CN103310210A - Character recognition device, recognition dictionary generation device and normalization method - Google Patents

Abstract

The invention provides a character recognition device and a normalization method is employed for normalizing a plurality of input images with various sizes in order to diminish the deviation in shapes of identical characters. A pretreatment for reducing disturbing factors is performed according to input images; the profile of the image which is pretreated is picked up; the pretreated image integrates with the profile picked up. The center of the composited image is within a range of a specific dimension and the pixel spread range of the composited image is within a range of a specific dimension. The composited image is mapped to the image with specific dimension after normalization. Normalization is performed on pretreated image according to mapping generated and the normalized image is turned into vector value in a vector space. The recognition dictionary based on a memory device can determine that what the character is according to the vector value and output the determination value.

Description

Character recognition device, recognition dictionary generating apparatus and method for normalizing

Technical field

The present invention relates to recognition dictionary generating apparatus and the character recognition device of literal identification usefulness, particularly the method for normalizing of character image.

Background technology

Character recognition device uses recognition dictionary, judges the literal classification of writing in the input picture, and the output result of determination.At this, when for example numeral was identified, the literal classification was 0～9 these 10 kinds of numerals.Make recognition dictionary by the recognition dictionary generating apparatus.

Character recognition device contains pre-service, normalization, feature extraction, identifies this four treatment steps to the treatment scheme of exporting result of determination from receiving input picture.

Fig. 2 is the process flow diagram of representing the processing of character recognition device execution in the past.

In character image input part 201, come input picture by the user or by the program that arithmetic unit is carried out.

For example carry out following steps in pretreatment portion 202: carry out the denoising of input picture, level and smooth (smoothing) of Fuzzy Processing (Pot か processing) etc., removing as much as possible becomes the disturbing factor that literal identification hinders.

Next, in normalization portion 203, carry out following steps: the pretreated image of various sizes is received as input, and make these size of images unanimities.Processing after can unifying thus.

Next, in feature extraction portion 209, carry out following steps: the image after the normalization is received as input, and be transformed to vector value on the vector space.Above-mentioned vector space is called feature space, above-mentioned vector value is called eigenvector.As feature extracting method, the special Cheng of extensive known extraction pixel characteristic, contour feature, gradient feature, gal cypress feature (ガ ボ ー Le) method (non-patent literature 1) such as.If desired, then use dimension reduction methods such as principal component analysis (PCA) or linear discriminant analysis, the dimension in compressive features space, and the dimension (non-patent literature 2) of reduction feature space.

By processing before this, input picture shows as the vector value on the feature space (eigenvector).

Next, in identification part 210, handle below the execution: judge the literal classification that eigenvector is affiliated with recognition dictionary 214.Recognition dictionary 214 is preserved the information which literal classification the each point that is used on the judging characteristic space belongs to.In non-patent literature 1 or non-patent literature 2, put down in writing the detailed description of the judgement of relevant use recognition dictionary 214.

In efferent 211, to output result of determination such as display device such as display or files.

In order to carry out high-precision literal identification, each processing of above-mentioned pretreatment portion 202, normalization portion 203, feature extraction portion 209 and identification part 210 has important effect.Therefore, the processing that is suitable for literal identification with each processing is important.

The preservation states of environment, papery or paper as the literal in the input picture of identifying object during according to Writing utensil equipped, amanuensis, font, scanning etc. are and different, even the same text type, size, shape or impairment grade also have deviation.In normalization, except the big or small unanimity that makes input picture, suppress the purpose of the deviation of the shape between the same literal type in this input picture in addition.Thus, can improve the discrimination of character recognition device.

In the method for normalizing of existing character image, linear normalization method (linear normalization method) is arranged, non-linear normalizing method (nonlinear normalization method), square normalization method (moment normalization method), two square normalization methods (bi-moment normalization method), CBA method (centroid-boundary alignment method: barycenter boundary alignment method), MCBA method (modified centroid-boundary alignment method), LDPF method (line density projection fitting method) etc.Deliver (non-patent literature 3) square normalization method and two square normalization method in the middle of these methods with paper etc. and had the benchmark result of this literal identification of high discrimination.

Fig. 5 is the key diagram with the example of the image of square normalization method and two square normalization method generations.Particularly, normalized image 502 and the normalized image 503 by in input picture 501, using two square normalization methods to generate that expression generates by use the square normalization method in the input picture 501 of normalized, input picture 501 in Fig. 5.

As mentioned above, to have high recognition capability be known to normalization method such as moments method and constant moments method.But, because these methods directly use the pixel value of original image to come the computing square, so influence the thickness degree of character stroke easily.Therefore, according to the thickness degree difference of literal, the value of square differs widely, and therefore, the position of the literal in the normalized image is according to the thickness degree of literal and difference.

Fig. 6 is the key diagram of example of the different fonts of same literal, particularly, and the image 601 of the literal " T " of expression different fonts.As shown in Figure 6, judging on the literal that the thickness degree of literal is not essence.Therefore, by the difference of the thickness degree of literal and the deviation (position of literal or size etc.) of the literal of the normalized image between the same literal classification that produces is disadvantageous to identification.

Contour feature amount square normalization method (patent documentation 1, non-patent literature 4, non-patent literature 5) is to extract the profile of literal and carry out normalized method based on the square of text profile.Be effectively in order to reduce the deviation the method that results between the literal of the length of literal and thickness degree, and in the identification of font literal, have high discrimination.

Figure 10 is the key diagram that carries out the image after the normalization with square normalization method and contour feature amount square normalization method (contour feature moment normalization method).

