JP5660294B2 - Character image identification device, character image identification method and program - Google Patents

Description

  The present invention relates to a character image identification device, a character image identification method, and a program.

  Non-Patent Document 1 uses a shading pattern called a Haar-Like feature, and the shape, size, and position where the pattern is applied on the target image are different, and the output value of each pattern is input and a predetermined identification value is output. A plurality of weak classifiers (generally called weak classifiers) are selected, and a weak classifier suitable for representing the intensity of facial brightness is selected from the plurality of weak classifiers, and the plurality of weak classifiers are selected. Discloses a method of performing face recognition using a classifier (in general, referred to as a strong classifier) connected in cascade.

  Regarding object identification techniques other than face identification, there are methods disclosed in Patent Document 1 and Patent Document 2.

  In Patent Document 1, in pattern recognition determination, sections defined by upper / lower thresholds are provided for n-dimensional feature quantities extracted from a determination target, respectively, and feature quantities extracted from the recognition target are sections of all dimensions. A pattern learning method for determining whether or not a pattern to be recognized is included in a specific category based on whether or not it falls within the pattern is disclosed.

  In Patent Document 2, in the pattern recognition system, instead of setting a unique threshold value for an input pattern, a threshold value based on a similarity distribution with the standard pattern is set for each input pattern. A pattern recognition system is disclosed in which a pattern matching an input pattern is narrowed down from a group.

Japanese Patent Laid-Open No. 06-124349 Japanese Patent Laid-Open No. 11-120291

Paul Viola, Michael Jones, `` Rapid Object Detection using a Boosted Cascade of Simple Features. '' IEEE CVPR, 2001.

  An object of the present invention is to provide a character image identification device, a character image identification method, and a program capable of identifying in a short time whether a cut character candidate image is a character image or a non-character image.

[Character image identification device]
According to the first aspect of the present invention, feature amounts based on a plurality of different image characteristics are sequentially extracted from a character candidate image that is determined to be a character image from image data to be processed and cut out. A multi-stage extraction means;
An identification value provided corresponding to each of the plurality of stages of extraction means, and indicating the possibility that the character candidate image is a character image or a non-character image based on the feature amount extracted by the corresponding extraction means And a comparison unit that compares the identification value calculated by the calculation unit with a threshold value set in advance for each image characteristic, and based on the comparison result in the comparison unit, the character candidate A plurality of stages of identifying means for identifying whether the image is a character image or a non-character image, or a character image, a non-character image, or an unknown unknown image;
When the character candidate image is identified as a character image or a non-character image by at least one of the plurality of stages of identification means, the character candidate image is determined to be a character image or a non-character image. Determination means for
Each of the plurality of stages of extraction means is a character image identification apparatus that extracts a feature amount based on the set image characteristics when the previous-stage identification means identifies the character candidate image as an unknown image .

According to a second aspect of the present invention, the plurality of stages of identification means further comprise addition means for adding the identification value in the preceding stage identification means and the identification value calculated by the current stage calculation means, respectively. The character image identification device according to claim 1 .

According to a third aspect of the present invention, the calculation means calculates an identification value from the feature amount using an identification function calculated in advance based on the frequency distribution of a plurality of character images and non-character images for the feature amount. It is the character image identification device of Claim 1 or 2 provided.

According to a fourth aspect of the present invention, from the first aspect, the plurality of stages of identification means are arranged in a preceding stage as the identification means having a higher identification ability for identifying whether the character candidate image is a character image or a non-character image. 3 a character image identification apparatus according to any one of.

[Character image identification method]
The present invention according to claim 5 is a step of extracting feature amounts based on a plurality of different image characteristics from character candidate images that have been determined to be character images from image data to be processed and cut out. When,
Calculating an identification value indicating the possibility that the character candidate image is a character image or a non-character image based on the extracted feature amount; and
Comparing the calculated identification value with a preset threshold value for each image characteristic;
Identifying whether the character candidate image is a character image, a non-character image, or a character image, a non-character image, or an unknown image based on a comparison result with a threshold value When,
When the character candidate image is identified as a character image or a non-character image, it is determined that the character candidate image is a character image or a non-character image, and the character candidate image is identified as an unknown image Includes a step of extracting a feature amount, a step of calculating an identification value, a step of comparing with a threshold value, and a step of repeating the step of identifying a character candidate image again.

[program]
The present invention according to claim 6 is a step of extracting feature amounts based on a plurality of different image characteristics from character candidate images that are determined to be character images from image data to be processed and cut out. When,
Calculating an identification value indicating the possibility that the character candidate image is a character image or a non-character image based on the extracted feature amount; and
Comparing the calculated identification value with a preset threshold value for each image characteristic;
Identifying whether the character candidate image is a character image, a non-character image, or a character image, a non-character image, or an unknown image based on a comparison result with a threshold value When,
When the character candidate image is identified as a character image or a non-character image, it is determined that the character candidate image is a character image or a non-character image, and the character candidate image is identified as an unknown image Is a program for causing a computer to execute a step of extracting a feature amount, a step of calculating an identification value, a step of comparing with a threshold value, and a step of repeating a step of identifying a character candidate image again.

