CN1741035A - Blocks letter Arabic character set text dividing method - Google Patents

Abstract

The present invention is characterized by that on the basis of character block classification selecting partial character blocks, making horizontal projection, dividing text image into subregions, then detecting multiple line subregion, combining character blocks and implementing literal line segmentation of multiple line subregion, then estimating baseline and top line positions of literal line, segmenting the literal line into body-conjunct character field, finally, utilizing feature of character tangent point to find the tangent points positioned on baseline, over baseline and under baseline, and utilizing structural rule to delete overtangential points. The invented character segmentation accuracy can be up to above 99%.

Description

Blocks letter Arabic character set text dividing method

Technical field

Blocks letter Arabic character set text dividing method belongs to the character cutting field in the optical character identification (OCR).

Background technology

Arabic is one of the United Nations's working language, is extensive use of in countries such as Egypt, Algeria, Morocco, Saudi Arabia.Uygur, Kazak, Kirgiz language are the important minority languages of China, though do not belong to the same family of languages with Arabic, all use Arabic character and write, and just character set is slightly different.In the present invention, the Arabic character set text written of use such as Arab, Uygur, Kazak are referred to as Arabic character set text, and brief note is the A Wen text.No matter be domestic or the world, identification A Wen text all has urgent current demand and application prospects.

In the A Wen text, Ah's Chinese character is write from right to left, and the literal of other language (for example Chinese character, English, numeral) is write from left to right.The standard Arabic has 28 base characters, and Uighur has 32 base characters.Each Ah's Chinese character has 1-4 kind written form: afterbody is write form with the head that next character is connected; The intermediate form that head and the tail are connected with adjacent character; Stem is write form with the tail that a last character is connected; Head and the tail and all disjunct absolute version of adjacent character.Each Ah's Chinese character is made up of a main part and several and the unconnected extention of main part, the position of extention can above the main part, below or the centre, the extention of standard Arabic character be 1-3 put and

The extention of Uighur also comprises

Deng.The main part of many characters is identical, only is different with the number or the shape of the disconnected extention of main part.No matter be hand-written or block letter A Wen text, the character that can connect always connects to be write or prints, and the part that character connects is called baseline.A word is made up of one or more disjunct conjuncted character fields.Fig. 1 is the base character table of standard Arabic, and Fig. 2 has reflected the Partial Feature of A Wen text.

General block letter A Wen text recognition system is as shown in Figure 3: preprocessing part strengthens the text image of input; Literal line cutting part is cut into literal line to the multiline text zone; Conjuncted character field cutting part is divided into disjunct conjuncted character field mutually to literal line; Cutting part resolves into conjuncted character field the basic elements of character such as character or stroke; Identification division extracts validity feature, the parts that utilize sorter identification to cut out; Aftertreatment partly utilizes means such as dictionary, language model, revises identification error, improves the text identification rate.

According to the relation between cutting and the identification, system can be divided into cut afterwards earlier and discern, the identification of limit cutting edge with do not cut integral body and discern three classes:

The first, cut afterwards earlier and discern (path among Fig. 3 (1) and path (2) is invalid).This type systematic at first cuts into character or stroke to conjuncted character field, the parts that obtain of identification cutting then, and to make up these parts be character identification result.These class methods are not subjected to the restriction of dictionary size, can discern the combination of arbitrary string, can handle big (infinitely) vocabulary situation.

The second, limit cutting edge identification (path among Fig. 3 (1) enabledisable all can, path (2) are effective).This type systematic must not carry out character cutting (when path 1 is invalid, then cutting in advance) before identification, but obtained the character cutting result by recognition result.The character boundary is the product of following of recognition result.These class methods depend on font, and calculated amount is big, to the deformation sensitivity.If middle certain character-recognition errors, perhaps part that character is other character, character identification result thereafter all can be affected.

The 3rd, do not cut whole identification (path among Fig. 3 (1) is effective, and path (2) are invalid).This type systematic does not need to identify each character, but the whole conjuncted character field of Direct Recognition, thereby avoided cutting.These class methods are come from the speech recognition development, can only handle little (limited) vocabulary situation.

Because general block letter A Wen text, have and have a large vocabulary, use font many, but set type the relatively characteristics of standard, be only so cut the method for afterwards discerning earlier.Therefore, design and printing body A Wen text dividing algorithm is the key that realizes block letter A Wen text recognition system.At present, the research of block letter A Wen text identification lags far behind the reason of other widely used literal (as the Latin alphabet, Chinese character, Japanese etc.), mainly with regard to the problem of the fine solution block letter A Wen text dividing that is to fail.For example: the method for horizontal projection cutting literal line can not be handled the situation that has the adhesion character between inclination text or the literal line; Boundary between a large amount of character appendix branch fuzzy literals is capable perhaps produces pseudo-literal line; The method of the conjuncted character field of vertical projection cutting can not be handled the overlapping situation of adjacent conjuncted character field level; General character cutting method is just sought the character point of contact on baseline, but in fact the part character may above the baseline or below adhesion.

Goal of the invention

Purpose of the present invention is exactly the method that a kind of reliable many font sizes of multi-font block letter A Wen text dividing will be provided.As shown in Figure 4: the present invention is on the basis of character block classification, and the cutting text image is a literal line, estimates a baseline and a line position of literal line then, and the cutting literal line becomes conjuncted character field, seeks all candidate point of contacts according to the characteristics of position, character point of contact at last.The block letter A Wen text dividing method of realizing according to the present invention has been applied to a block letter A Wen text recognition system.

Technical scheme

The invention is characterized in by following measure the block letter A Wen text image of input correctly is cut into single character picture.

1 literal line cutting

The purpose of literal line cutting is to be cut into continuous literal line text filed.Horizontal projection is the most direct capable cutting method, but can not handle the inclination text.In addition, owing to have a large amount of and the disjunct character extention of character main part in the A Wen text, so the blank gap that horizontal projection obtains both may be the blank between literal line, also may be the blank between interior character extention of literal line and the main part.Part projection and partial contour are followed the tracks of the method that combines can handle the constant text of vergence direction, but can not handle the situation of character adhesion in the ranks.

In order to address the above problem, the present invention proposes following literal line cutting method (Fig. 5): on the basis of character block classification, select part character block horizontal projection, the input text image segmentation is become subregion; Detect the multirow subregion then, the merger character block is realized the literal line cutting of multirow subregion; Cut the character block of adhesion in the ranks at last, and the small characters piece is included into affiliated literal line according to range information.This literal line cutting method not only can be handled nonangular text, can also handle inclination, there is the text of adhesion in character in the ranks.