The character image of a plurality of " T " that the thickness degree that illustrative original image 1001 is respectively horizontal line among Figure 10 is different.And, the image 1002 after expression uses the square normalization method with these original image 1001 normalization.As can be known, in normalized image 1002, along with the horizontal line chap, the skew of the top of the middle mind-set T of image, and the whole decline downwards in the position of literal.In addition, in normalized image 1002, in original image 1001, also produce deviation in the thickness degree of the ordinate of the T of identical thickness degree.Such deviation shows as the deviation of the vector point on the feature space after the feature extraction, and is the reason that discrimination reduces.Relative therewith, in the image 1003 of contour feature amount square normalization method after with original image 1001 normalization, these deviations have reduced.

The prior art document

Patent documentation

Patent documentation 1: TOHKEMY 2010-108113 communique

Non-patent literature

Non-patent literature 1:Mohammed Cheriet, Nawwaf Kharma, Cheng lin Liu, and Ching Suen, " Character Recognition Systems:A Guide for Students and Practitioners ", Wiley-Interscience, 2007.

Non-patent literature 2: Shi Jing is good for a youth, goes up Tian Xiugong, Maeda Hide works, village Lai Yang, " figure identification ", ohm bureau of publication of company, in August, 1998

Non-patent literature 3:Cheng-Lin Liu, Kazuki Nakashima, Hiroshi Sako, and Hiromichi Fujisawa, " Handwritten digit recognition:investigation of normalization and feature extraction techniques ", Pattern Recognition, Vol.37, No.2, pp.265_279,2004.

Non-patent literature 4:Toshinori Miyoshi, Takeshi Nagasaki, and Hiroshi Shinjo, " Character Normalization Methods using Moments of Gradient Features and Normalization Cooperated Feature Extraction ", Proceedings of the 2009 Chinese Conference on Pattern Recognition and the First CJK Joint Workshop on Pattern Recognition, pp.934-938,2009.

5: three good sharp Noboru of non-patent literature, rugged strong forever, Xin Zhuanguang, " using the literal normalization method of the square of gradient characteristic quantity ", and electronic information communication association technical research report, PRMU, figure identification/medium are understood 108(432), pp.187-192,2009.

Contour feature amount square normalization method is to extract the profile of literal and carry out normalized method based on the square value of the text profile portion that extracts.The method produces effect for the thickness degree that suppresses literal or the deviation of length, and is special effective in the identification of font literal.But in the font literal of handwriting and a part, the profile of literal is lost sometimes.

Figure 13 is the key diagram of example that the character image of a profile part is lost in expression.Character image 1301 and 1302 shown in Figure 13 all is the handwriting image of literal " Agencies ".On the other hand, character image 1303 and 2304 is respectively the image of the profile that extracts from character image 1301 and 1302.In character image 1302, because the wearing and tearing of literal, the part of profile has disappeared.In this case, contour feature amount square becomes unstable.

Summary of the invention

Following is a representational example of the present invention.That is, a kind of character recognition device is characterized in that, has: arithmetic unit, contain processor and memory storage; Input media is connected on the above-mentioned arithmetic unit; And output unit, be connected on the above-mentioned arithmetic unit, above-mentioned arithmetic unit is carried out following steps: first step, according to input images stored in the input picture of importing by above-mentioned input media or the above-mentioned memory storage, carry out the pre-service for reducing the disturbing factor that becomes literal identification obstruction; Second step is carried out normalization to having carried out above-mentioned pretreated image; Third step is transformed to vector value on the vector space with the image after the above-mentioned normalization; The 4th step judges that based on the recognition dictionary of storing in the above-mentioned memory storage which literal above-mentioned vector value is; The 5th step is exported the result of above-mentioned judgement by above-mentioned output unit, and above-mentioned second step contains: the 6th step, extract the profile of having carried out above-mentioned pretreated image; The 7th step, the image of having carried out the profile that above-mentioned pretreated image and said extracted arrive is synthetic; The 8th step, with the picture of the center of gravity of the above-mentioned image that synthesizes near the picture of the scope of the pixel-expansion of the center of the scope of afore mentioned rules size and the above-mentioned image that the synthesizes mode near the scope of afore mentioned rules size, the mapping of the image after generating from the above-mentioned image that synthesizes to the normalization of above-mentioned given size; And the 9th step, carry out normalization according to the above-mentioned mapping that generates to having carried out above-mentioned pretreated image.

The effect of invention

According to an embodiment of the present invention, carry out normalization by the composograph based on text profile image and original image, the normalized instability in the time of can reducing text profile and lose can improve the discrimination in font and the handwriting.

Description of drawings

Fig. 1 is the block diagram of an example of hardware configuration of the character recognition device of expression embodiments of the present invention.

Fig. 2 is the process flow diagram of representing the processing of character recognition device execution in the past.

Fig. 3 is that expression is by the process flow diagram of the summary of the literal identification processing of the arithmetic unit execution of embodiments of the present invention.

Fig. 4 is the key diagram of being handled by the identification that the arithmetic unit of embodiments of the present invention is carried out.

Fig. 5 is the key diagram with the example of the image of square normalization method and two square normalization method generations.

Fig. 6 is the key diagram of example of the different fonts of same literal.

Fig. 7 is with the center of gravity of the character image of square normalization method decision and the key diagram on border.

Fig. 8 is the key diagram by first example of the extracting method of the text profile of the arithmetic unit use of embodiments of the present invention.

