.

(Appendix 14)
The device according to any one of appendices 8 to 12,
The recognition unit is based on at least one of the language type context feature and the character width context feature and at least one of a single character recognizer feature, a semantic context feature, and a geometric context feature. An apparatus for recognizing a character string included in the character string region by the route search strategy.

(Appendix 15)
A program for causing a computer to execute each step of the method described in Appendix 1.

(Appendix 16)
A computer-readable storage medium storing the program according to attachment 15.

The preferred embodiment of the present invention has been described above, but the present invention is not limited to this embodiment, and all modifications to the present invention belong to the technical scope of the present invention unless departing from the spirit of the present invention.

Claims (10)

A method for recognizing a character string in an image,
An extraction step of extracting a character string region in the image;
A division step of performing over-segmentation on the character string region;
Recognizing a character string included in the character string region by a route search strategy based on at least one of a language type context feature and a character width context feature. The method of claim 1, comprising
The language type context feature is:
Determining whether each character in the search path and its one adjacent character belong to the same type of language;
The method is determined by calculating the language type context feature based on the result of the determination. The method of claim 1, comprising
The language type context feature is:
Determining whether each character in the search path and each adjacent character in the plurality of adjacent characters belong to the same type of language;
The method is determined by calculating the language type context feature based on the result of the determination. The method of claim 1, comprising
The character width context feature is:
Initial recognition is performed on the character string area,
Based on the result of the initial recognition, the weighted average character width is estimated by the following formula 7,
Determining the difference between the character width of each character in the search path and the weighted average character width;
Determined by calculating the character width context feature based on the difference;

here,
(Outside 23)

Is the weighted average character width;
(Outside 24)

Is the character width of the i-th character in the result of the initial recognition,
(Outside 25)

Is
(Outside 26)

Is the confidence of the method. A method according to claim 4, wherein
The character width is the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the character itself And the width of the gap on the right side of the character and the sum of the width of the character itself and the width of the gap on the left side of the character. A device for recognizing a character string in an image,
An extraction unit for extracting a character string region in the image;
A division unit that performs over-segmentation on the character string region;
A recognition unit for recognizing a character string included in the character string region by a route search strategy based on at least one of a language type context feature and a character width context feature. The apparatus according to claim 6, wherein
The recognition unit is
A language type determination subunit for determining whether each character in the search path and its one adjacent character belong to the same type of language;
A first calculation subunit for calculating the language type context feature based on the result of the determination. The apparatus according to claim 6, wherein
The recognition unit is
A language type determination subunit for determining whether each character in the search path and each adjacent character in the plurality of adjacent characters belong to the same type of language;
A first calculation subunit for calculating the language type context feature based on the result of the determination. The apparatus according to claim 6, wherein
The recognition unit is
An initial recognition subunit for performing initial recognition on the character string area;
Based on the result of the initial recognition, an average character width estimation subunit that estimates a weighted average character width by the following formula 8;
A difference determination subunit for determining a difference between a character width of each character in the search path and the weighted average character width;
A second calculation subunit that calculates the character width context feature based on the difference, and

here,
(Outside 27)

Is the weighted average character width;
(Outside 28)

Is the character width of the i-th character in the result of the initial recognition,
(Outside 29)

Is
(Outside 30)

The reliability of the device. The apparatus according to claim 9, wherein
The character width is the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the sum of the width of the gap on both sides of the character and the width of the character itself, the character itself And the width of the gap on the right side of the character, and the sum of the width of the character itself and the width of the gap on the left side of the character.

Images