  According to the first aspect of the present invention, characters that can be identified in a short time compared with the case where the extracted character candidate image is a character image or a non-character image is not provided. An image identification device can be provided.

According to the second aspect of the present invention, in addition to the effect obtained by the first aspect of the present invention, there is provided a character image identification device capable of performing identification processing taking into account the identification results up to the previous stage. Can do.

According to the third aspect of the present invention, in addition to the effect obtained by the first or second aspect of the present invention, there is provided a character image identification device capable of easily calculating an identification value by calculation using an identification function. can do.

According to the fourth aspect of the present invention, it is possible to provide a character image identification device capable of determining whether a character candidate image is a character image or a non-character image at an early stage among a plurality of identification means.

According to the fifth aspect of the present invention, it is possible to identify whether the extracted character candidate image is a character image or a non-character image in a short time compared to the case where this configuration is not provided. An image identification method can be provided.

According to the sixth aspect of the present invention, a program capable of identifying in a short time whether the extracted character candidate image is a character image or a non-character image as compared with the case where the present configuration is not provided. Can be provided.

It is a block diagram which shows the hardware constitutions of the character image identification device 10 in the 1st Embodiment of this invention. It is a block diagram which shows the function structure of the character image identification device 10 in the 1st Embodiment of this invention. It is a figure explaining an example of image size normalization in the image size normalization part. It is a diagram illustrating a specific example of the feature extractor 31 ₁ to 31 _N. It is a figure showing the specific example of the frequency distribution of the character image with respect to black pixel density, and a non-character image. It is a figure showing the specific example of the frequency distribution of the character image with respect to a black pixel perimeter, and a non-character image. It is a figure showing the specific example of the frequency distribution of the character image with respect to an average run length feature-value, and a non-character image. It is a figure showing the specific example of the frequency distribution of the character image with respect to the maximum run length, and a non-character image. It is a figure showing the specific example of the frequency distribution of the character image with respect to complexity, and a non-character image. It is a figure showing the specific example of the frequency distribution of the character image with respect to the number of connected components, and a non-character image. It is a figure showing the specific example of the frequency distribution of the character image with respect to a horizontal black-and-white change score, and a non-character image. It is a figure showing the specific example of the frequency distribution of the character image with respect to a vertical black-and-white change score, and a non-character image. It is a diagram for explaining the structure of a character discriminator 32 ₁ to 32 _N. It is a diagram for explaining the operation of the character discriminator 321 _to 323 identified value of _N calculating unit 41 ₁ to 41 _N. It is a figure for demonstrating the setting method of an identification function. It is a diagram for explaining the operation of the comparator 42 ₁ through 42 _N. It is a figure for demonstrating operation | movement of the comparator 42 _N when only one threshold value thN is used in the last stage. It is a figure for demonstrating the setting method of a threshold value. It is a figure for demonstrating the setting method of a threshold value in another frequency distribution. It is a figure for demonstrating the setting method of a threshold value in another frequency distribution. It is a flowchart for demonstrating the identification operation | movement in the character image identification device 10 of the 1st Embodiment of this invention. It is a block diagram which shows the function structure of the character image identification device 10a in the 2nd Embodiment of this invention. It is a diagram for explaining the structure of a character classifier 132 ₁ to 132 _N in the second embodiment of the character image identification apparatus 10a of the present invention.

Next, embodiments of the present invention will be described in detail with reference to the drawings.
[First embodiment]
FIG. 1 is a diagram showing a hardware configuration of a character image identification device 10 according to the first embodiment of the present invention.

  As shown in FIG. 1, the character image identification device 10 of the present embodiment transmits data to and from a CPU 11, a memory 12, a storage device 13 such as a hard disk drive (HDD), and an external device via a network. And a communication interface (IF) 14 that performs reception, and a user interface (UI) device 15 including a touch panel or a liquid crystal display and a keyboard. These components are connected to each other via a control bus 16.

  The CPU 11 controls the operation of the character image identification device 10 by executing a predetermined process based on the character image identification program stored in the memory 12 or the storage device 13. In the present embodiment, the CPU 11 has been described as reading and executing the character image identification device program stored in the memory 12 or the storage device 13, but the program is stored in a storage medium such as a CD-ROM. It is also possible to provide it to the CPU 11.

  FIG. 2 is a block diagram showing a functional configuration of the character image identification device 10 realized by executing the character image identification program.

As shown in FIG. 2, the character image identification device 10 of the present embodiment includes a character candidate image cutout unit 21, an image size normalization unit 22, N feature quantity extractors 31 _{1 to} 31 _N , Character discriminators 32 _{1 to} 32 _N and a determiner 24 are provided.

Hereinafter, in the case where any one of a plurality of components that can exist such as “feature amount extractors 31 _{1 to} 31 _N ” is indicated without being specified, it may be simply abbreviated as “feature amount extractor 31”. .

