CN100487722C - Method for determining connection sequence of cascade classifiers with different features and specific threshold - Google Patents

Abstract

The invention relates to a method for fixing the connecting sequence and character threshold value of cascade classifiers with different characters, wherein said cascade classifiers are used to extract selected connecting components from the candidate connecting components decomposed from image; said method comprises: first decomposing at least one image to obtain several connecting components as present array; then parallel feeding present sample into cascade classifier, to circulated train, to fix the connecting sequence and character threshold value of cascade classifiers. The invention also provides a method for obtaining selected image from one image, which uses the cascade classifiers via aforementioned method, to quickly remove non-selected connecting components, to spend more cost on selected connecting components, to improve the image processing speed and improve the obtaining accuracy.

Description

A kind of order of connection of cascade classifier of definite feature and the method for characteristic threshold value

Technical field

This relates to a kind of digital image processing field clearly, relate in particular to a kind of order of connection of the cascade classifier of determining one group of different characteristic and the method for characteristic threshold value, and utilize the cascade classifier group that forms by this method from image, to obtain the method for selected image.

Background technology

Text detection in the natural scene has a lot of application with cutting apart.Along with the increase of high-performance, low price, portable digital image documentation equipment, the application of scene text identification is expanded rapidly.By the video camera that use links to each other with mobile phone, PDA or other special-purpose digital equipments, we can catch the text of getting at one's side easily: for example road name, advertisement, traffic warning, restaurant menu or the like.Automatic identification, translation and pronunciation to these texts can be played very big help to overseas tourist, dysopia personage and video frequency searching program, meeting processing etc.

Particularly automatically extracting text the natural scene image from image, is a challenging problem all the time.Its difficult point comprises: the font of character, size, complex background, inhomogeneous illumination, shade and picture noise etc.And, also more and more higher to the requirement of image processing speed.

In recent years, the work of obtaining at the natural scene image Chinese version has had development faster.There are two classes from natural scene image, to obtain the method for text at present.

The first kind is based on the method for texture.People such as Shin use star-like template pixel to disclose the internal characteristics of text in " based on the digital video text detection of support vector machine " delivered in 2000.In " use localization measure obtain text filed " of delivering in September, 2000, people such as P.Clark have carefully proposed the measure of 5 kinds of localization, and incompatible to try to achieve the candidate text filed with these set of measurements.Frequency domain method also is used to the texture of detection type like text, for example: the Fourier transform of short scan line, discrete cosine transform, Gabor conversion, wavelet decomposition, Multiresolution Edge Detection.We find that as the line of text in menu or the document, these methods are respond well for less relatively character, because small text has strong texture response usually.Yet for big character, for example roadside or trade name can mislead these algorithms for the strong texture response of complex background, are not found thereby stay a lot of big characters.

Second class methods are based on connected component (Connected-Component, method CC).Color quantization, mathematical morphology operation and symmetric neighborhood filtering are normally used for original image is decomposed into candidate's connected component.These algorithms can be handled big character and small characters effectively.But how extracting the text connected component from candidate's connected component, people often use didactic method, for example: length breadth ratio, alignment and combined analysis, topological analysis, the multilayer connected component is analyzed.The shortcoming of these class methods is that all didactic rules are the orders of fixing, and its threshold value is the manual empirical value of input, and is unstable usually, can not guarantee to obtain the testing result of robust.In addition, can also (support vector machine for example, SupportVector Machine SVM) extract the text connected component from candidate's connected component with a kind of strong classifier, the shortcoming of these class methods is must calculate its whole features to each connected component, and calculated amount and consumed time are all too big.

The present invention is subjected to the inspiration of human face detection tech, and the connected component (for example, the text connected component) that utilizes the cascade classifier group to extract from candidate's connected component will to select can obtain very high verification and measurement ratio when improving image processing speed.

Summary of the invention

One of purpose of the present invention is to propose a kind of one group of different characteristic (F of determining ₁, F ₂..., F _M) cascade classifier (h ₁, h ₂..., h _M) the order of connection and the method for characteristic threshold value.This cascade classifier group is used for extracting the connected component that will select from the candidate's connected component that is got by picture breakdown, and the different characteristic here is relevant with the image that will select, and this method may further comprise the steps:

A. piece image obtains a plurality of connected components as current sample by decomposing at least, and with the cascade classifier of M different characteristic cascade classifier as current each feature, described current sample comprises positive example set P and counter-example set N, described positive example is the connected component that will select, and described counter-example is non-selected connected component;

B. current sample is walked abreast and send in the cascade classifier of current each feature, carry out the once training in i the circuit training, wherein i is the positive integer of 0＜i≤M, choose maximum false alarm rate characteristic of correspondence in current all features that participate in each training successively, determine the order of connection of the cascade classifier of each different characteristic thus, wherein said false alarm rate for be cascaded in each training sorter think by mistake the connected component that will select actual be the ratio of quantity with the current counter-example quantity of non-selected connected component;

C. after each selected characteristic, again current sample is sent in the cascade classifier of the feature correspondence of this time choosing and trained, in this training process, false alarm rate and verification and measurement ratio all constantly change, and obtain the threshold interval of described feature according to the minimum detection rate that this feature is allowed to, determine the characteristic threshold value interval of the cascade classifier of each different characteristic thus; Described verification and measurement ratio is the ratio of quantity with the positive example quantity of the correct detected selected connected component of a cascade classifier; And

D. after execution in step b and c, the feature that is selected among the deletion step b and the sorter of this feature are to upgrade the sorter of current feature and current each feature, and the set of the positive example in this time training remained unchanged and be cascaded that sorter is thought the connected component that will select by mistake during with the threshold interval that obtains feature among the step c and actual the set as new counter-example for non-selected connected component upgraded current sample, be used for next circuit training.

Be that connected component further may further comprise the steps with picture breakdown among the above-mentioned steps a:

A1. handle described image with non-linear Niblack thresholding method; With

A2. be connected component with the picture breakdown after the described processing.