1.1 character block classification

Make I represent the text image of importing, H is the height of I, and W is the width of I.Character block C on the I is expressed as

$C={\left[\begin{array}{cccc}C\left(\mathrm{0,0}\right)& C\left(\mathrm{0,1}\right)& L& C(0,w-1)\\ C\left(\mathrm{1,0}\right)& C\left(\mathrm{1,1}\right)& L& (1,w-1)\\ M& M& L& M\\ C(h-\mathrm{1,0})& C(h-\mathrm{1,1})& L& C(h-1,w-1)\end{array}\right]}_{[l,t,r,b]}.$

Wherein (y x)=1 represents black pixel to C, and (y x)=0 represents white pixel to C, and l, t, r, b, w and h represent left margin, coboundary, right margin, lower boundary, width and the height (Fig. 6) of C respectively.The last all character block BLOCK={C of I _NB| nB=1,2 ..., nBlock}.In the present invention, use and to have added various target C up and down and represent specific character block (C for example _NB, C ^(line)Deng), added left margin, coboundary, right margin, lower boundary, width and the height that same target l, t, r, b, w and h up and down represent this character block respectively (h for example _NB, r ^(line)Deng).

According to height and width information, can be divided into all character blocks three classes: the MIDDLE class is the general conjuncted character field of A Wen, and aspect ratio is more fixing; The BIG class comprises the character of adhesion in the ranks, and is highly the highest; The SMALL class comprises less isolated character, character extention and punctuate etc., and is highly short, and width is also short.

The present invention determines C by following formula _NBClassification,

Wherein

$\stackrel{\‾}{h}=\frac{1}{\mathrm{nBlock}}\underset{\mathrm{nB}=1}{\overset{\mathrm{nBlock}}{\mathrm{\Σ}}}{h}_{\mathrm{nB}},$

It is the average height of all character blocks.

1.2 subregion cutting

In order to remove character extention and the influence of adhesion character in the ranks, only keep MIDDLE class character block and carry out horizontal projection.Then to be 0 position be divided into a plurality of subregions text filed to the horizontal projection value.Each subregion comprises one or more literal line.

1.3 the multirow subregion detects

As shown in Figure 7, necessarily there is overlapping, the vertical nonoverlapping conjuncted character field of many groups level in the multirow subregion, and there is not the conjuncted character field group that satisfies condition in the single file subregion.

Therefore, in each subregion, find overlapping, the vertical nonoverlapping MIDDLE class character block group of all levels

CC＝{(C _I，D _J)|XOL(C _I，C _J)＞0，YOL(C _I，C _J)＝0，C _I∈MIDDLE，C _J∈MIDDLE}。

Wherein,

XOL(C _I，C _J)＝max(min(r _I，r _J)-max(l _I，l _J)，0)，

Expression C _IAnd C _JOverlap length in the horizontal direction,

YOL(C _I，C _J)＝max(min(b _I，b _J)-max(t _I，t _J)，0)，

Expression C _IAnd C _JThe overlap length of in the vertical direction.If ‖ CC is ‖＞and 10, just can judge that this subregion is the multirow subregion.

1.4 multirow subregion cutting

Because the A Wen text is write from right to left, so according to the inner all MIDDLE class character blocks of right margin descending sort multirow subregion.Take out a character block in order, the character block of the n time taking-up is expressed as C at every turn _nC _nAll character block C that taken out ₁, C ₂..., C _N-1Relatively, find overlapping maximum character block C on the vertical direction _Y, promptly

$C_Y=\underset{C_i}{\arg \max }\left(\mathrm{YOL}\left(C_n{,C}_i\right)\right),i=\mathrm{1,2},...,n-1.$

If C _nAnd C _YVertically lap is enough big, promptly

YOL(C _n，C _Y)＞ h/2，

C _nAnd C _YJust belong to same literal line, otherwise C _nIt is capable to belong to new literacy.After taking out all character blocks, just obtained the literal line cutting result of multirow subregion.

1.5 cutting BIG class character block

Because BIG class character block is the character block of adhesion in the ranks, so need carry out cutting at leap literal line place.If character block C _B∈ BIG has crossed over many literal lines, so in every literal line crossing over, at C _BNear one establish a capital and have character block and C _BVertically overlapping.

Therefore, the present invention adopts following way cutting BIG class character block:

For each character block C _B∈ BIG is at literal line L ^(nL)The interior character block C that seeks ^(nL)(nL=1,2 .., nLine), promptly

$C^{\left(\mathrm{nL}\right)}=\underset{C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}}{\arg \min }\left(\right|\frac{l_B+r_B}{2}-\frac{l_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}{+r}_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}}{2}\left|\right),C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\∈N{B}^{\left(\mathrm{nL}\right)}.$

Wherein

$N{B}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\∈L^{\left(\mathrm{nL}\right)},C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\∈\mathrm{MIDDLE},$

$\mathrm{HDIS}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)<\stackrel{\&OverBar;}{h}\&times;5,\mathrm{YOL}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)>\stackrel{\&OverBar;}{h}/3\},$

Represent in the nL bar literal line, at C _BNear and C _BVertical all overlapping MIDDLE class character blocks,

$\mathrm{HDIS}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)=\max (\max (l_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},l_B)-\min (r_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},r_B),0),$

Expression C _NB ^(nL)And C _BBetween horizontal range.C like this ^(nL)Be exactly at NB ^(nL)In and C _BThe character block that horizontal range is nearest.

If C ^(m)And C ^(m+1)Exist, C just is described _BCrossed over m bar and m+1 bar literal line, need

$y=\frac{b^{\left(m\right)}{+t}^{(m+1)}}{2}$

Just cross over the local cutting C of literal line _B

Cross over literal line place's cutting C at all _B, each part of mark is the MIDDLE class, is assigned to corresponding literal line then.

1.6 insert SMALL class character block

For each character block C _S∈ SMALL calculates C _SWith the distance of every literal line, and it is assigned to nearest literal line.Calculate C _SWith literal line L ^(nL)The method of distance as follows:

At first, at L ^(nL)Interior C _SThe left side seek horizontal range C _SNearest MIDDLE class character block C _l, promptly

$C_l=\underset{C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}}{\arg \min }(l_S-l_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}),C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;L{B}^{\left(\mathrm{nL}\right)},$

Wherein

$L{B}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;L^{\left(\mathrm{nL}\right)},C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;\mathrm{MIDDLE},l_S>l_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\}.$

Then, at L ^(nL)Interior C _SThe right side seek horizontal range C _SNearest MIDDLE class character block C _r, promptly

$C_r=\underset{C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}}{\arg \min }(r_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}-r_S),C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;R{B}^{\left(\mathrm{nL}\right)},$

Wherein

$R{B}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;L^{\left(\mathrm{nL}\right)},C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;\mathrm{MIDDLE},r_S>r_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\}.$

So, C _SWith literal line L ^(nL)Distance be

2 baselines and a line are estimated

Because Ah's Chinese character connects on baseline, so literal line baseline location is very crucial for character cutting.When the literal line image of pre-treatment is represented H with L _LBe the height of L, W _LIt is the width of L.

2.1 baseline Height Estimation

As can be seen from Figure 2, the height of the baseline of concatenation character is identical among the literal line L.Therefore, the length of the vertically black pixel distance of swimming that the frequency of occurrences is the highest among the L is exactly the height of baseline.

The present invention tlv triple (y ^(s), y ^(e), x ^(se)) expression vertically black the pixel distance of swimming, wherein y ^(s)Be the vertical reference position of the distance of swimming, y ^(e)Be the vertical final position of the distance of swimming, x ^(se)It is the horizontal level of the distance of swimming. $\mathrm{VRUN}=\left\{(y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)})\right|\mathrm{nR}=\mathrm{1,2},...,\mathrm{nRun}\}$ Be all the vertically black pixel distances of swimming among the L.