  The character image identification device 10 of the present embodiment is based on binarized image data and an image cut out as an image that appears to be a character image on the premise of character extraction processing for character recognition (hereinafter referred to as a character candidate image). Is called a character image or a non-character image.

  The character candidate image cutout unit 21 cuts out an image of an area determined to be a character image from the image data to be processed as a character candidate image.

Various methods have been proposed for the character cutting technique, and any one of them may be used in the present embodiment. For example, there are techniques disclosed in the following documents.
JP-A-5-114047 JP-A-4-100189 JP-A-4-92992 JP-A-4-68481 JP-A-9-54814 JP-A-9-185681 Paragraph [0021] Character boundary candidate extraction method Character cutting position determination method described in paragraph [0005] of Japanese Patent Laid-Open No. Hei 5-128308 It should be noted that, as a matter of course, any technique that can cut out a character candidate image can be used. Other methods may be used.

The image size normalization unit 22 uses the size of the character candidate image cut out by the character candidate image cutout unit 21 for processing in the feature quantity extractors 31 _{1 to} 31 _N and the character identification periods 32 _{1 to} 32 _N. Standardize to a certain size.

  FIG. 3 is a diagram for explaining an example of image size normalization in the image size normalization unit 22. Character candidate images that are extracted by binarizing the document image and extracted by the character extraction technology are various character sizes included in the document image, illegal character extraction positions, and illegal images and line drawing areas included in the document image. It becomes various sizes under the influence of cutting out. For example, FIG. 3A shows a different character size, FIG. 3B shows an example of an illegal cutout position, and FIG. 3C shows an example of an illegal cutout of a photograph or a line drawing area. The image size normalization unit 22 converts the character candidate images input in such various sizes into uniform sizes as shown in FIGS. 3 (A), 3 (B), and 3 (C). Standardize. The standardized size is, for example, a 64 × 64 pixel size. Of course, not only normalization to the square size, but also normalization to a rectangle may be used if normalization to the same size is performed.

The feature amount extractors 31 _{1 to} 31 _N are configured in a plurality of stages, and sequentially extract feature amounts based on a plurality of different image characteristics from the character candidate images normalized by the image size normalization unit 22.

The character identifiers 32 _{1 to} 32 _N are also configured in a plurality of stages, and the feature quantity extractors 31 _{1 to} 31 _N and the character identifiers 32 _{1 to} 32 _N are provided in a one-to-one correspondence. Yes. Therefore, description will be made assuming that one stage is constituted by a set of one feature quantity extractor 31 and one character classifier 32. For example, the feature quantity extractor 31 ₁ and the character classifier 32 ₁ are arranged in front of the feature quantity extractor 31 ₂ and the character classifier 32 ₂ , and the feature quantity extractor 31 ₃ and the character classifier 32 ₃ are The feature amount extractor 31 ₂ and the character identifier 32 ₂ are arranged at the subsequent stage.

FIG. 4 is a diagram illustrating a specific example of the feature quantity extractors 31 _{1 to} 31 _N. As shown in FIG. 4, for example, when the character discriminator 32 ₁ uses the black pixel density for discrimination judgment, the feature amount extractor 31 ₁ that extracts the feature amount functions as the black pixel density extractor 31 ₁ , and character candidates A black pixel density Db represented by the following equation (1) is calculated from the image as a feature quantity and output to the character classifier 32 ₁ .

  Db = number of black pixels in the character candidate image / character candidate image area (1)

Also, if the character discriminator 32 ₂ uses a black pixel perimeter identification decision, the feature quantity extractor 31 ₂ for extracting the characteristic amount similarly functions as a black pixel perimeter extractor 31 _2, from the character candidate image The black pixel perimeter La is calculated as the feature amount and output to the character discriminator 32 ₂ .

Similarly, the character classifiers 32 _{3 to} 32 _{6 use the} 0-degree direction black pixel average run length, the 45-degree direction black pixel average run length, the 90-degree direction black pixel average run length, and the 135-degree direction black pixel average, respectively, for identification determination. When the run length is used, the feature quantity extractors 31 _{3 to} 31 ₆ respectively include a 0 degree direction black pixel average run length extractor 31 ₃ , a 45 degree direction black pixel average run length extractor 31 ₄ , and a 90 degree direction black pixel. The average run length extractor 31 ₅ functions as a 135-degree direction black pixel average run length extractor 31 ₆ .

In this case, the 0 degree direction black pixel average run length extractor 31 ₃ outputs the 0 degree direction black pixel average run length L ₀ to the character identifier 32 ₃ as the feature quantity, and the 45 degree direction black pixel average run length L ₃ is output. The length extractor 31 ₄ outputs the 45-degree direction black pixel average run length L ₄₅ as a feature quantity to the character identifier 32 ₄ . Further, 90-degree direction black pixel average run length extractor 31 ₅ outputs the 90-degree direction black pixel average run length L ₉₀ as the feature amount on the character discriminator 32 _5, 135-degree direction black pixel average run length extractor 31 ₆ outputs the 135-degree direction black pixel average run length L ₁₃₅ as features in the character discriminator 32 _6.