Wherein, non-linear Niblack thresholding method has respectively increased a statistics Order Statistic Filters in the background filter of standard Niblack method and prospect wave filter.

Another object of the present invention is to provide a kind of method of from image, obtaining the image that will select, may further comprise the steps:

A. be connected component with picture breakdown;

B. this connected component is sent into the first order of the cascade classifier of the one group of different characteristic that cascades up according to preceding method, this feature is relevant with the image that will select, each cascade classifier abandons non-selected connected component, and exports the connected component that will select to the next stage sorter; And

C. the connected component that will select with afterbody sorter output in the cascade classifier group is combined to form the image that will select.

Another purpose of the present invention is to provide a kind of device that obtains the image that will select from image, and this device comprises:

Decomposer, being used for picture breakdown is connected component;

The cascade classifier of the one group of different characteristic that cascades up according to preceding method is imported the first order of this cascade classifier with this connected component, and each cascade classifier abandons non-selected connected component, and the connected component that will select to the output of next stage sorter; And

Image synthesizer is used for the connected component that will select of cascade classifier group afterbody sorter output is combined to form the image that will select.

Because the inventive method has used new non-linear Niblack method to handle original image, can efficiently gray level image be decomposed into a plurality of candidate's connected components, has improved the quality of connected component.In addition, the cascade classifier group that said method is trained to can easily be removed most of non-text connected components, and pays close attention to think it may is the connected component of text fast.Like this, both reduce the calculated amount of this method, improved image processing speed, and can obtain very high verification and measurement ratio again.

Description of drawings

Fig. 1 is a process flow diagram of determining the method for the order of connection of cascade classifier of one group of different characteristic and characteristic threshold value according to an embodiment of the invention;

Fig. 2 is the process flow diagram according to the method for obtaining text image from image of second embodiment of the invention; And

Fig. 3 is the installation drawing that obtains text image from image according to third embodiment of the invention.

Embodiment

The present invention is described in further detail below in conjunction with the drawings and specific embodiments.

Mention as described above, the inventive method is subjected to the inspiration of human face detection tech, utilizes the cascade classifier group to extract the text connected component from candidate's connected component, and candidate's connected component obtains by decomposing original image, and original image can be a natural scene image.With the text connected component formation text image that combines, like this, we just can obtain text image from natural scene image.

So, how could above-mentioned cascade classifier group extract the text connected component from candidate's connected component?

At first, we have proposed 12 kinds of different features, and these 12 kinds of features can be distinguished text or non-text connected component effectively.Train again with each sorter in the corresponding cascade set of classifiers of these 12 features, and to this cascade classifier group, organize the order of connection and the characteristic threshold value of the cascade classifier of different characteristic to determine this.The cascade classifier group that cascades up like this can abandon non-text connected component apace, output text connected component.

Next specifically describe 12 kinds of features of the inward nature's characteristic that discloses the text connected component earlier, comprising: geometric properties, contrast on border feature, shape canonical feature, stroke feature and Space Consistency feature.

1. geometric properties

Geometric properties comprises area ratio (Area Ratio), length ratio (Length Ratio) and length breadth ratio (Aspect Ratio).They can get rid of a large amount of very effectively obviously is the connected component of non-text, and the cost of calculating is very little.Therefore they can sharply reduce the execution time of whole algorithm.

Area is the area of connected component and the ratio of image area than AreaRatio, is used to get rid of too big or too little connected component, and its formula is:

$\mathrm{Feature}\_\mathrm{AreaRatio}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{\mathrm{Area}\left(\mathrm{Picture}\right)}---\left(1\right)$

Length is used to get rid of oversize or too short connected component than LengthRatio:

$\mathrm{Feature}\_\mathrm{LengthRatio}=\frac{\max \{w,h\}}{\max \{\mathrm{PicW},\mathrm{PicH}\}}---\left(2\right)$

Length breadth ratio AspectRatio is used to propose too thin connected component:

$\mathrm{Feature}\_\mathrm{AspectRatio}=\max \{\frac{\mathrm{width}\left(\mathrm{CC}\right)}{\mathrm{height}\left(\mathrm{CC}\right)},\frac{\mathrm{height}\left(\mathrm{CC}\right)}{\mathrm{width}\left(\mathrm{CC}\right)}\}---\left(3\right)$

In the above-mentioned formula (2), in (3), w represents described connected component bounding box width, and h represents the height of described connected component bounding box, the width of W presentation video, and the height of H presentation video.

2. contrast on border feature

The contrast on border feature comprises contrast on border (Edge Contrast), and contrast on border is the ratio on registration and the border of connected component of the edge image of the border of connected component and original image, and its formula is:

$\mathrm{EdgeContrast}=\frac{\mathrm{Border}\left(\mathrm{CC}\right)\∩\mathrm{Edge}\left(\mathrm{Picture}\right)}{\mathrm{Border}\left(\mathrm{CC}\right)}---\left(4\right)$

Wherein, Border (CC) is the boundary pixel of connected component, and Edge (Picture) is the edge-detected image of original image, is the union of Canny operator and Sobel operator, and its formula is:

Edge(Picture)＝Canny(Picture)∪Sobel(Picture) (5)

The contrast on border feature is a most important characteristic.Propose this feature and be based on very general visual angle, do not consider complex background and inhomogeneous illumination, the text connected component " is highly surrounded " by its skirt response usually.Therefore, we use equation (4) to measure the surrounded by edges degree of a connected component.This feature has been utilized the advantage based on the texture detection algorithm very fully, and it also has very strong response for big character.And this feature provides a kind of method that is independent of each connected component contrast on border of measurement of image.

3. shape normalization feature

The text connected component often has more regular shape than the noise connected component in the natural scene.Based on this viewpoint, we have proposed 4 features: empty number, profile roughness, degree of compacting and dutycycle.We find that the text connected component has less value on empty number and profile roughness, have bigger value on degree of compacting and dutycycle; But not the connected component of text is then on the contrary.These features are used to suppress to have irregularly shaped but have the connected component that strong texture responds.