Fig. 9 is the key diagram of the pixel of reference for the profile that extracts character image in embodiments of the present invention.

Figure 10 is the key diagram with square normalization method and the normalized image of contour feature amount square normalization method.

Figure 11 is the key diagram of the filtrator that uses for the profile that extracts character image in embodiments of the present invention.

Figure 12 is the key diagram by the example of the contour images of the arithmetic unit extraction of embodiments of the present invention.

Figure 13 is the key diagram of example of the literal of an expression part of losing profile.

Symbol description

100 character recognition devices

101 input medias

102 arithmetic units

103,214 recognition dictionaries

104 display device

105 D graphics B

201 character image input parts

202 pretreatment portions

203,301 normalization portions

204,302 text profile extraction units

206,304 square value operational parts

207 normalized mapping generating units

208 normalized image generating units

209 feature extraction portions

210 identification parts

211 efferents

212 character image DB

213 recognition dictionary study portions

303 composograph generating units

Embodiment

Fig. 1 is the block diagram of an example of hardware configuration of the character recognition device of expression embodiments of the present invention.

Character recognition device 100 of the present invention has: input media 101, arithmetic unit 102, recognition dictionary 103, display device 104 and graphic data base (DB) 105.

Input media 101 is for the keyboard of input command etc. or mouse and the devices such as scanner that are used for the image input.

Arithmetic unit 102 reads the image of importing, and judges the literal in the input picture.CPU (central processing unit)), internal memory and memory storage etc. arithmetic unit 102 has: CPU(Central Processing Unit:.

Recognition dictionary 103 is dictionary databases of preserving recognition dictionary.

Display device 104 is devices of the contents processing of output arithmetic unit 102, for example is devices such as display.When not needing to show contents processing, can not have display device 104 yet, can replace with display device output unit in addition as required yet.

The figure that D graphics B105 storage is imported with input media 101.

Also can store recognition dictionary 103 and D graphics B105 in the memory storage in arithmetic unit 102.

The arithmetic unit 102 of embodiments of the present invention has: word recognition unit.Particularly, for example, realize word recognition unit by program stored in the execution internal memory of the CPU in the arithmetic unit 102 or the memory storage.

Next, transfer to the explanation of the treatment scheme in the embodiments of the present invention.

Fig. 3 is that expression is by the process flow diagram of the summary of the literal identification processing of arithmetic unit 102 execution of embodiments of the present invention.

Character image input part 201 shown in Figure 3, pretreatment portion 202, normalization portion 301, feature extraction portion 209, identification part 210, efferent 211 and recognition dictionary study portion 213 are by arithmetic unit 102(namely, carry out program stored in the internal memory etc. by CPU) in other words the function that realizes is exactly to be equivalent to the treatment step carried out by arithmetic unit 102 respectively.The text profile extraction unit 302 that normalization portion 301 is contained, composograph generating unit 303, square value operational part 304, normalized mapping generating unit 207 and normalized image generating unit 208 also are same.

Character recognition device 100 reads the image of importing, and judges the literal in the input picture, and the output result of determination.As described in illustrating, Fig. 2 is the process flow diagram that the literal identification of contour feature amount square normalization method is in the past handled.In the middle of the literal identification that the character recognition device 100 of present embodiment is carried out was handled, the text profile extraction unit 302 in the normalization portion 301 and the processing of composograph generating unit 303 were handled different with literal identification in the past.

In character image input part 201, by the user or become the image of identifying object by the program input of being carried out by arithmetic unit 102.For example, the scanner that input media 101 contains reads text, and the data of the character image that arithmetic unit 102 will obtain thus are stored in internal memory or the memory storage.In addition, when in memory storage etc., storing the data of character image in advance, also it can be used as identifying object.

Pretreatment portion 202 is by implementing denoising, Fuzzy Processing etc. to input picture, can reduce the disturbing factor that noise or the literal in the process decision chart picture such as fuzzy become obstacle.For example, remove the isolated point of certain size below the threshold value with denoising.Also can temporarily be stored in the memory storage having implemented pretreated input picture.

Normalization portion 301 implements the image that pretreated input picture is transformed to preassigned fixed size with each.Image after the conversion is called normalized image.One of normalized fundamental purpose is, is transformed to the image of fixed measure, the processing after unifying by the input picture with various sizes.In addition, normalized another fundamental purpose is, the input picture of different shape is transformed to the image of fixed measure for the deviation that makes literal shape between same literal diminishes.Thus, can reduce the deviation between the character image of same literal type, and help the raising of accuracy of identification.Describe aftermentioned in detail.Also the normalized image that generates with normalization portion 301 temporarily can be stored in the memory storage.

Feature extraction portion 209 will be received as input by the normalized image that normalization portion 301 generates, and the normalized image of importing is transformed to vector value on the vector space.The vector space of change target is called feature space, and the vector value after the conversion is called eigenvector.Sometimes also cut down the dimension of feature space by the dimension compression.At this moment, as far as possible remove the little composition of contribution to identification from feature space, and eigenvector is showed as the eigenvector on the feature space of low-dimensional more.

Identification part 210 usefulness recognition dictionaries 214 are judged the literal classification under the eigenvector.Recognition dictionary 214 is kept for feature space is divided into the information in the shared zone of each literal classification.Thus, will with eigenvector under regional corresponding literal classification return as result of determination.

Fig. 4 is the key diagram of being handled by the identification that the arithmetic unit 102 of embodiments of the present invention is carried out.