  As described above, the feature quantity extractors 31 are configured by the number of feature quantities based on the image characteristics used by the character classifier 32 for discrimination determination.

  Therefore, the feature quantity extractor 31 is not limited to the one that extracts the feature quantity as shown in FIG. 4, but various feature quantities are extracted from the input binarized character candidate images. It is possible to configure. For example, a feature quantity extractor as shown below can be applied.

(1) Maximum run length extractor (0 degree, 45 degree, 90 degree, 135 degree directions)
(2) Complexity extractor (3) Connected component number extractor 53
(4) 0 degree direction (horizontal direction) black and white change point extractor (5) 90 degree direction (vertical direction) black and white change point extractor

  Here, (1) the maximum run length extractor is a feature quantity extractor independent of each direction, and calculates the maximum run length in each direction.

(2) The complexity extractor calculates the complexity. The complexity C is calculated by the following equation (2), for example.
C = (perimeter) ² / number of black pixels (2)

  (3) The connected component extractor calculates the number of independent connected components in the character candidate image. For example, 1 for the character “A” and 2 for the character “I”.

  (4) The 0-degree direction black-and-white change point extractor calculates, for example, the total number of times of changing from black pixels to white pixels in the 0-degree direction (horizontal direction or main scanning direction) in the character candidate image.

  (5) The 90-degree direction black-and-white change point extractor calculates, for example, the total number of times of changing from black pixels to white pixels in the 90-degree direction (vertical direction or sub-scanning direction) in the character candidate image.

  As described above, the feature quantity extractor 31 extracts a plurality of feature quantities having different image characteristics from the binarized character candidate images, but the range (value range) of the extracted feature quantities is different. For example, the extracted feature value x of the real value is normalized in the range of (0 ≦ x ≦ 127) or in the range of (0 ≦ x ≦ 1), and each feature amount is extracted by aligning the range. The feature amount output by the device 31 can be handled in the same manner.

  The specific example of the feature amount extracted by the feature amount extractor 31 has been described so far. Here, what feature amount is used as the feature amount for character identification in the feature amount extractor 31 will be described. .

  5, 6, and 7 are diagrams illustrating specific examples of the frequency distribution of the character image and the non-character image with respect to the black pixel density, the black pixel perimeter, and the average run length described above. Here, in FIGS. 5, 6, and 7, the horizontal axis represents a value obtained by normalizing each feature amount in a range of (0 ≦ x ≦ 127), and the vertical axis corresponds to the number of character images and non-character images. If the total number of character images is N and the total number of non-character images is M, 1 / 2N if the feature amount is extracted from the character image, and 1 / 2M if the feature amount is extracted from the non-character image. It represents as.

  As can be seen from FIGS. 5, 6, and 7, the character image and the non-character image represent different frequency distributions in any of the feature amounts of black pixel density, black pixel perimeter, and average run length. I understand that. That is, it can be seen that the character image distribution and the non-character image distribution have different distribution peaks, and it is possible to identify characters and non-characters by using a predetermined threshold for these feature values.

  Similarly, the maximum run length, complexity, number of connected components, horizontal black-and-white change points, and vertical black-and-white change points, which are other feature quantities described so far, are also shown in FIGS. As shown in a specific example of the frequency distribution, it can be seen that the character image and the non-character image represent different frequency distributions.

  In other words, when determining whether a character candidate image is a character image or a non-character image, the feature amount is selected so that the frequency distribution of the character image and the non-character image represents different distributions. A plurality of feature quantity extractors 31 for extracting quantities may be configured.

Further, as described above, each of the character identifiers 32 _{1 to} 32 _N performs character image identification as to whether the character candidate image is a character image or a non-character image based on feature amounts based on different image characteristics. Do.

When the character candidate image is identified as a character image or a non-character image by at least one character identifier among the _N character identifiers 32 _{1 to} 32 _N , the determiner 24 selects the character candidate. The image is finally determined to be a character image or a non-character image.

Next, the configuration of the character discriminators 32 _{1 to} 32 _N will be described with reference to FIG.
As shown in FIG. 13, the character classifier 32 ₁ includes an identification value calculation unit 41 ₁ and a comparator 42 ₁ . The character identifier 32 ₂ includes an identification value calculator 41 ₂ , a comparator 42 _2, and an adder 43 ₂ . Similarly, the character discriminators 32 _{3 to} 32 _N are configured by discrimination value calculation units 41 _{3 to} 41 _N , comparators 42 _{3 to} 42 _N , and adders 43 ₃ to 43 _N , respectively.

The identification value calculation units 41 _{1 to} 41 _N calculate and output the identification values y1 to yN based on the feature amounts extracted from the corresponding feature amount extractors 31 _{1 to} 31 _N , respectively. The identification values y1 to yN are values indicating the possibility that the character candidate image is a character image or a non-character image, and details thereof will be described later.

Adder 43 ₃ and ~ 43 _N, respectively, and a previous character discriminator 32 ₁ ~32 _N-1 comparators 42 ₁ ~42 _N-1 identification value is input to Y1~Y (N-1), The identification values y2 to yN calculated by the identification value calculation units 41 _{2 to} 41 _N of the current stage provided in the same character identifiers 32 _{2 to} 32 _N are added and compared as identification values Y2 to YN. Outputs to devices 42 _{2 to} 42 _N.