$\mathrm{Featuer}\_\mathrm{ContourRoughness}=\frac{|\mathrm{CC}-\mathrm{open}(\mathrm{imfill}\left(\mathrm{CC}\right),2\×2)|}{\left|\mathrm{CC}\right|}---\left(6\right)$

Feature_CCHoles＝|imholes(CC)| (7)

$\mathrm{Feature}\_\mathrm{Compact}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{{\left[\mathrm{Perimeter}\left(\mathrm{CC}\right)\right]}^2}---\left(8\right)$

$\mathrm{Feature}\_\mathrm{OccupyRatio}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{\mathrm{Area}\left(\mathrm{BoundingBox}\left(\mathrm{CC}\right)\right)}---\left(9\right)$

In the above-mentioned formula, imfill (CC) is the operation of filling up the inner hole of connected component, 2 * 2nd, and the structural element of morphology opening operation (structure element), morphology opening operation (open) is that connected component is carried out level and smooth operation.

4. stroke statistical nature

Character is made up of stroke, so we propose 2 features of calculating relative complex, discloses the stroke statistical information of connected component.These two features are to check " the non-systematicness " of connected component aspect character stroke in fact.

First feature is average stroke width MeanStrokeWidth, and we are based on a kind of like this viewpoint: the stroke width of character is all smaller usually:

Feature_Stroke_Mean＝Mean(strokeWidth(skeleton(CC))) (10)

Second feature is normalized stroke width standard deviation, and we are based on such viewpoint: the stroke with a character often has similar width, and the connected component that has very large value on stroke width standard deviation feature more may be a noise:

$\mathrm{Featuer}\_\mathrm{Stroke}\_\mathrm{Std}=\frac{\mathrm{Deviation}\left(\mathrm{strokeWidth}\left(\mathrm{skeleton}\left(\mathrm{CC}\right)\right)\right)}{\mathrm{Mean}\left(\mathrm{strokeWidth}\left(\mathrm{skeleton}\left(\mathrm{CC}\right)\right)\right)}---\left(11\right)$

In the above-mentioned formula, skeleton is a morphology skeleton algorithm, connected component is taken out skeleton and obtains skeleton diagram, the stroke width of strokeWidth for obtaining for every bit on the described skeleton diagram, Mean obtains mean breadth for for averaging a little on the described skeleton diagram.

5. Space Consistency feature

Latter two Space Consistency feature has been explored Space Consistency information, comes the non-text connected component of filtering.Noise often has less space systematicness and polymerism, so we have proposed this two features.The Space Consistency feature comprises Space Consistency area ratio (Spatial Coherence Area Ratio) and Space Consistency boundary characteristic (Spatial CoherenceBoundary Touching), wherein,

$\mathrm{Feature}\_\mathrm{AreaRatio}\_S=\frac{\mathrm{Area}\left(\mathrm{imdilate}(\mathrm{CC},5\×5)\right)}{\mathrm{Area}\left(\mathrm{Picture}\right)}---\left(12\right)$

Feature_Boundary_S＝Bound(imdilate(CC，5×5)) (13)

In the above-mentioned formula, imdilate is the operation that connected component is expanded, and 5 * 5 are the structural element of expansive working (structureelement).

Under the situation that much connected component of obvious non-texts has been excluded, in each layer (Niblack has black and white two color layers), if after the expansion through a minor structure element, certain connected component is expanded very severely, and it probably is the random noise of space correlation so.The text connected component then can be not like this, because the structure essence of character string, intercharacter often has the spacing of a bit, not adhesion and expand to a very big connected component mutually after expanding.By usage space consistance wave filter, we can effectively reduce picture noise.

After having proposed above-mentioned 12 features that can distinguish text or non-text connected component effectively,, and this cascade classifier group trained each sorter in the corresponding cascade set of classifiers of 12 features.Our training method will solve two problems, and one, what arranges these features in proper order with; Two, how many threshold values on each feature should be.Its advantage is to make the cascade classifier group both to have guaranteed to obtain the precision of image with the strong mode cascade in weak back earlier, has improved image processing speed again.Fig. 1 is the process flow diagram according to the method for the order of connection of the cascade classifier of determining one group of different characteristic of first embodiment of the invention and characteristic threshold value.

Train, at first will determine training examples (step 110).For example, we can be from picture library picked at random 200 width of cloth pictures, this 200 width of cloth picture is resolved into a plurality of connected components as training examples (method that original image is decomposed into connected component will be described in more detail below).This training sample rate comprises positive example set P and counter-example set N.Positive example is the connected component that we manually are labeled as text, and counter-example is the connected component that we manually are labeled as non-text.

For each training examples (i.e. connected component), it has two Booleans: one is mark true value (GroundTruth), and just whether this sample is text, and true is a text, and false is non-text; Another is the sorter output valve, and just sorter thinks whether this sample is text, and output positive is a text, and negative is non-text.By this meaning, false alarm rate false-positive represents to be classified device and thinks it is the non-text sample of text and the ratio of all non-text samples by mistake; Verification and measurement ratio detection rate is actual to be exactly true-positive, and expression is classified device and correctly thinks the text sample of text and the ratio of all text samples; False-rejection is exactly false-negative, and expression is classified the ratio that device is just being confirmed as the non-text sample that is not text and all non-text samples.

P gathers as positive example, does not change in whole training process, because we expect that each positive example (text connected component) can that is to say necessary " understanding " these positive examples of each sorter by all sorters, promptly will learn them.And for counter-example set N, because each sorter all can " be tackled " a part of counter-example, for each sorter in the cascade, the counter-example that they are seen is different.First sorter is seen all counter-examples, can only see those counter-examples that are divided into text by first mistake for second ... from the angle of a back sorter, it only needs those sorters of concern front not have the problem that can correctly distinguish, and its counter-example to be processed only is the non-text connected component by all sorters of front in other words.So we need change counter-example set N in each circulation of training.