As an example, in Fig. 4, be illustrated in classification A, classification B and shared regional 402A, 402B and the 402C of classification C difference in the feature space 401.Of all categories corresponding to a literal.In this example, unknown input (being the eigenvector of the normalized image imported) 403 is not included in the zone of arbitrary classification.At this moment, identification part 210 also can be judged as the classification A corresponding with the regional 402A of the most approaching unknown input 403 the affiliated classification of unknown input.Perhaps, identification part 210 also can be judged as unknown input 403 and not belong to arbitrary classification, makes the judgement of abandoning.The result (for example " classification A " or " abandoning ") that identification part 210 outputs are judged.

Referring again to Fig. 3.Efferent 211 is to the result of determination of output identification parts 210 such as display device such as display 104 or memory storage.

Next, before the explanation of the processing of transferring to normalization of the present invention portion 301, the processing of the normalization portion 203 of contour feature amount square normalization method is described.

If through pretreatment portion 202 and be input to the original image f of text profile extraction unit 204 (x, size y) is width W 0, height H 0.At this, 0≤x＜W0,0≤y＜H0, the x and the y that establish each net point of expression are round valuess, the pixel value of several k1 on a left side, following several k2 net points is expressed as f(k1-1, k2-1).Explanation is normalized into this original image the example of the image size of width L, height L.

When using contour feature amount square normalization method, at first, text profile extraction unit 204 is extracted original image f(x, the contour images fc(x of literal y), y).Below enumerate two examples of the extracting method of profile.

Enumerate first example of the extracting method of text profile.At first, according to character image f(x, y) extract the horizontal ingredient f x(x of profile, y) and vertical ingredient f y(x, y).

Fig. 8 is the key diagram by first example of the extracting method of the text profile of the arithmetic unit use of embodiments of the present invention.

Input picture 801, contour images 802, horizontal contour images 803 in Fig. 8, have been represented and contour images 804 is as an example longitudinally.At this, input picture 801 is images of literal " B ", contour images 802, horizontal contour images 803 and longitudinally contour images 804 all are examples of the contour images that extracts from input picture 801.Input picture 801 is corresponding to f(x, and y), horizontal contour images 803 is corresponding to fx(x, and y), contour images 804 is corresponding to fy(x, y) longitudinally.

At first, text profile extraction unit 204 is set fx(x, y)=0, fy(x, and y)=0.Next, text profile extraction unit 204 is selected input picture f(x in order, net point y), and extract the feature of contour direction for each net point.Vergence direction is at the counting of both direction in length and breadth.When the pixel of the net point of present concern is black pixel, i.e. p=f(x, y)=1 o'clock, feature is extracted near the information of the pixel that is arranged in pixel p shown in Figure 9 in text profile extraction unit 204 usefulness formula (1)～(3).

[formula 1]

Vergence direction

$\left.\begin{array}{c}\mathrm{fori}=\mathrm{0,2,4,6},\\ \mathrm{if}{d}_i=0\mathrm{and}{d}_{i+1}=1,\\ \mathrm{then}{f}_x(x,y)=f_x(x,y)+1,\\ f_y(x,y)=f_y(x,y)+1\\ \mathrm{endif}\\ \mathrm{endfor}.\end{array}\right\}---\left(1\right)$

[formula 2]

Vertically

$\left.\begin{array}{c}\mathrm{fori}=\mathrm{0,4}\\ \mathrm{if}{d}_i=0,d_{i+1}=0\mathrm{and}{d}_{(i+2)\%8}=1,\\ \mathrm{then}{f}_y(x,y)=f_y(x,y)+1\\ \mathrm{endif}.\\ \mathrm{endfor}.\end{array}\right\}---\left(2\right)$

[formula 3]

Laterally

$\left.\begin{array}{c}\mathrm{fori}=\mathrm{2,6}\\ \mathrm{if}{d}_i=0,d_{i+1}=0\mathrm{and}{d}_{(i+2)\%8}=1,\\ \mathrm{then}{f}_x(x,y)=f_x(x,y)+1\\ \mathrm{endif}\\ \mathrm{endfor}.\end{array}\right\}---\left(3\right)$

Fig. 9 is the key diagram of the pixel of reference for the profile that extracts character image in embodiments of the present invention.

Particularly, in Fig. 9 the pixel p of certain net point of expression with and its adjacency around the position of pixel d1 to d7 of net point concern 901.For example, when the coordinate of the net point of pixel p is (x, y) time, the coordinate of the net point of pixel d1, d2, d3, d4, d5, d6 and d7 be respectively (x+1, y+1), (x, y+1), (x-1, y+1), (x-1, y), (x-1, y-1), (x, y-1) and (x+1, y-1).

Thus, generate the horizontal ingredient f x(x of profile, y) and vertical ingredient f y(x, y).By calculating fc(x, y)=fx(x, y)+fy(x, y) (this=be substitution) obtain contour images fc(x, y).

Enumerate second example of the extracting method of text profile.At first, text profile extraction unit 204 be set at fx (x, y)=0, fy(x, y)=0.Next, text profile extraction unit 204 is selected input picture f(x in order, net point y), and for each net point extraction feature.Vergence direction is at the counting of both direction in length and breadth.(x y) extracts feature to text profile extraction unit 204 usefulness formula (4) to net point.