The number that then among the L highly is the vertically black pixel distance of swimming of runH is

$\mathrm{VH}\left(\mathrm{runH}\right)=\left|\right|\{y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)}|(y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)})\&Element;\mathrm{VRUN},y_{\mathrm{nR}}^{\left(e\right)}-y_{\mathrm{nR}}^{\left(s\right)}=\mathrm{runH}\left\}\right||,$

runH＝1，2，...，H _L。

The height that can get baseline thus is

$H_0=\underset{\mathrm{runH}}{\arg \max }\left(\mathrm{VH}\left(\mathrm{runH}\right)\right),\mathrm{runH}=\mathrm{1,2},...,H_L.$

2.2 baseline position is estimated

Ideally, the baseline of interior all the conjuncted character fields of literal line is all on same horizontal linear.But in the actual scan text image, literal line exists low-angle inclination or crooked deformation usually, causes the baseline position of different conjuncted character fields not necessarily identical.In order to address the above problem, the present invention is divided into a plurality of short parts to long literal line and estimates baseline position respectively, so just can think that the baseline position in each part is identical.

L is divided into literal line

Individual part, the length of each part are α * H ₀Make HP ^(nP)(y) (y=0,1 ..., H _L-1) be the horizontal projection result that L only keeps the nP parts of images, the baseline position of nP part is so,

$B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}=\underset{y}{\arg \max }\left(\underset{k=0}{\overset{H_0-1}{\mathrm{\&Sigma;}}}{\mathrm{HP}}^{\left(\mathrm{nP}\right)}(y+k)\right),y=\mathrm{0,1},...,H_L-H_0,$

$B_{\mathrm{Btm}}^{\left(\mathrm{nP}\right)}=B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}+H_0-1,$

B wherein _Top ^(nP)And B _Btm ^(nP)Coboundary and the lower boundary of representing baseline respectively.In real system, α can be between 10～15 value.So just can obtain the baseline position of L any place,

$\left\{\begin{array}{cc}{B}_{\mathrm{Top}}\left(x\right)=B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}& l^{\left(\mathrm{nP}\right)}x{r}^{\left(\mathrm{nP}\right)}\\ B_{\mathrm{Btm}}\left(x\right)=B_{\mathrm{Btm}}^{\left(\mathrm{nP}\right)}& l^{\left(\mathrm{nP}\right)}x{r}^{\left(\mathrm{nP}\right)}\end{array},\right.$

L wherein ^(nP)And r ^(nP)It is respectively the left and right border of nP part.

2.3 a line position is estimated

As shown in Figure 2, after the removal extention, the distance of the top of Ah's Chinese character and baseline coboundary can obviously be divided into two classes: a class is the distance of high character top and baseline coboundary; One class is the distance of low character top and baseline coboundary.Line position refers to the position on low character top.

Make U _L(x) (x=0,1 ..., W _L-1) is last profile after literal line L removes extention, then

E _L＝{x|U _L(x)＜U _L(x-1)，U _L(x)＜U _L(x+1)，1≤x＜W _L-1}，

Be U _L(x) set of minimal point.

The present invention utilizes the method for asking apart from average twice, distinguishes high character top and the low character top distance to the baseline coboundary.

Definition

$\mathrm{BU}\left(x\right)=\left\{\begin{array}{cc}{B}_{\mathrm{Top}}\left(x\right)-U_L\left(x\right)& x{\&Element;E}_L\\ 0& x\&NotElement;E_L\end{array}\right.,$

Then can get $h^{\&prime;}=\frac{1}{\left|\right|E_L\left|\right|}\underset{\left\{x\right|x\&Element;E_L\}}{\mathrm{\&Sigma;}}\mathrm{BU}\left(x\right).$ Make E '={ x|BU (x)＜h ', x ∈ E again _L, can get ${\mathrm{<figure-callout\; id="1"\; label="H"\; filenames="A20051008647800241.png,A20051008647800251.png"\; state="\{\{state\}\}">H</figure-callout>}}_1=\frac{1}{\left|\right|E^{\&prime;}\left|\right|}\underset{\left\{x\right|x\&Element;E^{\&prime;}\}}{\mathrm{\&Sigma;}}\mathrm{BU}\left(x\right).$

H ₁It is exactly the distance between a line and the baseline coboundary.

3 conjuncted character field cuttings

After obtaining literal line, need be cut into conjuncted character field to literal line.Each conjuncted character field is made up of a main part and some extentions.The method of vertical projection can not be handled the overlapping in the horizontal direction situation of adjacent conjuncted character field, therefore the present invention is according to the main part of baseline position and the conjuncted character field of character block classification mark, then according to the main part and the extention of the conjuncted character field of range information merger.

Two types character block is the main part of conjuncted character field below the face identification of the present invention: the MIDDLE class character block that is passed by baseline; Passed by baseline and SMALL class character block that other character blocks of getting along well in the horizontal direction are overlapping.Why part SMALL class character block is considered to conjuncted character field main part, is because these character blocks may be punctuate or isolated character.Remaining SMALL class character block then belongs to the extention of conjuncted character field.

Each is belonged to the SMALL class character block C of conjuncted character field extention _SSearch for the nearest character block C that belongs to conjuncted character field main part _M, C then _SAnd C _MBe the extention and the main part of same conjuncted character field.C _SAnd C _MDistance definition be

DIS(C _S，C _M)＝|l _S+r _S-l _M-r _M|。

4 character cuttings

The character cutting is the process that conjuncted character field is cut into single character.When the conjuncted character field image of pre-treatment is represented H with P _PBe the height of P, W _PIt is the width of P.The present invention is divided into point of contact on baseline to the character point of contact, at the point of contact above the baseline and point of contact three classes below baseline.According to shown in Figure 8, the present invention searches for position, all possible point of contact of every class successively, checks the legitimacy at point of contact then, at last extention is distributed to corresponding character main part.

4.1 the candidate point of contact on baseline

As shown in Figure 9, place, the point of contact on baseline satisfies one of following situation:

(A) distance of swimming number changes.The character left margin, vertically distance of swimming number is increased to more than 2 from 1; The character right margin, vertically distance of swimming number is from being reduced to 1 more than 2.

(B) distance of swimming invariable number, but bigger variation takes place in length.The character left margin, the position of last profile and/or bottom profiled is away from b extent; The character right margin, the position of last profile and/or bottom profiled returns b extent.

(C) distance of swimming invariable number, last outline position gradually changes, and causes the outline position accumulation that bigger variation takes place.The character left margin, last outline position is gradually away from b extent; The character right margin, last outline position returns b extent gradually.

Defined function D (x) describes the upper and lower profile of P and the distance between the baseline,

D(x)＝max(B _Top(x)-U _P(x)，0)+max(V _P(x)-B _Btm(x)，0)，x＝0，1，...，W _P-1，

U wherein _P(x) and V _P(x) be last profile and bottom profiled after P removes extention respectively.