As above mention, next describe the method that original image is decomposed into connected component in detail.

As everyone knows, be that connected component is based on a step very crucial in the connected component method with picture breakdown.If the result that decomposition step obtains is very poor, the performance of whole algorithm will sharply descend so.Existing decomposition method is mainly pursued validity and robustness.

Present embodiment has used a kind of new for the decomposition method of picture breakdown as connected component, comprises two steps: at first with non-linear Niblack thresholding method processing original image; Picture breakdown after will handling again is a connected component.

The key of Niblack method is: it thinks those text pixels that people were concerned about, its intensity can and its neighborhood averaging intensity a certain distance be arranged, this gap is greater than the k of its neighborhood strength criterion difference doubly.Its former being used to is carried out binary conversion treatment to image.In the present embodiment, we handle image earlier with this method, and then the picture breakdown after will handling is candidate's connected component, can also obtain the low-complexity of high efficiency and realization like this on the basis of existing validity and robustness.

Wherein, non-linear Niblack thresholding method has respectively increased a statistics Order Statistic Filters in the background filter of standard Niblack method and prospect wave filter.The formula of non-linear Niblack thresholding method is:

$\mathrm{NLNiblack}(x,y)=\left\{\begin{array}{c}1,f(x,y)<T_{+}(x,y)\\ -1,f(x,y)<T_{-}(x,y)\\ \begin{array}{cc}0,& \mathrm{other}\end{array}\end{array}\right.---\left(14\right)$

$T_{\&PlusMinus;}(x,y)={\widehat{\mathrm{\&mu;}}}_{p1}(x,y,W_B)\&PlusMinus;k\&CenterDot;{\widehat{\mathrm{\&sigma;}}}_{p2}(x,y,W_F)$

${\widehat{\mathrm{\&mu;}}}_{p1}=\mathrm{Order}[\mathrm{Mean}(f(x,y),W_B),p1,W_B]$

${\widehat{\mathrm{\&sigma;}}}_{p2}=\mathrm{Order}[\mathrm{Deviation}(f(x,y),W_F),p2,W_F]$

Wherein: k is the empirical value according to standard Niblack method, is set as the numerical value between the 0.17-0.19, preferably, is set as 0.18 in the present embodiment.F (x, y) be input picture (x, y) the pixel intensity of position, Mean (, W) be that window width is the mean filter of W, Deviate (, be that window width is the standard deviation wave filter of W W), Order[, p, W] be to be number percent with p, W is the order statistics wave filter of width.

In the present embodiment, in background filter

In, filter width W _BBe made as 1/16 of original image width, number percent p1 is made as 50%.This is because big median filter can not got rid of their high fdrequency component when extracting background object.This background filter can be dealt with the inhomogeneous illumination situation in the natural scene.

At the prospect wave filter

In, width W _FBe W _B1/5, p2 is made as 80%.For the pocket with big variance, the wave filter of this high number percent can propagate into contiguous zone with its influence effectively, can suppress local noise effectively simultaneously.

Certainly, above-mentioned filter width and number percent can be adjusted according to actual needs.

In addition, what deserves to be mentioned is, above-mentioned picture breakdown step also can non-linear Niblack method be handled image, and with existing be the technology of connected component with picture breakdown, equally also can reach purpose of the present invention, but because more of poor quality, thereby make the general effect of this method also can descend to some extent with the connected component of prior art acquisition.

Next, set and initialization operation (step 120).

Set this cascade classifier group (h ₁, h ₂... h ₁₂) overall system target detection rate D _Target=0.95; And manually import this target detection rate.

Initializing variable: overall verification and measurement ratio D is set ₀=1.0, counter-example set N ₁=N, cycle index i=0, the scope of i is 0＜i≤M, i.e. the set of 0＜i≤12, and initialization feature, this characteristic set comprises 12 feature (F ₁, F ₂... F ₁₂).Sorter is corresponding one by one with feature.

Make cycle index i=i+1 (step 130).

Judge that whether i is greater than M (step 140).If i is not more than M, then carry out in i cycle calculations once.For example, i=1 so just carries out the cycle calculations first time.Below with the first time cycle calculations be that example is elaborated.

With this positive example set P and current counter-example set N ₁In parallel the sending in each cascade classifier of sample train (step 150).Each sorter all calculates the eigenwert of all samples.For example, if first sorter characteristic of correspondence is geometric properties " area ratio ", so just calculate the area ratio of all samples, i.e. the ratio of the area of the area of sample connected component bounding box and whole picture.

After obtaining the eigenwert of all samples, be horizontal ordinate with the eigenwert, the quantity of connected component is ordinate, forms positive example P and counter-example N ₁Eigenvalue distribution figure.

At each feature, establish an initial value for (∞ ,+∞) threshold interval, if the eigenwert of a sample outside this threshold interval, then this sample is judged to the connected component of non-text by the cascade classifier of this feature correspondence; If the eigenwert of a sample is outside this threshold interval, then this sample is judged to the connected component of text by the cascade classifier of this feature correspondence.

This threshold interval (∞ ,+∞) time, all samples all meet this threshold interval, therefore, the verification and measurement ratio d of each sorter is 1, false alarm rate f also is 1.At each feature, this threshold interval is constantly dwindled, make the eigenwert of more and more samples not meet this threshold interval, positive example and counter-example constantly are judged to non-text connected component, the verification and measurement ratio d of each cascade classifier _1jWith false alarm rate f _1jConstantly descend, as the verification and measurement ratio d of certain sorter of the 1st circuit training _1iDrop to the overall verification and measurement ratio D that is not less than after the previous cycle _I-1The time, stop to dwindle described threshold interval.Here D _I-1=D ₀=1.0.Since the discreteness that distributes during actual computation, d _1iCan not drop to and equal D ₀, only can be big a little.