In the _first character discriminator 32 ₁ , the preceding character discriminator does not exist, and therefore, the discrimination value y ₁ calculated by the discrimination value calculation unit 41 ₁ is directly output to the comparator 42 ₁ as the discrimination value Y ₁ .

That is, the identification value inputted to the comparator 42 with, it all the identification values becomes a value obtained by accumulating in front than, for example, the identification value Y2 which is input to the comparator 42 ₂ was added to y1 and y2 a value, identification value Y3 input to the comparator 42 ₃ is a value obtained by adding the y1, y2, y3.

Then, the comparators 42 _{1 to} 42 _N respectively compare the input identification values Y1 to YN with threshold values set in advance for each image characteristic. As a result, the character identifiers 32 _{1 to} 32 _N identify the character candidate images in three ways based on the comparison results in the comparators 42 _{1 to} 42 _N. That is, the character identifiers 32 _{1 to} 32 _N identify whether the character candidate image is a character image, a non-character image, or an unknown image. If the character image or non-character image is identified, the identification result is output and the identification operation for the current character candidate image ends here. When the comparators 42 _{1 to} 42 _{N-1 determine} that the character candidate image is a character image or a non-character image, it is determined that the result is a feature quantity extractor 32 _{2 to} 32 _{N in the} next stage. Will be notified.

The feature quantity extractors 32 _{2 to} 32 _N extract the feature quantity set from the character candidate image only when it is unclear whether the character candidate image is a character image in the preceding character identification process. .

That is, the character discriminators 32 _{1 to} 32 _N sequentially identify the character candidate images to identify whether the character candidate images are character images or are unknown, and are unknown in the current identification process. If it is, the identification process in the next stage is performed.

However, the character identifier 32 _N at the last stage identifies the character image or non-character image from the character image in consideration of the use of the character image identification device 10, the required accuracy, etc. You may operate | move so that it may perform, and you may operate | move so that it may identify with a non-character image.

As described above, in each of the character classifiers 32 _{1 to} 32 _N , the classification operation is performed in the subsequent-stage character classifier 32 only when it is determined that the previous-stage character classifier 32 is unknown. When the character classifiers 32 _{1 to} 32 _N determine whether or not the image is a character image at a certain stage, the identification process is terminated at that point. In other words, in the character image identification device 10 of the present embodiment, for the input character candidate image, for the character candidate image that is easy to identify, the character / non-character determination is determined by the character identifier 32 at the previous stage as much as possible. For character candidate images that are difficult to identify, the operation is performed so as to perform identification based on the judgment of many character classifiers 32.

Next, the operation of the identification value calculation units 41 _{1 to} 41 _{N in} the character identifiers 32 _{1 to} 32 _N will be described with reference to FIG.
FIG. 14 schematically shows the operation of the identification value calculation units 41 _{1 to} 41 _N. As illustrated in FIG. 14, the identification value calculation units 41 _{1 to} 41 _N receive the output xi from the feature quantity extractor 31 _i (i = 1 to N), and are set for each of the identification value calculation units 41 _{1 to} 41 _N. The discriminant value yi is output according to the discriminated discriminant function f (x). Here, the discriminant function f (x) is a one-input one-output function expressed by the following equation (3).

    y = f (x) (3)

That is, the identification value yi is uniquely output according to the equation (3) for the feature quantity xi that is the output from the feature quantity extractor 31 _i .

  Consider a case where the distribution of the character image and the non-character image corresponding to the feature amount extracted from the specific feature amount extractor 31 is represented in FIG. For example, this is the case of the black pixel density, the black pixel perimeter, the average run length, the maximum run length, the complexity, and the horizontal black-and-white change points shown in the specific examples in FIGS. 5 to 9 and 11 described above.

A distribution as shown in FIG. 15A is obtained by extracting specific feature amounts from a large number of character images and non-character images in advance and generating histograms of characters and non-character images for the extracted feature amounts. It is done. In the example shown in FIG. 15A, the feature amount x is (0 ≦ x ≦ 127) and the value is normalized. On the premise that such a distribution is obtained, the discrimination function y = f (x) set in each discrimination value calculation unit 41 _i is set by a function represented by the following equation (4).

  y = (character image distribution (x) −non-character image distribution (x)) ÷ (character image distribution (x) + non-character image distribution (x)) (4)

  In this equation (4), the character image distribution (x) is the number of character images corresponding to the feature quantity x, and the non-character image distribution (x) is the number of non-character images corresponding to the feature quantity x. is there.

FIG. 15B shows the discrimination function f (x) set in this way. As shown in FIG. 15B, the discriminant function y = f (x) set by the equation (4) is in the range of the feature quantity x (0 ≦ x ≦ 127), and the value range (−1 ≦ y ≦ 1). Assuming that the lower limit value x0 of the character image distribution and the upper limit value x1 of the non-character image distribution in FIG. 7, the value of the identification value y is as shown in the following equation.
y = 1 (x> x1)
= -1 (x <x0)

Thus, by setting the discrimination function y = f (x) as shown in FIG. 15B, a unique discrimination value yi is obtained for the feature quantity output from the feature quantity extractor 31 _i .