[formula 4]

$\left.\begin{array}{c}{f}_x(x,y)=f(x+1,y-1)+2f(x+1,y)\\ +f(x+1,y+1)-f(x-1,y-1)\\ -2f(x-1,y)-f(x-1,y+1),\\ f_y(x,y)=f(x-1,y+1)+2f(x,y+1)\\ +f(x+1,y+1)-f(x-1,y-1)\\ -2f(x,y-1)-f(x+1,y-1)\end{array}\right\}---\left(4\right)$

Figure 11 is the key diagram of the filtrator that uses for the profile that extracts character image in embodiments of the present invention.The filtrator 1101 of Figure 11 is corresponding to the fy(x of formula (4), computing formula y), and filtrator 1102 is corresponding to fx(x, computing formula y).

By calculating fc(x, y)=fx(x, y)+fy(x, y) (this=be substitution) obtain contour images fc(x, y).

Figure 12 is the key diagram by the example of the contour images of arithmetic unit 102 extractions of embodiments of the present invention.For example, from literal type “ Aya ” “ Catfish " " or " original image 1201 of the handwriting of " grain " and " making rich " extracts contour images 1202.

Next, square value operational part 206 calculates contour images fc(x, square value y).At this, with the center of gravity shown in formula (5) and formula (7) computing formula (6) (xc, yc) and the δ x shown in the formula (8), the value of δ y.This δ x and δ y are the parameters of scope of the pixel-expansion of expression original image, use for the border that determines original image described later.

[formula 5]

m _pq＝∑ _x∑ _yx ^py ^qf _c(x，y) …(5)

[formula 6]

x _c＝m ₁₀/m ₀₀， y _c＝m ₀₁/m ₀₀ …(6)

[formula 7]

$\left.\begin{array}{c}{\mathrm{\μ}}_{20}={\mathrm{\Σ}}_x{\mathrm{\Σ}}_y{(x-x_c)}^2{f}_c(x,y),\\ {\mathrm{\μ}}_{02}={\mathrm{\Σ}}_x{\mathrm{\Σ}}_y{(y-y_c)}^2{f}_c(x,y).\end{array}\right\}---\left(7\right)$

[formula 8]

${\mathrm{\δ}}_x=\mathrm{\α}\sqrt{{\mathrm{\μ}}_{20}/m_{00}},$ ${\mathrm{\δ}}_y=\mathrm{\α}\sqrt{{\mathrm{\μ}}_{02}/m_{00}},\·\·\·\left(8\right)$

Next, normalized mapping generating unit 207 generates the mapping that is used for original image is painted into normalization plane ［ 0, L ］ * ［ 0, L ］.In contour feature amount square normalization method, (xc has horizontal δ x centered by yc), vertically the zone of the width of δ y enlarges or the size that dwindles into L * L generates normalized image to the center of gravity by will calculating with square value operational part 206.That is, shine upon the part of ［ xc-δ x/2, xc+ δ x/2 ］ * ［ yc-δ y/2, the yc+ δ y/2 ］ in the original image to normalization plane ［ 0, L ］ * ［ 0, L ］.Represent the mapping that this is used with formula (9).

[formula 9]

$\left.\begin{array}{c}u\left(x\right)=L(x-x_c)/{\mathrm{\δ}}_x+L/2,\\ v\left(y\right)=L(y-y_c)/{\mathrm{\δ}}_y+L/2.\end{array}\right\}---\left(9\right)$

Next, the relation formula of normalized image generating unit 208 usefulness formula (10) generation normalized image f ' (x ', y ').Under this routine situation, as, obtain normalized image by the part of ［ xc-δ x/2, xc+ δ x/2 ］ in the original image * ［ yc-δ y/2, yc+ δ y/2 ］ is enlarged the size that dwindles into L * L.

[formula 10]

$\left.\begin{array}{c}{f}^{\′}(x^{\′},y^{\′})=f(x,y),\\ x^{\′}=u\left(x\right),\\ y^{\′}=v\left(y\right).\end{array}\right\}---\left(10\right)$

As mentioned above, contour feature amount square normalization method y) is extracted contour images fc(x from original image f(x, y), and uses contour images fc(x, and square y) determines center of gravity and the border of character image.

When as in the past, using the square normalization method namely based on the normalization method of the square value of original image itself, with the picture of the center of gravity of the pixel of the original image center near the scope of normalized image, and the picture of the scope of the pixel-expansion of original image generates mapping from the original image to the normalized image near the mode of the scope of normalized image.

Fig. 7 is with the center of gravity of the character image of square normalization method decision and the key diagram on border.

Particularly, expression pretreated image (being the original image in the above-mentioned explanation) 701 and character image 702 in Fig. 7, character image 702 comprise the center of gravity that determines based on pretreated image and the demonstration on border.For example, according to literal type " 0 " corresponding original image 701A determines center of gravity 703A and border 704A.At this, border 704A is among the original image 701A, shows the zone of the literal that is equivalent to literal type " 0 " and the border in the zone beyond it, in other words is exactly the scope that is equivalent to the pixel-expansion of the literal suitable with literal type " 0 ".When using the square normalization method, use the second moment value δ x that calculates with formula (8), δ y is as the parameter that the scope of the pixel-expansion of literal is represented, in order to have horizontal δ x centered by the center of gravity 703A, vertically the zone of the width of δ y defines border 704A.

By using the mapping that generates as described above to carry out normalization, even when the size of the character image of importing and shape have deviation, if the deviation of these to be images of same literal type then can expect suppress characteristic quantities of normalized character image.

But, when using aforesaid square normalization method, easily according to the change of the thickness degree of the line of the character image of importing, produce the deviation of the normalized image shown in the normalized image 1002 of Figure 10.This be because, the position of the center of gravity of the pixel of original image is subjected to the influence of thickness degree of line of literal of original image and change etc., so the square value becomes unstable, the mapping of Sheng Chenging also changes thus.