When this threshold interval, calculate the verification and measurement ratio d of each cascade classifier _1j, false alarm rate f _1jAnd the probability FR that correctly abandons non-text connected component _1j, wherein, FR _1j=1-f _1j, be the ratio that a cascade classifier correctly abandons quantity with the current counter-example quantity of non-text connected component.

In current characteristic set, promptly in 12 features, choose maximum false alarm rate f _1jCharacteristic of correspondence feature _k(step 160).The feature feature that this is selected _kBe first feature, its corresponding sorter is first sorter of this cascade classifier group.

Choose maximum false alarm rate characteristic of correspondence, be because take turns calculating as can be seen by above-mentioned one, under equal condition, maximum false alarm rate characteristic of correspondence is thought the text sample with non-text sample most, this feature just is considered to the most invalid feature so, and its classification capacity is the poorest, therefore it will be placed on the foremost of cascade classifier group, the rest may be inferred, so that the set of classifiers that cascades up with this method has the strong cascade system in weak earlier back.

Next, calculate the feature feature that this is selected _kMass ratio with and the minimum detection rate (step 170) that is allowed to.

The feature feature that is selected _kMass ratio γ=FR _k/ ∑ FR _1j, wherein, FR _kBe the feature feature that is selected in the 1st circuit training _kCorresponding cascade classifier correctly abandons the probability of non-text connected component, is equivalent to the quality of this sorter, and this value is by obtaining in the step 160; ∑ FR _1jRepresent that the cascade classifier of all feature features correspondences in the 1st circuit training correctly abandons the probability sum of non-text connected component.Ratio between two is the mass ratio of the cascade classifier of this feature correspondence that is selected, and is used to weigh the ability power of this feature differentiation text connected component and non-text connected component.

Distribute formula d according to verification and measurement ratio _i=(D _Target/ D _I-1) ^γ, calculate this feature feature _kThe minimum detection rate d that is allowed to _Min, D _I-1Be the overall verification and measurement ratio after the previous cycle training, i is a cycle index.Owing to be first time circuit training, D here _I-1=D ₀=1.0, d _Min=(D _Target) ^γ

Specifically describing this verification and measurement ratio below distributes formula how to obtain.

Suppose that we will send into some connected component serials in the cascade classifier of one group of M different characteristic, classify to the one-level level, if any one sorter thinks that a connected component is non-text connected component, be about to its removal, if think the text connected component, promptly export to the next stage sorter and classify once more.Like this, we are easy to obtain following relation:

$F=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{f}_i$ $D=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{d}_i---\left(15\right)$

For in M the sorter each a verification and measurement ratio d is arranged all _i, d hereto _iA false alarm rate f is arranged _i, in order to simplify expression, we are d _iForm a vector { d ₁, d ₂..d _M, this moment, overall verification and measurement ratio was $D=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{d}_i,$ Overall false alarm rate is $F=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{f}_i.$ If we set another group verification and measurement ratio for this M sorter

Then Dui Ying false alarm rate is

Overall verification and measurement ratio $D\&prime;=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{d}_i\&prime;,$ Overall false alarm rate is $F\&prime;=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{f\&prime;}_i.$ Under the situation of D=D ', F=F ' may not be arranged.Our purpose is, at overall verification and measurement ratio D=D _TargetSituation in, select to have that group verification and measurement ratio vector of minimum false alarm rate F.Do you so how under the situation that D fixes, minimize F?

By to equation (15) citation form to number conversion, we find that overall verification and measurement ratio distributes to all sorters linearly:

$\log \left(F\right)=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\log \left(f_i\right)$ $\log \left(D\right)=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\log \left(d_i\right)---\left(16\right)$

Hypothetical universe verification and measurement ratio D distributes to all sorters linearly according to " quality " of sorter, and " quality " of i sorter is Q, and all sorter quality sums are $Q=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{Q}_i,$ The mass ratio γ of i sorter _iBe defined as:

${\mathrm{\&gamma;}}_i=\frac{Q_i}{\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{Q}_j}---\left(17\right)$

Make that D is overall verification and measurement ratio, we can distribute equation expression as follows, i the verification and measurement ratio d that sorter is assigned to _iFor:

$d_i={\left(D\right)}^{{\mathrm{\&gamma;}}_i}---\left(18\right)$

By equation 1) we have:

$D=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{d}_i=\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}{\left(D\right)}^{{\mathrm{\&gamma;}}_i}=D^{\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_i}=D^{\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\frac{Q_i}{Q}}=D---\left(19\right)$

This illustrates that my allocation algorithm at first numerically is correct.

Because D is ∈ [0,1], its exponential function is a monotonic decreasing function." quality " of a sorter is good more, and γ is big more, and the verification and measurement ratio d that is assigned to is more little.Because " quality " means that well this sorter can get rid of non-text most effectively,, allow it can have more space to remove to get rid of non-text connected component so we allow its verification and measurement ratio d smaller.Reduce verification and measurement ratio and represented the more strict condition of setting, so just can get rid of more non-text connected component." quality " of sorter can be weighed by the probability of the non-text of correct eliminating.

After the minimum detection rate that the feature featurek that obtains choosing is allowed to, with all positive example set P and current counter-example set N ₁In sample send into the cascade classifier h of the feature correspondence of choosing _kIn train (step 180).

The eigenwert of these all samples of classifier calculated.For example,, then calculate the length ratio of all samples, computing formula reference description above if this feature is a length ratio.

If an initial value be (∞ ,+∞) threshold interval, when the eigenwert of a sample outside this threshold interval, then this sample is cascaded sorter h _kBe judged to non-text connected component.

This threshold interval is constantly dwindled, make positive example and counter-example constantly are judged to non-text connected component, cascade classifier h _kVerification and measurement ratio d _kWith false alarm rate f _kConstantly descend, work as d _kDrop to and be not less than the minimum detection rate d that is allowed to that obtains in the step 180 _MinThe time, stop to dwindle described threshold interval; The threshold interval of this moment is the feature feature that this is chosen _kThreshold interval.