Like this, the candidate point of contact x on baseline _OSatisfy following condition (1) and (2) or condition (1) and (3), "+" computing during condition (1) (2) (3) is various and "-" computing be the point of contact, left side and the point of contact, right side of corresponding character respectively:

(1) point of contact x _OOn baseline:

D(x _O)≤2；

(2) point of contact situation (A) or (B):

D(x _O±1)-D(x _O)＞1.5×H ₀；

(3) point of contact situation (C):

$\left\{\begin{array}{cc}D(x_O\&PlusMinus;i)-D\left(x_O\right)>0.75\&times;H_0& i=e\\ D(x_O\&PlusMinus;i)\&GreaterEqual;D(x_O\&PlusMinus;\mathrm{iml})& i=\mathrm{1,2},K,e\end{array}\right.,$

Wherein e satisfies x _O± e ∈ E _P, E _PBe R _P(x) set of minimal point;

4.2 the candidate point of contact of baseline top

As shown in figure 10, if position x _TBe the candidate point of contact of baseline top, x so _TMust satisfy following condition:

(1) point of contact itself is a utmost point low spot of going up profile:

U _P(x _T)＞U _P(x _T-1)，U _P(x _T)＞U _P(x _T+1)；

(2) go up profile above the position, point of contact is in a line:

U _P(x _T)＜B _Top(x _T)-H ₁；

(3) minimum point of left side, point of contact bottom profiled is below baseline:

$\underset{x{<x}_T}{\max }\left(V_P\left(x\right)\right)>B_{\mathrm{Btm}}\left(x_T\right);$

(4) minimum point of right side, point of contact bottom profiled is below baseline:

$\underset{x_T<x<w_P-1}{\max }\left(V_P\left(x\right)\right)>B_{\mathrm{Btm}}\left(x_T\right).$

4.3 the candidate point of contact of baseline below

As shown in figure 11, if position x _BBe the candidate point of contact of baseline below, x so _BMust satisfy following condition:

(1) point of contact left side 3 list profile peak TL above baseline:

$(\mathrm{TL}=\underset{x_B-3<x{<x}_B}{\min }\left(V_P\left(x\right)\right))<B_{\mathrm{Top}}\left(x_B\right);$

(2) point of contact itself and right side 3 list profile peak TR below baseline:

$(\mathrm{TR}=\underset{x_B\&le;k{<x}_B+3}{\min }\left(V_P\left(x\right)\right))<B_{\mathrm{Btm}}\left(x_B\right);$

(3) TL and TR difference in height are greater than twice baseline height:

TR-TL＞2×H ₀。

4.4 check the legitimacy at point of contact

Check the legitimacy at each point of contact from right to left successively, and delete illegal point of contact.If less than 2 times of baseline height, Zuo Ce point of contact is exactly an illegal point of contact to the distance between adjacent two point of contacts so less than the character height between the height of baseline and two point of contacts.Concrete grammar is as follows: establish x _rAnd x _lBe two adjacent point of contacts, wherein x _r＞x _l, if

$\left\{\begin{array}{c}{x}_r-x_l\&le;\frac{H_0}{2}\\ \underset{x_l<x<x_r}{\max }\left(V_P\left(x\right)\right)-\underset{x_l<x<x_r}{\min }\left(U_P\left(x\right)\right)\&le;H_0\&times;2\end{array}\right.,$

Then delete point of contact x _l

4.5 extention is distributed

When continuous two characters all had extention, extention (Figure 12) also might adhesion.If the main part of an extention and a plurality of characters is overlapping in the horizontal direction, then this extention is adhesion.If this extention is crossed over x _RAnd x _LTwo adjacent point of contacts are then at x _BAnd x _LBetween the minimum place's cutting of extention vertical projection value extention.

After the extention cutting of adhesion finishes, will distribute to nearest character main part to each extention.Extention C _AWith character main part C _MDistance be

DIS(C _A，C _M)＝|l _A+r _A-l _M-r _M|。

The A Wen text dividing algorithm of realizing according to the present invention has been applied to a block letter A Wen text recognition system.The black and white two-value Arabic of 300DPI resolution scan and the newspaper of Uighur and periodical are used for testing this system, recognition result such as table 1.Wherein have only the mistake about 30% to cause because of the cutting mistake, the accuracy rate that is to say character cutting is more than 99%.Test findings can illustrate the validity of the block letter A Wen text dividing method that the present invention proposes.

	Number of characters	Discrimination
			Test set 1	11966	95.38％
Test set 2	11790	97.79％
			Test set 3	9912	96.75％
Test set 4	10788	95.80％
			Amount to	44456	96.43％

Table 1 test findings

Many font sizes of multi-font block letter A Wen text dividing method that the present invention proposes has obtained excellent text dividing performance in experiment, be with a wide range of applications.In sum, the present invention has the following advantages:

First, the present invention propose on the basis of character block classification, elder generation's horizontal projection is the text filed subregion that is divided into, realize that at the inner merger character block of multirow subregion the method for literal line cutting not only can the nonangular text of processing horizontal then, it is big to handle the angle of inclination, has the text of character adhesion in the ranks.

The second, baseline, a line position method of estimation accuracy rate height that the present invention proposes.

The 3rd, the conjuncted character field cutting method that the present invention proposes can be handled the overlapping in the horizontal direction situation of adjacent conjuncted character field.

The 4th, the present invention can seek the point of contact on baseline, reaches the point of contact below baseline at the point of contact above the baseline, can cut the adhesion extention, and utilizes the cut-off of crossing of the frequent appearance of tactical rule deletion.Character point of contact recall ratio height, pseudo-point of contact number is few.

Description of drawings

Fig. 1 standard Arabic basic character set

Fig. 2 A Wen text characteristics

Fig. 3 block letter A Wen text recognition system structural drawing

The flow process of Fig. 4 INVENTION IN GENERAL

Fig. 5 literal line cutting flow process

Fig. 6 coordinate system and variable declaration

Fig. 7 multirow subregion detects

Fig. 8 character cutting flow process

The point of contact characteristics of Fig. 9 on baseline

Point of contact, Figure 10 baseline top characteristics

Point of contact, Figure 11 baseline below characteristics

The cutting of Figure 12 extention

Embodiment

A block letter A Wen text recognition system is made of image capture device and computing machine two parts on hardware.Image capture device generally is a scanner, is used for obtaining the digital picture of text to be identified.Computing machine is used for digital picture is handled, and finishes the final identification of text.

For one piece of A Wen text specimen page, at first it is swept computing machine by scanner, make it to become digital picture.Digital picture is taked pre-service measures such as binaryzation, removal noise, obtained bianry image.Capable cutting obtains line of text to input picture again, estimates baseline, a line position of line of text, and line of text is cut into conjuncted character field.On this basis each disjunctor character field is carried out the individual character cutting, obtain single character, discern single character then.The mistake in each stage mode by hand corrects.