Next, details of the identification process based on the identification value yi output from the identification value calculation unit 41 _i will be described.

In the comparators 42 _{1 to} 42 _N , a plurality of threshold values, specifically to the identification values y1 to yN output from the identification value calculation units 41 _{1 to} 41 _N and added by the adders 43 ₂ to 43 _N , specifically, There are three ways of determining whether a character candidate image is a character image, a non-character image, or a character image is a non-character image using two threshold values th (L) and th (H). Make a decision. The two threshold values th (L) and th (H) are preferably set in advance separately from the character candidate image identification process.

Specifically, as shown in FIG. 16, two threshold values th (L) and th (H) having different values are set in the comparators 42 _{1 to} 42 _N , respectively, and th (L) <Th (H). For example, the larger one of the two threshold values set in the comparator 42 ₁ is expressed as th1 (H), and the smaller value is expressed as th1 (L).

Then, the comparators 42 _{1 to} 42 _N compare the input identification value Yi with the two threshold values thi (L) and thi (H) to determine whether the character candidate image is a character image. Identify whether it is a character image or unknown.

Specifically, the comparators 42 _{1 to} 42 _N use the input identification value Yi and the two threshold values thi (L) and thi (H) described above to generate three types of character candidate images as follows. To identify. (I = 1 to N)
Character image (thi (H) ≦ Yi)
Non-character image (Yi ≦ thi (L))
Unknown image (thi (L) <Yi <thi (H))

As shown in FIG. 17, the threshold value used in the comparator 42 _N is set so that there are two classification results of “non-character” and “character” in the final-stage character classifier 32 _N. Only one threshold value thN may be set so as to be identified as either “non-character” or “character”.

  Next, a method for setting the threshold values th (L) and th (H) will be described.

  Each of the two threshold values th (L) and th (H) is determined based on an identification function y = f (x) set for each identification value calculation unit 41. FIG. 18 shows a specific example of threshold setting. In the specific example shown in FIG. 18B, the output values of the discrimination function y = f (x) in the feature quantity x2 and the feature quantity x3 are set as threshold values th (L) and th (H), respectively. . Here, as a specific example of the determination method of the feature amount x2 and the feature amount x3, for example, the boundary feature amount including the lower 5% number in the character image distribution is set as the feature amount x2, and the upper 5% number in the non-character image distribution. A method in which the boundary feature amount including the feature amount x3 is used. However, the method is not limited to such a method, and the determination method of the boundary feature amounts x2 and x3 may be changed depending on the use of the character image identification device 10 or the required accuracy.

As a simple determination method, the feature amount x2 may be the lower limit value of the character image distribution and the feature amount x3 may be the upper limit value of the non-character image distribution as described below.
x2 = x0
x3 = x1

  A specific example of the threshold setting shown in FIG. 18 is an example on the assumption that the distribution of character images and non-character images for a specific feature amount is represented in FIG. However, the distribution of the character image and the non-character image with respect to the feature amount is not limited to the case shown in FIG.

Therefore, a specific method for setting the threshold when the distribution of the character image and the non-character image is different from that in FIG. 18A will be described with reference to FIG.
FIG. 19A shows the frequency distribution of characters and non-character images when the number of connected components shown in FIG. In FIG. 19A, the character image distribution is localized in the feature value range having a large value, and the character image distribution is substantially included in the non-character image distribution. Here, x0 is a lower limit value of the character image distribution.

FIG. 19B shows an identification function set by the function represented by the above formula (4) and a threshold value set for the character and non-character image distribution shown in FIG. Th (L) and th (H) are illustrated. In the threshold value th (L) and th (H) setting method illustrated in FIG. 19B, the threshold value th (L) includes, for example, the number of upper and lower 5% in the character image distribution. The boundary feature amounts x1 and x2 are determined, and the function values for the boundary feature amounts x1 and x2 are set as a threshold value th (L) and a threshold value th (H). As a simple determination method, the feature amount x2 may be the lower limit value of the character image distribution, and the feature amount x3 may be the maximum feature amount value as described below.
x1 = x0
x2 = 127

  Further, a specific example for another feature amount will be described with reference to FIG. FIG. 20A shows a case of frequency distribution of characters and non-character images when, for example, the vertical black-and-white change score shown in the specific example in FIG. 12 is used as the feature amount. In FIG. 20A, the character image distribution and the non-character image distribution are different distributions, but the non-character image distribution is localized in the range of the feature amount having a small value. Here, x0 and x1 are the upper limit value of the non-character image distribution and the upper limit value of the character image distribution, respectively.