Up to the present, the work in selected characteristic and definite characteristic threshold value interval all finishes.

Next new variables more is to be used for circuit training next time (step 190).

The sorter of deleting the above-mentioned feature that is selected and this feature is to upgrade the sorter of current characteristic set and current each feature.Be cascaded sorter during with the threshold interval that obtains feature in the step 180 and think that the non-text connected component of text connected component is as new counter-example set N by mistake _I+1, positive example set P remains unchanged, thereby upgrades current sample.Upgrade current overall verification and measurement ratio D again _i=D _I-1* d _Min, be used for next circuit training.

Ensuing cycle calculations and above-mentioned primary identical is selected a feature at every turn and is obtained the threshold interval of this feature.The sorter sequence number of the feature correspondence of at every turn selecting is this time round-robin number of times i.Greater than M, then end loop is calculated up to i.

The cascade classifier group that cascades up of the order of connection of Que Dinging can be got rid of non-text connected component apace as stated above, and more time is spent on the connected component that calculating may be text.

The feature that proposes in the present embodiment is relevant with text image, can distinguish text or non-text connected component effectively, therefore, the cascade classifier group of this stack features correspondence can be obtained the text connected component from candidate's connected component, thereby, by the combine text connected component, obtain the text image that we need.But, those skilled in the art should know, if the feature that proposes is relevant with other content that will select, this content can be that we wish any content of obtaining from original image, corresponding with this stack features so cascade classifier group can be obtained the connected component that will select from candidate's connected component, thereby be combined to form the image that we will select, and be not limited to text image.Therefore, the cascade classifier group of being determined by the method in the present embodiment can be obtained the connected component that will select according to the feature relevant with the content that will select.

Fig. 2 is the process flow diagram according to the method for obtaining text image from image of second embodiment of the invention.

At first, original image is decomposed into a plurality of candidate's connected components (step 210).The original image here can be a natural scene image.Can handle this original image with non-linear Niblack thresholding method earlier in this step; And then the picture breakdown after will handling is a plurality of connected components.Here handling the method for this original image with non-linear Niblack thresholding method is identical with disposal route among first embodiment, repeats no more herein.With non-linear Niblack thresholding method can be fast and robust obtain candidate's connected component.

Secondly, a plurality of candidate's connected components are sent into the first order according to the cascade classifier of one group of different characteristic that method cascaded up of first embodiment, this feature is relevant with text image, each cascade classifier abandons non-text connected component, and to next stage sorter output text connected component (step 220).The order of connection of this cascade classifier group and characteristic threshold value are determined according to the method for first embodiment.

Particularly, after the first order of a plurality of candidate's connected component input cascade set of classifiers, first sorter calculates the eigenwert of all connected components that receive according to own characteristic of correspondence.The eigenwert of all connected components is compared with the threshold interval of this feature respectively; At last the connected component of eigenwert outside this threshold interval abandoned as non-text connected component; The connected component of eigenwert in this threshold interval exported to second level sorter as the text connected component.That is to say that the connected component by first sorter refusal will no longer be transfused to second sorter, need further not calculate it and judge, therefore, can save a large amount of computing times.

After second sorter receives the connected component of first sorter output, carry out identical calculating and classification work again, the rest may be inferred, and to the last a sorter abandons non-text connected component, output text connected component.

Alternatively, the text connected component of above-mentioned cascade classifier group output can also be imported a strong classifier (step 230) again.The sorter of this strong classifier for being trained by standard A daboost method, the feature of this strong classifier is identical with the feature of aforementioned cascade classifier group.This strong classifier carries out linear combination and judges whether this connected component is the text connected component all eigenwerts of each connected component of aforementioned cascade classifier group output, thereby abandons non-text connected component, output text connected component.Because all eigenwerts of each connected component were all calculated in the cascade classifier group, as long as therefore in this strong classifier, carry out linear combination, total just can obtain this connected component eigenwert.Like this, can spend less computing time, further improve precision.

Certainly, do not use strong classifier here, can reach purpose of the present invention yet, add strong classifier, can further improve the precision of final formation image.

At last, the text connected component with output in the step 230 is combined to form text image (step 240).Like this, we have just obtained the text image that we need from original image.

In the method for present embodiment, owing to used new non-linear Niblack method to handle original image, can efficiently gray level image be decomposed into a plurality of candidate's connected components, improved the quality of connected component.In addition, the cascade classifier group can easily be removed most of non-text connected components, and pays close attention to think it may is the connected component of text fast.Like this, reduce the calculated amount of this method, improved image processing speed, and can obtain very high verification and measurement ratio.

Those skilled in the art should know, though the feature of present embodiment cascade set of classifiers is relevant with text image, but this feature also can be relevant with other content that will select, method in the present embodiment also can be used for obtaining any image that will select from image so, and is not limited to text image.

Fig. 3 is the installation drawing that obtains text image from image according to third embodiment of the invention.Device 300 comprises decomposer 310, cascade classifier group 320, strong classifier 330 and image synthesizer 340.

Decomposer 310 is used for original image is decomposed into a plurality of connected components.This decomposer 310 also comprises treating apparatus 312 and picture breakdown device 316.The non-linear Niblack thresholding method for the treatment of apparatus 312 usefulness is handled original image earlier, and non-linear here Niblack thresholding method is identical with first embodiment.Picture breakdown after picture breakdown device 316 will be handled is a plurality of connected components.

Cascade classifier group 320 is one group of different characteristic (F that the method according to first embodiment cascades up ₁, F ₂..., F ₁₂) cascade classifier (h ₁, h ₂... h ₁₂), these features are relevant with text image.With the first order of this connected component input cascade set of classifiers, each cascade classifier abandons non-text connected component, and to next stage sorter output text connected component.