Therefore, realize practical block letter A Wen text recognition system, aspect text dividing, need consider following several aspect: the literal line cutting; A baseline and a line position are estimated; Conjuncted character field cutting; Character cutting.Respectively these four aspects are described in detail below:

The first, the literal line cutting

1.1 character block classification

According to height and width information, all the character block { C on the text image I _NB| nB=1,2 ..., nBlock} is divided into three classes:

Wherein

$\stackrel{\&OverBar;}{h}=\frac{1}{\mathrm{nBlock}}\underset{\mathrm{nB}=1}{\overset{\mathrm{nBlock}}{\mathrm{\&Sigma;}}}{h}_{\mathrm{nB}}.$

1.2 subregion cutting

Only keep MIDDLE class character block horizontal projection, to be 0 position be divided into a plurality of subregions text filed to the horizontal projection value.Each subregion comprises one or more literal line.

1.3 the multirow subregion detects

Each subregion is obtained

CC＝{(C _I，C _J)|XOL(C _I，C _J)＞0，YOL(C _I，C _J)＝0，C _I∈MIDDLE，C _J∈MIDDLE}。

Wherein,

XOL(C _I，C _J)＝max(min(r _I，r _J)-max(l _I，l _J)，0)，

YOL(C _I，C _J)＝max(min(b _I，b _J)-max(t _I，t _J)，0)。

If ‖ CC is ‖＞and 10, just judge that this subregion is the multirow subregion.

1.4 multirow subregion cutting

According to the inner all MIDDLE class character blocks of right margin descending sort multirow subregion.Take out a character block C in order at every turn _nAll character block C that taken out ₁, C ₂..., C _N-1Relatively, find overlapping maximum character block C on the vertical direction _Y, promptly

$C_Y=\underset{C_i}{\arg \max }\left(\mathrm{YOL}(C_n,C_i)\right),i=\mathrm{1,2},...,n-1.$

If C _nAnd C _YVertically lap is enough big, promptly

YOL(C _n，C _Y)＞ h/2，

C _nAnd C _YJust belong to same literal line, otherwise C _nIt is capable to belong to new literacy.After taking out all character blocks, just obtained the literal line cutting result of multirow subregion.At last all literal lines of all subregions are arranged in accordance with the order from top to bottom.

1.5 cutting BIG class character block

For each character block C _B∈ BIG is at literal line L ^(nL)The interior character block C that seeks ^(nL)(nL=1,2 .., nLine), promptly

$C^{\left(\mathrm{nL}\right)}=\underset{C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}}{\arg \min }\left(\right|\frac{l_B+r_B}{2}-\frac{l_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}+r_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}}{2}\left|\right),C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\&Element;{\mathrm{NB}}^{\left(\mathrm{nL}\right)}.$

Wherein

${\mathrm{NB}}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\&Element;L^{\left(\mathrm{nL}\right)},C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\&Element;\mathrm{MIDDLE},$

$\mathrm{HDIS}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)<\stackrel{\&OverBar;}{h}\&times;5,\mathrm{YOL}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)>\stackrel{\&OverBar;}{h}/3\},$

$\mathrm{HDIS}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)=\max (\max (l_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},l_B)-\min (r_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},r_B),0).$

If C ^(m)And C ^(m+1)Exist, C just is described _BCrossed over m bar and m+1 bar literal line, need

$y=\frac{b^{\left(m\right)}+t^{(m+1)}}{2},$

Place's cutting C _BCross over the position of literal line C at all _BBe cut into a plurality of parts, each part of mark is the MIDDLE class, is assigned to corresponding literal line then.

1.6 insert SMALL class character block

For each character block C _S∈ SMALL calculates C _SWith the distance of every literal line, and it is assigned to nearest literal line.

Calculate C _SWith literal line L ^(nL)The method of distance as follows: seek the character block C that satisfies condition _lAnd C _r, promptly

$C_l=\underset{C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}}{\arg \min }(l_S-l_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}),C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;{\mathrm{LB}}^{\left(\mathrm{nL}\right)},$

$C_r=\underset{C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}}{\arg \min }(r_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}-r_S),C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;{\mathrm{RB}}^{\left(\mathrm{nL}\right)},$

Wherein

${\mathrm{LB}}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;L^{\left(\mathrm{nL}\right)},C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;\mathrm{MIDDLE},l_S>l_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\},$

${\mathrm{RB}}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;L^{\left(\mathrm{nL}\right)},C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;\mathrm{MIDDLE},r_S<r_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\}.$

C then _SWith literal line L ^(nL)Distance be

The second, a baseline and a line are estimated

When the literal line image of pre-treatment is represented H with L _LBe the height of L, W _LIt is the width of L.

2.1 baseline Height Estimation

Order $\mathrm{VRUN}=\left\{(y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)})\right|\mathrm{nR}=\mathrm{1,2},...,\mathrm{nRun}\}$ Be all the vertically black pixel distances of swimming among the L,

$\mathrm{VH}\left(\mathrm{runH}\right)=\left|\right|\{(y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)})|(y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)})\&Element;\mathrm{VRUN},y_{\mathrm{nR}}^{\left(e\right)}-y_{\mathrm{nR}}^{\left(s\right)}=\mathrm{runH}\}\left|\right|,$

runH＝1，2，...，H _L。

The height that then can get baseline is

$H_0=\underset{\mathrm{runH}}{\arg \max }\left(\mathrm{VH}\left(\mathrm{runH}\right)\right),\mathrm{runH}=\mathrm{1,2},...,H_L.$

2.2 baseline position is estimated

L is divided into literal line

Individual part, the length of each part are 15 * H ₀Make HP ^(nP)(y) (y=0,1 ..., H _L-1) be the horizontal projection result that L only keeps the nP parts of images, the baseline position of nP part is so,

$B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}=\underset{y}{\arg \max }\left(\underset{k=0}{\overset{H_0-1}{\mathrm{\&Sigma;}}}{\mathrm{HP}}^{\left(\mathrm{nP}\right)}(y+k)\right),y=\mathrm{0,1},...,H_L-H_0,$

$B_{\mathrm{Btm}}^{\left(\mathrm{nP}\right)}=B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}+H_0-1,$

B wherein _Top ^(nP)And B _Btm ^(nP)Coboundary and the lower boundary of representing baseline respectively.So just can obtain the baseline position of L any place,

$\left\{\begin{array}{cc}{B}_{\mathrm{Top}}\left(x\right)=B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}& l^{\left(\mathrm{nP}\right)}\&le;x<r^{\left(\mathrm{nP}\right)}\\ B_{\mathrm{Btm}}\left(x\right)=B_{\mathrm{Btm}}^{\left(\mathrm{nP}\right)}& l^{\left(\mathrm{nP}\right)}\&le;x<r^{\left(\mathrm{nP}\right)}\end{array}\right.,$

L wherein ^(nP)And r ^(nP)It is respectively the left and right border of nP part.

2.3 a line position is estimated

Make U _L(x) (x=0,1 ..., W _L-1) is last profile after literal line L removes extention, then

E _L＝{x|U _L(x)＜U _L(x-1)，U _L(x)＜U _L(x+1)，1≤x＜W _L-1}，

Be U _L(x) set of minimal point.