Relative therewith, when using contour feature amount square normalization method (when carrying out the normalization based on the square value of contour images), with the picture of the center of gravity of the pixel of the profile of the original image center near the scope of normalized image, and make the picture of scope of pixel-expansion of profile of original image near the mode of the scope of normalized image, generate the mapping from the original image to the normalized image.At this moment, owing to deleted the pixel of the part beyond the profile in the middle of the original image, so the position of the center of gravity of the pixel of profile is not subject to the influence of thickness degree of line of the literal of original image.Therefore, the thickness degree of the mapping of square value and generation and the line of literal is irrespectively stable, and deviation such shown in the normalized image 1003 of Figure 10, normalized image becomes and is difficult to produce.

But shown in the example of the character image 1302 of Figure 13 and 1304, when the structure of the profile of literal was lost, the part of profile can't be extracted.Because thereby the part of profile is lost the position change of center of gravity of the pixel of profile, thus when the structure of profile is lost, the value of the square that the calculates instability that becomes, it is big that the deviation of the normalized image of generation becomes between same literal type.Such deviation shows as the deviation of the vector point on the feature space after the feature extraction, and is the reason that discrimination reduces.

Next, the normalization of normalization portion 301 execution of embodiments of the present invention is described.

Text profile extraction unit 302 can be used with the same method (for example above-mentioned first or second example) of text profile extraction unit 204 and extract text profile image fc(x, y), and also can be with other method.At this, other the example of method as extracting text profile illustrates the 3rd and the 4th example.

Initial explanation the 3rd example.At first, text profile extraction unit 302 y) is made as g0(p for whole white pixel p=(x)=g1(p)=...=g7(p).Next, text profile extraction unit 302 is for all black pixel p=(x is y) with formula (11) computing g0(p), g1(p) ..., g7(p).

[formula 11]

$\left.\begin{array}{c}{g}_{k+1}\left(p\right)=\left\{\begin{array}{cc}1& \mathrm{iff}\left(d_k\right)=0\mathrm{andf}\left(d_{k+1}\right)=1\\ 0& \mathrm{otherwise}\end{array}\right.\\ g_{(k+2)\%8}\left(p\right)=\begin{array}{}\end{array}\left\{\begin{array}{cc}1& \mathrm{iff}\left(d_k\right)=f\left(d_{k+1}\right)=0\\ & \mathrm{andf}\left(d_{(k+2)\%8}\right)=1\\ 0& \mathrm{otherwise}\end{array}\right.\end{array}\right\}---\left(11\right)$

As shown in Figure 9, d0, d1 ..., d7 is near the pixel of pixel p.Text profile extraction unit 302 is used fc(x, y)=Σ gk(x, y) generate contour images fc(x, y).At this, k=0,1 ..., computing Σ gk(x in 7 the scope, y).

Next the 4th example is described.At first, text profile extraction unit 302 is for whole pixel p=(x is y) with formula (12) computing gx(p), gy(p).Next, text profile extraction unit 302 usefulness formula (13) generate contour images fc(x, y).At this, as shown in Figure 9, d0, d1 ..., d7 is near the pixel of pixel p.

[formula 12]

$\left.\begin{array}{c}{g}_x\left(p\right)=[f\left(d_1\right)+2f\left(d_0\right)+f\left(d_7\right)\\ -f\left(d_3\right)-2f\left(d_4\right)-f\left(d_5\right)]/8,\\ g_y\left(p\right)=[f\left(d_1\right)+2f\left(d_2\right)+f\left(d_3\right)\\ -f\left(d_5\right)-2f\left(d_6\right)-f\left(d_7\right)]/8,\end{array}\right\}---\left(12\right)$

[formula 13]

$f_c(x,y)=\sqrt{g_x^2(x,y)+g_y^2(x,y)}\·\·\·\left(13\right)$

Above-mentioned first～the 4th example is to extract the example of method of the profile of character image, and text profile extraction unit 302 also can be extracted the profile of character image with the method beyond the above-mentioned illustrated method.As mentioned above, can extract the profile of character image in order to following method: when the pixel value of the net point around the net point of original image satisfies defined terms, increase the pixel value (being equivalent to above-mentioned first example and the 3rd example) of the image of the profile in the net point of this original image, perhaps multiply by the pixel value (being equivalent to above-mentioned second example and the 4th example) etc. of the image of the profile in the net point that value behind the coefficient of regulation added up to calculate this original image by the pixel value to the net point around the net point of original image.

So far, the explanation of text profile extraction unit 302 is finished, next, the processing after the synthetic image production part 303 is illustrated.Composograph generating unit 303 usefulness formula (14) are created in the text profile extraction unit 302 the text profile image fc(x of each net point that generates, y) with original image f(x from each net point of pretreatment portion 202 outputs, and composograph fs(x y), y).

[formula 14]

f _s(x，y)＝γ ₁f(x，y)+γ ₂f _c(x，y) …(14)

At this, γ 1 and γ 2 are positive numbers, and satisfy γ 1+ γ 2=1.This composograph is the image of emphasizing the original image outline portion, in other words, is equivalent to come than the big mode of pixel value of the part beyond it with the pixel value of original image outline portion the image of weighting exactly.

Square value operational part 304 is used fs(x, y) replaces fc(x, y) calculates the square value.That is, square value operational part 206 usefulness formula (15) replace formula (5) come the center of gravity shown in the computing formula (6) (xc, yc) and the value of the δ x shown in the formula (8), δ y.