Also comprise calculation element, comparison means and output unit in each sorter.Calculation element is used for according to this sorter characteristic of correspondence, calculates the eigenwert of all connected components that receive.Comparison means compares the eigenwert of all connected components respectively with the threshold interval of this feature.Output unit abandons the connected component of eigenwert outside this threshold interval as non-text connected component; The connected component of eigenwert in this threshold interval exported to the next stage sorter as the text connected component.

Strong classifier 330, the sorter of this strong classifier for training by standard A daboost method, the feature of this strong classifier is identical with the feature of cascade classifier group 320, and promptly the feature of this strong classifier comprises all features of cascade classifier group 320.All eigenwerts of the connected component of 330 pairs of cascade classifier groups of strong classifier, 320 outputs are carried out linear combination, and judge whether this connected component is the text connected component, thereby abandon non-text connected component, output text connected component.

Image synthesizer 340 is used for text connected component with strong classifier 330 output and is combined to form and wants text image.

Those skilled in the art should know, though the feature of present embodiment cascade set of classifiers 320 is relevant with text image, but this feature also can be relevant with other content that will select, device in the present embodiment also can be used for obtaining any image that will select from image so, and is not limited to text image.

The present invention describes in detail in conjunction with above-mentioned exemplary embodiments, various selections, modification, variation, improvement and/or basic equivalent technologies, and content at present known or (possibility) the unknown is known those of ordinary skill in the art.Therefore, above-mentioned exemplary embodiments of the present invention is not lying in restriction the present invention with illustrating.Can make multiple change not breaking away within the spirit and scope of the present invention.Therefore, the present invention can comprise all selection, modification, variation, improvement and/or basic equivalent technologies known or development later on.

Claims (44)

1. determine one group of different characteristic (F for one kind ₁, F ₂..., F _M) cascade classifier (h ₁, h ₂..., h _M) the order of connection and the method for characteristic threshold value, the cascade classifier group that described method forms is used for obtaining the image that will select from image, described different characteristic is relevant with the image that will select, wherein, M is 〉=1 positive integer, it is characterized in that, may further comprise the steps:

A. piece image obtains a plurality of connected components as current sample by decomposing at least, and with the cascade classifier of M different characteristic cascade classifier as current each feature, described current sample comprises positive example set P and counter-example set N, described positive example is to be labeled as the connected component that will select, and described counter-example is to be labeled as non-selected connected component;

B. current sample is walked abreast and send in the cascade classifier of current each feature, carry out the once training in i the circuit training, wherein i is the positive integer of 0＜i≤M, choose maximum false alarm rate characteristic of correspondence in current all features that participate in each training successively, determine the order of connection of the cascade classifier of each different characteristic thus, wherein said false alarm rate for be cascaded in each training sorter think by mistake the connected component that will select actual be the ratio of quantity with the current counter-example quantity of non-selected connected component;

C. after each selected characteristic, again current sample is sent in the cascade classifier of the feature correspondence of this time choosing and trained, in this training process, false alarm rate and verification and measurement ratio all constantly change, and obtain the threshold interval of described feature according to the minimum detection rate that this feature is allowed to, determine the characteristic threshold value interval of the cascade classifier of each different characteristic thus; Described verification and measurement ratio is the ratio of quantity with the positive example quantity of the correct detected selected connected component of a cascade classifier; And

D. after execution in step b and c, the feature that is selected among the deletion step b and the sorter of this feature are to upgrade the sorter of current feature and current each feature, and the set of the positive example in this time training remained unchanged and be cascaded that sorter is thought the connected component that will select by mistake during with the threshold interval that obtains feature among the step c and actual the set as new counter-example for non-selected connected component upgraded current sample, be used for next circuit training.

2. the method for claim 1 is characterized in that, is that connected component further may further comprise the steps with picture breakdown among the step a:

A1. handle described image with non-linear Niblack thresholding method;

A2. be connected component with the picture breakdown after the described processing.

3. method as claimed in claim 2 is characterized in that, described non-linear Niblack thresholding method has respectively increased a statistics Order Statistic Filters in the background filter of standard Niblack method and prospect wave filter.

4. method as claimed in claim 3 is characterized in that, the formula of described non-linear Niblack thresholding method is:

$\mathrm{NLNiblack}(x,y)=\left\{\begin{array}{cc}1,& f(x,y)>T_{+}(x,y)\\ -1,& f(x,y)<T_{-}(x,y)\\ 0,& \mathrm{other}\end{array}\right.$

$T_{\&PlusMinus;}(x,y)={\widehat{\mathrm{\&mu;}}}_{p1}(x,y,W_B)\&PlusMinus;k\&CenterDot;{\widehat{\mathrm{\&sigma;}}}_{p2}(x,y,W_F)$

${\widehat{\mathrm{\&mu;}}}_{p1}=\mathrm{Order}[\mathrm{Mean}(f(x,y),W_B),p1,W_B]$

${\widehat{\mathrm{\&sigma;}}}_{p2}=\mathrm{Order}[\mathrm{Deviation}(f(x,y),W_F),p2,W_F]$

Wherein: k is set as 0.17-0.19 according to standard Niblack method;

(x y) is (x, y) the pixel intensity of position of input picture to f;

Mean (, be that window width is the mean filter of W W);

Deviation (, be that window width is the standard deviation wave filter of W W);

Order[, p, W] be to be number percent with p, W is the order statistics wave filter of width.

5. method as claimed in claim 4 is characterized in that, in background filter In, filter width W _BBe made as 1/16 of picture traverse, number percent p1 is made as 50%; At the prospect wave filter In, width W _FBe W _B1/5, p2 is made as 80%.

6. method as claimed in claim 5 is characterized in that, sets earlier and initialization operation before step b carries out circuit training, further may further comprise the steps:

Set described cascade classifier group (h ₁, h ₂... h _j) overall system target detection rate D _Target, 0＜j＜=M wherein, M〉and 1;

Initializing variable: overall verification and measurement ratio D ₀=1.0, counter-example set N ₁=N, cycle index i=0, the scope of i is 0＜i≤M, and the initialization feature set, described characteristic set comprises j feature (F ₁, F ₂... F _j), 0＜j＜=M, M〉1.