Definition

$\mathrm{BU}\left(x\right)=\left\{\begin{array}{cc}{B}_{\mathrm{Top}}\left(x\right)-U_L\left(x\right)& x\&Element;E_L\\ 0& x\&NotElement;E_L\end{array}\right.,$

Then can get $h^{\&prime;}=\frac{1}{\left|\right|E_L\left|\right|}\underset{\left\{x\right|x\&Element;E_L\}}{\mathrm{\&Sigma;}}\mathrm{BU}\left(x\right).$ Make E '={ x|BU (x)＜h ', x ∈ E again _L, can get ${\mathrm{<figure-callout\; id="1"\; label="H"\; filenames="A20051008647800241.png,A20051008647800251.png"\; state="\{\{state\}\}">H</figure-callout>}}_1=\frac{1}{\left|\right|E^{\&prime;}\left|\right|}\underset{\left\{x\right|x{\&Element;E}^{\&prime;}\}}{\mathrm{\&Sigma;}}\mathrm{BU}\left(x\right).$

H ₁It is exactly the distance between a line and the baseline coboundary.

The 3rd, conjuncted character field cutting

Each conjuncted character field is made up of a main part and some extentions.The MIDDLE class character block that is passed by baseline and passed by baseline and SMALL class that other character blocks of getting along well in the horizontal direction are overlapping is the main part of conjuncted character field.Remaining SMALL class character block then belongs to the extention of conjuncted character field.

Each is belonged to the SMALL class character block C of conjuncted character field extention _SSearch for the nearest character block C that belongs to conjuncted character field main part _M, C then _SAnd C _MBe the extention and the main part of same conjuncted character field.C _SAnd C _MDistance definition be

DIS(C _S，C _M)＝|l _S+r _S-l _M-r _M|。

The 4th, character cutting

When the conjuncted character field image of pre-treatment is represented H with P _PBe the height of P, W _PIt is the width of P.

4.1 the candidate point of contact on baseline

Definition

D(x)＝max(B _Top(x)-U _P(x)，0)+max(V _P(x)-B _Btm(x)，0)，x＝0，1，...，W _P-1，

U wherein _P(x) and V _P(x)) be last profile and bottom profiled after P removes extention respectively.

Candidate point of contact x on baseline _OSatisfy (1) and (2) or (1) and (3), below "+" computing and "-" computing in various distinguish the point of contact, left side and the point of contact, right side of corresponding character:

(1) point of contact x _OOn baseline:

D(x _O)≤2；

(2) point of contact situation (A) or (B):

D(x _O±1)-D(x _O)＞1.5×H ₀；

(3) point of contact situation (C):

$\left\{\begin{array}{cc}D(x_O\&PlusMinus;i)-D\left(x_O\right)>0.75\&times;H_0& i=e\\ D(x_O\&PlusMinus;i)\&GreaterEqual;D(x_O\&PlusMinus;\mathrm{im}1)& i=\mathrm{1,2},K,e\end{array}\right.,$

Wherein e satisfies x _O± e ∈ E _P, E _PBe U _P(x) set of minimal point.

4.2 the candidate point of contact of baseline top

The point of contact x of baseline top _TMust satisfy following condition:

(1) point of contact itself is a utmost point low spot of going up profile:

U _P(x _T)＞U _P(x _T-1)，U _P(x _T)＞U _P(x _T+1)；

(2) go up profile above the position, point of contact is in a line:

U _P(x _T)＜B _Top(x _T)-H ₁；

(3) minimum point of left side, point of contact bottom profiled is below baseline:

$\underset{x{<x}_T}{\max }\left(V_P\left(x\right)\right)>B_{\mathrm{Btm}}\left(x_T\right);$

(4) minimum point of right side, point of contact bottom profiled is below baseline:

$\underset{x_T<x{<w}_P-1}{\max }\left(V_P\left(x\right)\right)>B_{\mathrm{Btm}}\left(x_T\right).$

4.3 the candidate point of contact of baseline below

The point of contact x of baseline below _BMust satisfy following condition:

(1) point of contact left side 3 list profile peak TL above baseline:

$(\mathrm{TL}=\underset{x_B-3<x<x_B}{\min }\left(V_P\left(x\right)\right))<B_{\mathrm{Top}}\left(x_B\right);$

(2) point of contact itself and right side 3 list profile peak TR below baseline:

$(\mathrm{TR}=\underset{x_B\&le;k{<x}_B+3}{\min }\left(V_P\left(x\right)\right)){>B}_{\mathrm{Btm}}\left(x_B\right);$

(3) TL and TR difference in height are greater than twice baseline height:

TR-TL＞2×H ₀。

4.4 check the legitimacy at point of contact

Check the legitimacy at each point of contact from right to left successively, and delete illegal point of contact.Concrete grammar is as follows: establish x _rAnd x _lBe two adjacent point of contacts, wherein x _r＞x _l, if

$\left\{\begin{array}{c}{x}_r-x_l\&le;\frac{H_0}{2}\\ \underset{x_l<x{<x}_r}{\max }\left(V_P\left(x\right)\right)-\underset{x_l<x{<x}_r}{\min }\left(U_P\left(x\right)\right)\&le;H_0\&times;2\text{'}\end{array}\right.$

Then delete point of contact x _l

4.5 extention is distributed

If the main part of extention and a plurality of characters is overlapping in the horizontal direction, then extention is adhesion.If this extention is crossed over x _RAnd x _LTwo adjacent point of contacts are then at x _RAnd x _LBetween the minimum place's cutting of extention vertical projection value extention.

At last each extention is distributed to nearest character main part.Extention C _AWith character main part C _MThe distance calculation formula be

DIS(C _A，C _M)＝|l _A+r _A-l _M-r _M|。

Claims (1)

1. the method for blocks letter Arabic Chinese character collection text dividing, it is characterized in that: at first on the basis of character block classification, select part character block horizontal projection that the input text image segmentation is become subregion earlier, detect the literal line cutting that multirow subregion merger character block is realized the multirow subregion then; Estimate a baseline and a line position of literal line then; Subsequently literal line is cut into conjuncted character field; Last characteristics according to the character point of contact are sought the point of contact on baseline, reach the point of contact below baseline at the point of contact above the baseline; In the system that is made up of image capture device and computing machine, this method contains following steps successively:

The 1st step, the literal line cutting

Make I represent the text image of importing, H is the height of I, W is the width of I, and the left margin of character block C, coboundary, right margin, lower boundary, width and be expressed as l, t, r, b, w and h have highly respectively added various target C up and down and represented specific character block (C for example _NB, C ^(line)Deng), added left margin, coboundary, right margin, lower boundary, width and the height that identical up and down target l, t, r, b, w and h represent this character block respectively (h for example _NB, r ^(line)Deng);

The 1.1st step, the character block classification

According to height and width information, character block { C all on the input text image I _NB| nB=1,2 ..., nBlock} is divided into three classes:

Wherein

$\stackrel{\&OverBar;}{h}=\frac{1}{\mathrm{nBlock}}\underset{\mathrm{nB}=1}{\overset{\mathrm{nBlock}}{\mathrm{\&Sigma;}}}{h}_{\mathrm{nB}},$

It is the average height of all character blocks;