FIG. 20B shows an identification function set by the function represented by the above formula (4) and a threshold value set for the character and non-character image distribution shown in FIG. Th (L) and th (H) are illustrated. In the threshold value th (L) and th (H) setting method illustrated in FIG. 20B, the threshold feature th (L) includes, for example, a boundary feature including the number of lower 5% in the character image distribution. When the amount x2 is the threshold th (H), the boundary feature amount x3 including the top 5% number in the non-character image distribution is determined, and the function value for the boundary feature amounts x2 and x3 is set to the threshold th ( L) and threshold value th (H). As a simple determination method, the feature quantity x2 may be the minimum feature quantity value and the feature quantity x3 may be the non-character distribution upper limit value as described below.
x2 = 0
x3 = x0

  Thus, the specific example of FIGS. 19A and 20A, which is a distribution different from the character and non-character image distribution shown in FIG. 18A, is the same as FIG. 18B. The thresholds th (L) and th (H) can be set by a setting method. That is, the threshold values th (L) and th (H) can be set in the same manner as described above, regardless of the distribution of characters and non-characters for a specific feature amount.

  Next, the identification operation in the character image identification apparatus 10 of this embodiment will be described in detail with reference to the flowchart shown in FIG.

First, the feature quantity extractor 31 _1, extracts a feature x (1) from the character candidate image (step S101). Then, in the character classifier 32 ₁ , the identification value calculation unit 41 ₁ calculates an identification value y1 based on the extracted feature quantity x (1) (step S102).

Then, in the character discriminator 32 ₁ , the comparator 42 ₁ compares the calculated discriminant value y1 with preset threshold values th1 (L) and th1 (H) (step S103, S104).

By comparison in the comparator 42 _1, the identification value y1 is th1 when (L) below (yes in step S103), the determination result is to be a non-character identification process ends (step S203). If the identification value y1 is greater than or equal to th1 (H) (yes in step S104), the determination result is assumed to be a character and the identification process ends (step S204).

When it is determined that the identification value y1 is greater than th1 (L) and smaller than th1 (H), the determination result is unknown, and the comparator 421 applies to the next-stage feature quantity extractor 32 ₂ . This is notified (both in steps S103 and S104, no).

As a result, the feature amount extractor 31 ₂ extracts the feature amount x (2) from the character candidate image (step S105), and the character identifier 32 ₂ extracts the feature amount x extracted by the identification value calculation unit 41 _2. An identification value y2 based on (2) is calculated. The adder 43 ₂ adds the identification value y1 (Y1) and the identification value y2 to calculate the identification value Y2 (step S106).

In the character discriminator 32 ₂ , the comparator 42 ₂ compares the calculated discriminant value Y2 with preset threshold values th2 (L) and th2 (H) (step S107, S108).

By comparison in the comparator 42 _2, the identification value Y2 is the case of th2 (L) or less (yes in step S107), the determination result is to be a non-character identification process ends (step S203). If the identification value Y2 is greater than or equal to th2 (H) (yes in step S108), the determination result is assumed to be a character and the identification process ends (step S204).

If the identification value Y2 is determined to th2 (L) less than the greater than th2 (H) is determined as the result is unknown, to the comparator 42 ₂ from the next stage feature quantity extractor 32 ₃ To that effect (both in steps S107 and S108, no). In this way, the same processing is repeated in step S109 and subsequent steps.

In this iterative process, the character identifiers 32 _{2 to} 32 _N are provided with adders 43 ₂ to 43 _N so that the sum of the identification value at the current stage and the identification value calculated up to the previous stage (for example, comparison) vessel 42 comparison ₃ in Y3 = y1 + y2 + y3) and the threshold value are to be made.

Finally, the character discriminator 32 _N compares the calculated identification value YN with the threshold thN (step S202), and finally determines whether the determination result is “non-character” (step S203). "(Step S204).

Next, the configuration order of the character identifiers 32 _{1 to} 32 _N will be described. As described above, in the character image identification device 10 in the present embodiment, the character identifier 32 is operated so that the character identifier 32 in the preceding stage can be identified as “character” and “non-character” as much as possible. The construction order of _{1 to} 32 _{N is} the order in which the character / non-character discrimination capability is high. That is, the character discriminator 32 having a higher discriminating ability is arranged in the preceding stage. The character discriminator 32 having a high character / non-character discriminating ability here is the character discriminator 32 having the lowest misidentification rate based on the set discriminating function and threshold value. It is obtained by identifying the same character candidate image group including characters / non-characters every time, counting the number of misidentified images, and dividing the number of misidentified images by the total number of character candidate images.

  Further, regarding the threshold value used in the character classifier 32, in order to identify the character / non-character with high accuracy in the preceding character classifier 32, for example, in the description of threshold setting in FIG. As in the case where the lower limit value of the character image distribution and the upper limit value of the non-character image are used (x2 = x0, x3 = x1), a threshold value that is expected to be identified with higher accuracy is set.

[Second Embodiment]
Next, a character image identification apparatus according to a second embodiment of the present invention will be described.

In the first embodiment, in the identification processing in the character classifiers 32 _{2 to} 32 _N , the identification values from the character classifiers 32 _{1 to} 32 _{N−1 up} to the previous stage and the character classifiers 32 from the current stage are used. Although the sum with the identification value is used for the identification processing, the present invention is not limited to such a configuration.