[formula 15]

m _pq＝∑ _x∑ _yx ^py ^qf _s(x，y) …(15)

Next, normalized mapping generating unit 207 of the present invention generates normalized mapping based on the square value that calculates with formula (15) etc., and the normalized mapping of normalized image generating unit 208 usefulness of the present invention generation generates normalized image (formula (10)).

In the embodiment of the invention described above, the scope (being the border) that decides the pixel-expansion of composograph based on second moment value δ x and the δ y of composograph.This scope is not necessarily consistent with the profile of the pixel of composograph.But, decide an only example of method based on the Decision of the scope of above-mentioned square value, in the present invention, also can decide the scope of the pixel-expansion of character image with the method beyond above-mentioned.For example, arithmetic unit 102 also can replace calculating the square value in square value calculating part 304, and will decide with the scope of the circumscribed rectangle scope of the profile of the pixel of composograph as the pixel-expansion of composograph.

Explanation so far relates to character recognition device 100, uses but character recognition device 100 also can be used as the recognition dictionary generating apparatus.At this moment, the memory storage of arithmetic unit 102 keeps character image DB212(Fig. 3), and pretreatment portion 202 carries out pre-service for the character image of storing among the character image DB212.The processing of normalization portion 301 and feature extraction portion 209 and above-mentioned character recognition device 100 are same.Recognition dictionary study portion 213 carries out the study of recognition dictionary based on the characteristic quantity that is extracted by feature extraction portion 209, and stores its result into recognition dictionary 103 that recognition dictionary 214(is equivalent to Fig. 1).In addition, with normalization portion 301 grades similarly, recognition dictionary study portion 213 is the functions that realize with arithmetic unit 102.

As mentioned above, by embodiments of the present invention, carry out the normalization based on the square value of the composograph of original image and contour images.That is, calculate the square value of composograph, and based on this mapping of generation from the original image to the normalized image.By synthetic, it is bigger than the pixel value of the part beyond it that the pixel value of the outline portion of character image becomes.Consequently, compare during with the normalization carried out based on the square value of original image itself, owing to increased the weight of the pixel of outline portion, so can alleviate the influence of thickness degree of the line of literal, and, compare during with the normalization carried out based on the square value of contour images, owing to also utilized the pixel of the part beyond the profile, can alleviate the influence that profile disappears.Like this, according to present embodiment, even each in disappearing for the thickness degree of line and profile also can be realized stable normalization, thus, can improve the discrimination of font and handwriting.

In addition, in order to maximize above-mentioned effect, expectation is carried out optimization to coefficient gamma 1 and γ 2.Optimum coefficient gamma 1 and the value of γ 2 can exist with ... the various conditions such as extracting method of profile, but the pixel value of the outline portion of the character image in the image that needs to select to be synthesized is than the big value of pixel value of the part beyond it.For example, the composograph generating unit 303 of present embodiment also can be used γ 1 and the γ 2 that satisfies γ 1＜γ 2.

Claims (15)

1. character recognition device is characterized in that having:

Arithmetic unit contains processor and memory storage;

Input media is connected on the above-mentioned arithmetic unit; And

Output unit is connected on the above-mentioned arithmetic unit,

Above-mentioned arithmetic unit is carried out following steps:

First step according to input images stored in the input picture of importing by above-mentioned input media or the above-mentioned memory storage, is carried out the pre-service for reducing the disturbing factor that becomes literal identification obstruction;

Second step is carried out normalization to having carried out above-mentioned pretreated image;

Third step is transformed to vector value on the vector space with the image after the above-mentioned normalization;

The 4th step based on the recognition dictionary of storing in the above-mentioned memory storage, judges that above-mentioned vector value is any literal; And

The 5th step is exported the result of above-mentioned judgement by above-mentioned output unit,

Above-mentioned second step contains:

The 6th step is extracted the profile of having carried out above-mentioned pretreated image;

The 7th step, the image of having carried out the profile that above-mentioned pretreated image and said extracted arrive is synthetic;

The 8th step, with the picture of the center of gravity of the above-mentioned image that synthesizes near the picture of the scope of the pixel-expansion of the center of the scope of given size and the above-mentioned image that the synthesizes mode near the scope of afore mentioned rules size, the mapping of the image after generating from the above-mentioned image that synthesizes to the normalization of above-mentioned given size; And

The 9th step according to the above-mentioned mapping that generates, is carried out normalization to having carried out above-mentioned pretreated image.

2. character recognition device as claimed in claim 1 is characterized in that,

Above-mentioned the 7th step has following steps: by will be to the execution in each net point the pixel value of above-mentioned pretreated image multiply by value addition after value and pixel value to the image of the above-mentioned profile in each net point behind first coefficient multiply by second coefficient, calculate the pixel value of the above-mentioned image that synthesizes in each net point.

3. character recognition device as claimed in claim 1 is characterized in that,

Above-mentioned second step also contains and has the following steps: calculate the square value of the above-mentioned image that synthesizes as the parameter of the scope of the pixel-expansion of the above-mentioned image that synthesizes of expression,

Above-mentioned the 8th step contains and has the following steps: generate the mapping that enlarges or dwindle the above-mentioned image that synthesizes according to above-mentioned square value.

4. character recognition device as claimed in claim 1 is characterized in that,

Above-mentioned the 6th step contains and has the following steps: when the pixel value of the net point around each net point of having carried out above-mentioned pretreated image satisfies defined terms, the pixel value of the image of the profile in above-mentioned each net point is increased.