The 1.2nd step, the subregion cutting

To MIDDLE class character block horizontal projection, to be 0 position be divided into a plurality of subregions text filed to the horizontal projection value, and each subregion comprises one or more literal line;

In the 1.3rd step, the multirow subregion detects

Each subregion is obtained overlapping, the vertical nonoverlapping MIDDLE class character block group of all levels

CC＝{(C _I，C _J|XOL(C _I，C _J)＞0，YOL(C _I，C _J)＝0，C _I∈MIDDLE，C _J∈MIDDLE}，

Wherein,

XOL(C _I，C _J)＝max(min(r _I，r _J)-max(l _I，l _J)，0)，

XOL (C _I, C _J) expression character block C _IWith character block C _JOverlap length in the horizontal direction,

YOL(C _J，C _J)＝max(min(b _I，b _J)-max(t _I，t _J)，0)，

YOL (C _I, C _J) expression character block C _IAnd character block C _JThe overlap length of in the vertical direction;

If ‖ CC is ‖＞and 10, just judge that this subregion is the multirow subregion;

The 1.4th step, the cutting of multirow subregion

According to the inner all MIDDLE class character blocks of right margin descending sort multirow subregion, take out a character block C in order at every turn _nAll character block C that taken out ₁, C ₂..., C _N-1Relatively, find overlapping maximum character block C on the vertical direction _Y, promptly

$C_Y=\underset{C_i}{\arg \max }\left(\mathrm{YOL}(C_n,C_i)\right),i=\mathrm{1,2},...,n-1,$

If C _nAnd C _YVertically lap is enough big, promptly

YOL(C _n，C _Y)＞h/2，

C _nAnd C _YJust belong to same literal line, otherwise C _nIt is capable to belong to new literacy, take out all character blocks after, just obtained the literal line cutting result of multirow subregion;

The 1.5th step, cutting BIG class character block

For each character block C _B∈ BIG is at literal line L ^(nL)The interior character block C that seeks ^(nL)(nL=1,2 .., nLine), promptly

$C^{\left(\mathrm{nL}\right)}=\underset{C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}}{\arg \max }\left(\right|\frac{l_B+r_B}{2}-\frac{l_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}+r_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}}{2}\left|\right),C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\&Element;{\mathrm{NB}}^{\left(\mathrm{nL}\right)},$

Wherein

${\mathrm{NB}}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\&Element;L^{\left(\mathrm{nL}\right)},C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)}\&Element;\mathrm{MIDDLE},$

$\mathrm{HDIS}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)<\stackrel{\&OverBar;}{h}\&times;5,\mathrm{YOL}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)>\stackrel{\&OverBar;}{h}/3\},$

NB ^(nL)Represent in the nL bar literal line, at C _BNear and C _BVertical all overlapping MIDDLE class character blocks,

$\mathrm{HDIS}(C_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},C_B)=\max (\max (l_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},l_B)-\min (r_{\mathrm{NB}}^{\left(\mathrm{nL}\right)},r_B),0),$

HDIS (C _NB ^(nL), C _B) expression C _NB ^(nL)And C _BBetween horizontal range, C then ^(nL)Be at NB ^(nL)In and C _BThe character block that horizontal range is nearest;

If C ^(m)And C ^(m+1)Exist, just explanation is in m bar and m+1 bar literal line, at C _BNear all have character block and C _BVertically overlapping, that is to say C _BCrossed over m bar and m+1 bar literal line, need

$y=\frac{b^{\left(m\right)}+t^{(m+1)}}{2},$

Place's cutting C _B, cross over the position of literal line C at all _BBe cut into a plurality of parts, each part of mark is the MIDDLE class, is assigned to corresponding literal line then;

In the 1.6th step, insert SMALL class character block

For each character block C _S∈ SMALL calculates C _SWith the distance of every literal line, and it is assigned to nearest literal line:

Calculate C _SWith literal line L ^(nL)The method of distance as follows: seek the character block C that satisfies condition _lAnd C _r, promptly

$C_l=\underset{C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}}{\arg \min }(l_S-l_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}),C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;{\mathrm{LB}}^{\left(\mathrm{nL}\right)},$

$C_r=\underset{C_{\mathrm{EB}}^{\left(\mathrm{nL}\right)}}{\arg \min }(r_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}-r_S),C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;{\mathrm{RB}}^{\left(\mathrm{nL}\right)},$

Wherein

${\mathrm{LB}}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;L^{\left(\mathrm{nL}\right)},C_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\&Element;\mathrm{MIDDLE},l_S>L_{\mathrm{LB}}^{\left(\mathrm{nL}\right)}\},$

${\mathrm{RB}}^{\left(\mathrm{nL}\right)}=\left\{C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\right|C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;L^{\left(\mathrm{nL}\right)},C_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\&Element;\mathrm{MIDDLE},r_S<r_{\mathrm{RB}}^{\left(\mathrm{nL}\right)}\},$

C then _SWith literal line L ^(nL)Distance be

In the 2nd step, a baseline and a line are estimated

When the literal line image of pre-treatment is represented H with L _LBe the height of L, W _LIt is the width of L;

The 2.1st step, the baseline Height Estimation

Order $\mathrm{VRUN}=\left\{(y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)})\right|\mathrm{nR}=\mathrm{1,2},...,\mathrm{nRun}\}$ Be all the vertically black pixel distances of swimming among the L, then

$\mathrm{VH}\left(\mathrm{runH}\right)=\left|\right|\left\{(y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)})\right|(y_{\mathrm{nR}}^{\left(s\right)},y_{\mathrm{nR}}^{\left(e\right)},x_{\mathrm{nR}}^{\left(\mathrm{se}\right)})\&Element;\mathrm{VRUN},y_{\mathrm{nR}}^{\left(e\right)}-y_{\mathrm{nR}}^{\left(s\right)}=\mathrm{runH}\left\}\right||,$

runH＝1，2，...，H _L，

The height that can get baseline thus is

$H_0=\underset{\mathrm{runH}}{\arg \min }\left(\mathrm{VH}\left(\mathrm{runH}\right)\right),\mathrm{runH}=\mathrm{1,2},...,H_L;$

In the 2.2nd step, baseline position is estimated

L is divided into literal line

Individual part, the length of each part are α * H ₀, α is value between 10～15, makes HP ^(nP)(y) (y=0,1 ..., H _L-1) be the horizontal projection result that L only keeps the nP parts of images, the baseline position of nP part is so,

$B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}=\underset{y}{\arg \max }\left(\underset{k=0}{\overset{H_0-1}{\mathrm{\&Sigma;}}}{\mathrm{HP}}^{\left(\mathrm{nP}\right)}(y+k)\right),y=\mathrm{0,1},...,H_L-H_0,$

$B_{\mathrm{Btm}}^{\left(\mathrm{nP}\right)}=B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}+H_0-1,$

B wherein _Top ^(nP)And B _Btm ^(nP)Coboundary and the lower boundary of representing baseline respectively so just can obtain the baseline position of L any place,