The configuration of the character image identification device 10a of this embodiment will be described with reference to FIGS.
Character image identification apparatus 10a of the present embodiment is different from the character image identification apparatus 10 of the first embodiment shown in FIG. 2, replacing the character discriminator 32 ₁ to 32 _N in the character discriminator 132 ₁ to 132 _N It becomes the composition.

In the character identifiers 132 _{1 to} 132 _N in the present embodiment, as can be seen with reference to FIG. 23, an adder for sequentially adding the identification values in the previous stage is not provided. Therefore, in the character classifiers 132 _{1 to} 132 _{N according to} the present embodiment, the identification values y1 to yN calculated by the identification value calculation units 41 _{1 to} 41 _N and the preset threshold values th (L) and th. Comparison with (H) is made.

Even when the character identifiers 132 _{1 to} 132 _N are configured as in the present embodiment, the operation of the character identification process is the same as the operation of the first embodiment, and steps S106 and S201 in the flowchart shown in FIG. It is only necessary to replace Y2, YN, etc. with y2, yn.

[Modification]
In the above embodiment, the case where a character candidate image is identified as a character image or a non-character image has been described. However, the present invention is not limited to this, and identification is performed as in face identification processing. The present invention can be applied in the same manner even when identifying whether or not the target image matches a specific pattern.

10 character image identification device 11 CPU
12 Memory 13 Storage Device 14 Communication Interface (IF)
15 User Interface (UI) device 16 controls the bus 21 character candidate image extraction unit 22 image size normalization section 24 judging unit 31 ₁ to 31 _N feature quantity extractor 32 ₁ to 32 _N character discriminators 41 ₁ to 41 _N identification value Calculation unit 42 _{1 to} 42 _N comparator 43 ₂ to 43 _N adder 132 _{1 to} 132 _N character identifier

Claims (6)

A plurality of stages of extraction means for sequentially extracting feature amounts based on a plurality of different image characteristics from character candidate images that have been determined to be possibly character images from the image data to be processed;
An identification value provided corresponding to each of the plurality of stages of extraction means, and indicating the possibility that the character candidate image is a character image or a non-character image based on the feature amount extracted by the corresponding extraction means And a comparison unit that compares the identification value calculated by the calculation unit with a threshold value set in advance for each image characteristic, and based on the comparison result in the comparison unit, the character candidate A plurality of stages of identifying means for identifying whether the image is a character image or a non-character image, or a character image, a non-character image, or an unknown unknown image;
When the character candidate image is identified as a character image or a non-character image by at least one of the plurality of stages of identification means, the character candidate image is determined to be a character image or a non-character image. Determination means for
Each of the plurality of stages of extraction means is a character image identification device that extracts a feature amount based on the set image characteristics when the previous-stage identification means identifies the character candidate image as an unknown image . It said plurality of stages of identification means, respectively, and the identification value in the preceding-stage identification means, character image identification apparatus according to claim 1, further comprising adding means for adding the calculated identification value by the calculation means in the current stage. The calculating means, the plurality with respect to the feature amount using the discriminant function which is previously calculated based on the frequency distribution of the character image and the non-character image, according to claim 1 or 2, wherein comprising calculating the identification value from the feature Character image identification device. The character according to any one of claims 1 to 3 , wherein the plurality of stages of identifying means are arranged in a preceding stage as an identifying means having a higher identification ability for identifying whether a character candidate image is a character image or a non-character image. Image identification device. Extracting a feature amount based on a plurality of different image characteristics from a character candidate image that is determined to be a character image from image data to be processed and cut out;
Calculating an identification value indicating the possibility that the character candidate image is a character image or a non-character image based on the extracted feature amount; and
Comparing the calculated identification value with a preset threshold value for each image characteristic;
Identifying whether the character candidate image is a character image, a non-character image, or a character image, a non-character image, or an unknown image based on a comparison result with a threshold value When,
When the character candidate image is identified as a character image or a non-character image, it is determined that the character candidate image is a character image or a non-character image, and the character candidate image is identified as an unknown image The step of extracting the feature amount, the step of calculating the identification value, the step of comparing with the threshold value, the step of identifying the character candidate image again,
A character image identification method comprising: Extracting a feature amount based on a plurality of different image characteristics from a character candidate image that is determined to be a character image from image data to be processed and cut out;
Calculating an identification value indicating the possibility that the character candidate image is a character image or a non-character image based on the extracted feature amount; and
Comparing the calculated identification value with a preset threshold value for each image characteristic;
Identifying whether the character candidate image is a character image, a non-character image, or a character image, a non-character image, or an unknown image based on a comparison result with a threshold value When,
When the character candidate image is identified as a character image or a non-character image, it is determined that the character candidate image is a character image or a non-character image, and the character candidate image is identified as an unknown image A program for causing a computer to execute a step of extracting a feature amount, a step of calculating an identification value, a step of comparing with a threshold value, and a step of repeating a step of identifying a character candidate image again.

Images