5. character recognition device as claimed in claim 1 is characterized in that,

Above-mentioned the 6th step contains and has the following steps: added up to by the value behind the coefficient that will multiply by regulation to the pixel value of the net point around each net point of having carried out above-mentioned pretreated image, calculate the pixel value of the image of the profile in above-mentioned each net point.

6. recognition dictionary generating apparatus is characterized in that having:

Processor; And

Memory storage is connected on the above-mentioned processor, and stores character image,

Above-mentioned arithmetic unit is carried out following steps:

First step is according to the pre-service of the character image execution of storing in the above-mentioned memory storage for reducing the disturbing factor that becomes literal identification obstruction;

Second step is carried out normalization to having carried out above-mentioned pretreated image;

Third step is transformed to vector value on the vector space with the image after the above-mentioned normalization;

The 4th step is learnt the recognition dictionary of use in literal identification based on above-mentioned vector value; And

The 5th step stores the result of above-mentioned study into above-mentioned memory storage,

Above-mentioned second step contains:

The 6th step is extracted the profile of having carried out above-mentioned pretreated image;

The 7th step, the image of having carried out the profile that above-mentioned pretreated image and said extracted arrive is synthetic;

The 8th step, with the picture of the center of gravity of the above-mentioned image that synthesizes near the picture of the scope of the pixel-expansion of the center of the scope of given size and the above-mentioned image that the synthesizes mode near the scope of afore mentioned rules size, the mapping of the image after generating from the above-mentioned image that synthesizes to the normalization of above-mentioned given size; And

The 9th step according to the above-mentioned mapping that generates, is carried out normalization to having carried out above-mentioned pretreated image.

7. recognition dictionary generating apparatus as claimed in claim 6 is characterized in that,

Above-mentioned the 7th step contains and has the following steps: by will be to the execution in each net point the pixel value of above-mentioned pretreated image multiply by value addition after value and pixel value to the image of the above-mentioned profile in each net point behind first coefficient multiply by second coefficient, calculate the pixel value of the above-mentioned image that synthesizes in each net point.

8. recognition dictionary generating apparatus as claimed in claim 6 is characterized in that,

Above-mentioned second step also contains and has the following steps: calculate the square value of the above-mentioned image that synthesizes as the parameter of the scope of the pixel-expansion of the above-mentioned image that synthesizes of expression,

Above-mentioned the 8th step contains and has the following steps: generate the mapping that enlarges or dwindle the above-mentioned image that synthesizes according to above-mentioned square value.

9. recognition dictionary generating apparatus as claimed in claim 6 is characterized in that,

Above-mentioned the 6th step contains and has the following steps: when the pixel value of the net point around each net point of having carried out above-mentioned pretreated image satisfies defined terms, the pixel value of the image of the profile in above-mentioned each net point is increased.

10. recognition dictionary generating apparatus as claimed in claim 6 is characterized in that,

Above-mentioned the 6th step contains and has the following steps: added up to by the value behind the coefficient that will multiply by regulation to the pixel value of the net point around each net point of having carried out above-mentioned pretreated image, calculate the pixel value of the image of the profile in above-mentioned each net point.

11. a method for normalizing is carried out by arithmetic unit, this arithmetic unit contains processor and is connected memory storage on the above-mentioned processor,

This method for normalizing is characterised in that to have:

The 6th step, above-mentioned arithmetic unit extracts the profile of the original image of storing in the above-mentioned memory storage;

The 7th step, above-mentioned arithmetic unit is synthetic with the image of the profile that above-mentioned original image and said extracted arrive;

The 8th step, above-mentioned arithmetic unit with the picture of the center of gravity of the above-mentioned image that synthesizes near the picture of the scope of the pixel-expansion of the center of the scope of given size and the above-mentioned image that the synthesizes mode near the scope of afore mentioned rules size, the mapping of the image after generating from the above-mentioned image that synthesizes to the normalization of above-mentioned given size; And

The 9th step, above-mentioned arithmetic unit carries out normalization according to the above-mentioned mapping that generates to above-mentioned original image, and stores its result into above-mentioned memory storage.

12. method for normalizing as claimed in claim 11 is characterized in that,

Above-mentioned the 7th step contains and has the following steps: by will be to the execution in each net point the pixel value of above-mentioned pretreated image multiply by value addition after value and pixel value to the image of the above-mentioned profile in each net point behind first coefficient multiply by second coefficient, calculate the pixel value of the above-mentioned image that synthesizes in each net point.

13. method for normalizing as claimed in claim 11 is characterized in that,

Above-mentioned method for normalizing also has following steps: calculate the square value of the above-mentioned image that synthesizes as the parameter of the scope of the pixel-expansion of the above-mentioned image that synthesizes of expression,

Above-mentioned the 8th step contains and has the following steps: generate the mapping that enlarges or dwindle the above-mentioned image that synthesizes according to above-mentioned square value.

14. method for normalizing as claimed in claim 11 is characterized in that,

Above-mentioned the 6th step contains and has the following steps: when the pixel value of the net point around each net point of having carried out above-mentioned pretreated image satisfies defined terms, the pixel value of the image of the profile in above-mentioned each net point is increased.

15. method for normalizing as claimed in claim 11 is characterized in that

Above-mentioned the 6th step contains and has the following steps: added up to by the value behind the coefficient that will multiply by regulation to the pixel value of the net point around each net point of having carried out above-mentioned pretreated image, calculate the pixel value of the image of the profile in above-mentioned each net point.

Images