$\left\{\begin{array}{cc}{B}_{\mathrm{Top}}\left(x\right)=B_{\mathrm{Top}}^{\left(\mathrm{nP}\right)}& l^{\left(\mathrm{nP}\right)}\&le;x<r^{\left(\mathrm{nP}\right)}\\ B_{\mathrm{Btm}}\left(x\right)=B_{\mathrm{Btm}}^{\left(\mathrm{nP}\right)}& l^{\left(\mathrm{nP}\right)}\&le;x<r^{\left(\mathrm{nP}\right)}\end{array}\right.,$

L wherein ^(nP)And r ^(nP)It is respectively the left and right border of nP part;

In the 2.3rd step, a line position is estimated

Make U _L(x) (x=0,1 ..., W _L-1) is last profile after literal line L removes extention, then

E _L={ x|U _L(x)＜U _L(x-1), U _L(x)＜U _L(x+1), 1≤x＜W _L-1} is U _L(x) set of minimal point;

Definition

$\mathrm{BU}\left(x\right)=\left\{\begin{array}{cc}{B}_{\mathrm{Top}}\left(x\right)-U_L\left(x\right)& x\&Element;E_L\\ 0& x\&NotElement;E_L\end{array}\right.,$

Then can get $h^{\&prime;}=\frac{1}{\left|\right|E_L\left|\right|}\underset{\left\{x\right|x\&Element;E_L\}}{\mathrm{\&Sigma;}}\mathrm{BU}\left(x\right),$ Make E '={ x|BU (x)＜h ', x ∈ E again _L, can get $H_1=\frac{1}{\left|\right|E^{\&prime;}\left|\right|}\underset{\left\{x\right|x\&Element;E^{\&prime;}\}}{\mathrm{\&Sigma;}}\mathrm{BU}\left(x\right),$ H ₁It is exactly the distance between a line and the baseline coboundary;

The 3rd step, conjuncted character field cutting

Each conjuncted character field is made up of a main part and some extentions, the MIDDLE class character block that is passed by baseline and passed by baseline and SMALL class that other character blocks of getting along well in the horizontal direction are overlapping is the main part of conjuncted character field, remaining SMALL class character block then belongs to the extention of conjuncted character field;

Each is belonged to the SMALL class character block C of conjuncted character field extention _SSearch for the nearest character block C that belongs to conjuncted character field main part _M, C then _SAnd C _MBe the extention and the main part of same conjuncted character field, C _SAnd C _MDistance definition be

DIS(C _S，C _M)＝|l _S+r _S-l _M-r _M|；

The 4th step, character cutting

When the image user of the conjuncted character field of pre-treatment represents H _PBe the height of P, W _PIt is the width of P;

The 4.1st step, the candidate point of contact on baseline

Place, point of contact on baseline satisfies one of following situation:

(A) distance of swimming number changes: the character left margin, and vertically distance of swimming number is increased to more than 2 from 1; The character right margin, vertically distance of swimming number is from being reduced to 1 more than 2;

(B) distance of swimming invariable number, but bigger variation takes place in length: the character left margin, the position of last profile and/or bottom profiled is away from b extent; The character right margin, the position of last profile and/or bottom profiled returns b extent;

(C) distance of swimming invariable number, last outline position gradually changes, and causes the outline position accumulation that bigger variation takes place: the character left margin, last outline position is gradually away from b extent; The character right margin, last outline position returns b extent gradually;

Defined function D (x) describes the upper and lower profile of P and the distance between the baseline,

D (x)=max (B _Top(x)-U _P(x), 0)+max (V _P(x)-B _Btm(x), 0), x=0,1 ..., W _P-1, U wherein _P(x) and V _P(x) be last profile and bottom profiled after P removes extention respectively;

Candidate point of contact x on baseline _OSatisfy following condition (1) and (2) or condition (1) and (3), "+" computing during condition (1) (2) (3) is various and "-" computing be the point of contact, left side and the point of contact, right side of corresponding character respectively:

(1) point of contact x _OOn baseline:

D(x _O)≤2；

(2) point of contact situation (A) or (B):

D(x _O±1)-D(x _O)＞1.5×H ₀；

(3) point of contact situation (C):

$\left\{\begin{array}{cc}D(x_O\&PlusMinus;i)-D\left(x_O\right)>0.75\&times;H& i=e\\ D(x_O\&PlusMinus;i)\&GreaterEqual;D(x_O\&PlusMinus;\mathrm{iml})& i=\mathrm{1,2},K,e\end{array}\right.,$

Wherein e satisfies x _O± e ∈ E _P, E _PBe U _P(x) set of minimal point;

The 4.2nd step, the candidate point of contact of baseline top

The point of contact x of baseline top _TMust satisfy following condition:

(1) point of contact itself is a utmost point low spot of going up profile:

U _P(x _T)＞U _P(x _T-1)，U _P(x _T)＞U _P(x _T+1)；

(2) go up profile above the position, point of contact is in a line:

U _P(x _T)＜B _Top(x _T)-H ₁；

(3) minimum point of left side, point of contact bottom profiled is below baseline:

$\underset{x<x_T}{\max }\left(V_P\left(x\right)\right)>B_{\mathrm{Btm}}\left(x_T\right);$

(4) minimum point of right side, point of contact bottom profiled is below baseline:

$\underset{x_T<x<w_P-1}{\max }\left(V_P\left(x\right)\right)>B_{\mathrm{Btm}}\left(x_T\right);$

The 4.3rd step, the candidate point of contact of baseline below

The point of contact X of baseline below _BMust satisfy following condition:

(1) point of contact left side 3 list profile peak TL above baseline:

$(\mathrm{TL}=\underset{x_B-3<x<x_B}{\min }\left(V_P\left(x\right)\right))<B_{\mathrm{Top}}\left(x_B\right);$

(2) point of contact itself and right side 3 list profile peak TR below baseline:

$(\mathrm{TR}=\underset{x_B\&le;k<x_B+3}{\min }\left(V_P\left(x\right)\right))>B_{\mathrm{Btm}}\left(x_B\right);$

(3) TL and TR difference in height are greater than twice baseline height:

TR-TL＞2×H ₀；

The 4.4th step, the legitimacy at inspection point of contact

Check the legitimacy at each point of contact from right to left successively, and delete illegal point of contact, basis for estimation is as follows: establish x _rAnd x _lBe two adjacent point of contacts, wherein x _r＞x _l, if

$\left\{\begin{array}{c}{x}_r-x_l\&le;\frac{H_0}{2}\\ \underset{x_l<x<x_r}{\max }\left(V_P\left(x\right)\right)-\underset{x_l<x<x_r}{\min }\left(U_P\left(x\right)\right)\&le;H_0\&times;2\end{array}\right.,$

Then delete point of contact x _l

In the 4.5th step, extention is distributed

If the main part of extention and a plurality of characters is overlapping in the horizontal direction, then extention is adhesion, if this extention is crossed over x _RAnd x _LTwo adjacent point of contacts are then at x _RAnd x _LBetween the minimum place's cutting of extention vertical projection value extention;

At last each extention is distributed to nearest character main part, extention C _AWith character main part C _MThe distance calculation formula be

DIS(C _A，C _M)＝|l _A+r _A-l _M-r _M|。

Images