7. method as claimed in claim 6 is characterized in that, the circuit training among the step b further may further comprise the steps:

B1. with parallel the sending in each cascade classifier of sample among described positive example set P and the current counter-example set Ni, calculate the eigenwert of all samples,

B2: at each feature, establish an initial value for (∞ ,+∞) first threshold interval, when the eigenwert of a sample outside described first threshold interval, then described sample is judged to non-selected connected component by the cascade classifier of described feature correspondence;

B3: at each feature, described first threshold interval is constantly dwindled, make positive example and counter-example constantly are judged to non-selected connected component, the verification and measurement ratio d of each cascade classifier _IjWith false alarm rate f _IjConstantly descend, as the verification and measurement ratio d of certain sorter of the i time circuit training _iDrop to the overall verification and measurement ratio D that is not less than after the previous cycle _I-1The time, stop to dwindle described first threshold interval; And

B4: the verification and measurement ratio d of each cascade classifier when obtaining current first threshold interval _Ij, false alarm rate f _IjAnd the probability FR that correctly abandons non-selected connected component _Ij, wherein, FR _Ij=1-f _Ij, be the ratio of quantity with the current counter-example quantity of the non-selected connected component that correctly abandons of a cascade classifier; And

B5. in current characteristic set, choose false alarm rate f _IjMaximum feature feature _k, the described feature feature that is selected _kThe sequence number of corresponding cascade classifier is current cycle time i.

8. method as claimed in claim 7 is characterized in that, and is further comprising the steps of after the step b5:

B6. according to the result of step b5, calculate the described feature feature that is selected _kMass ratio γ=FR _k/ ∑ FR _Ij, wherein, FR _kBe the feature feature that is selected in the i time circuit training _kCorresponding cascade classifier correctly abandons the probability of non-selected connected component, is equivalent to the quality of described cascade classifier, ∑ FR _IjIt is the probability sum that all cascade classifiers correctly abandon non-selected connected component in the i time circuit training; And

B7. distribute formula d according to verification and measurement ratio _i=(D _Target/ D _I-1) ^γ, calculate described feature feature _kThe minimum detection rate d that is allowed to _Min, D _I-1Be the overall verification and measurement ratio after the previous cycle training, i is a cycle index; And

B8. upgrade current overall verification and measurement ratio D _i=D _I-1* d _Min

9. method as claimed in claim 8 is characterized in that, the verification and measurement ratio among the described step b7 distributes formula to obtain with the following method:

Hypothetical universe verification and measurement ratio D distributes to all cascade classifiers linearly according to " quality " of cascade classifier, and the quality of i cascade classifier is Q _i, all cascade classifier quality sums are $Q=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{Q}_i,$ The mass ratio γ of i cascade classifier _iBe defined as:

${\mathrm{\&gamma;}}_i=\frac{Q_i}{\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{Q}_j}$

It is as follows that then verification and measurement ratio distributes equation expression, i.e. i the verification and measurement ratio d that sorter is assigned to _iFor:

$d_i={\left(D\right)}^{{\mathrm{\&gamma;}}_i}$

10. method as claimed in claim 9 is characterized in that step c further may further comprise the steps:

C1. the sample among described positive example set P and the current counter-example set Ni is sent into the cascade classifier h of the described feature correspondence of choosing _kIn, calculate the eigenwert of all samples;

C2. establish an initial value for (∞ ,+∞) second threshold interval, when the eigenwert of a sample outside described second threshold interval, then described sample is by described cascade classifier h _kBe judged to non-selected connected component; And

C3. described second threshold interval is constantly dwindled, make positive example and counter-example constantly are judged to non-selected connected component, described cascade classifier h _kVerification and measurement ratio d _kWith false alarm rate f _kConstantly descend, work as d _kDrop to and be not less than the minimum detection rate d that is allowed to that obtains among the step b7 _MinThe time, stop to dwindle described second threshold interval; Second threshold interval of this moment is the described feature feature that chooses _kThreshold interval.

11. as each described method of claim 1-10, it is characterized in that, the described image that will select is a text image, the described connected component that will select is the text connected component, described non-selected connected component is the connected component of non-text, and the feature relevant with text image comprises: geometric properties, contrast on border feature, shape canonical feature, stroke feature and Space Consistency feature.

12. method as claimed in claim 11 is characterized in that, described geometric properties comprises area ratio, length ratio and length breadth ratio, wherein,

Area ratio is the area of connected component and the ratio of image area, and its formula is:

$\mathrm{Feature}\_\mathrm{AreaRatio}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{\mathrm{Area}\left(\mathrm{Picture}\right)}$

The formula of length ratio is:

$\mathrm{Feature}\_\mathrm{MLengthRatio}=\frac{\max \{w,h\}}{\max \{\mathrm{PicW},\mathrm{PicH}\}}$

The formula of length breadth ratio is:

Feature_AspectRatio＝max{w/h，h/w}

In the above-mentioned formula, CC represents connected component, max{a, b} represents to get the higher value among a and the b, and w represents described connected component bounding box width value, and h represents the height value of described connected component bounding box, the width of PicW presentation video, and the height of PicH presentation video.

13. method as claimed in claim 11 is characterized in that, described contrast on border feature comprises contrast on border, and described contrast on border is the ratio on registration and the border of connected component of the edge image of the border of connected component and original image, and its formula is:

$\mathrm{EdgeContrast}=\frac{\mathrm{Border}\left(\mathrm{CC}\right)\&cap;\mathrm{Edge}\left(\mathrm{Picture}\right)}{\mathrm{Border}\left(\mathrm{CC}\right)}$

Wherein, Border (CC) is the border of connected component, and Edge (Picture) is the edge-detected image of original image, is the union of Canny operator and Sobel operator, and its formula is:

Edge(Picture)＝Canny(Picture)∪Sobel(Picture)

14. method as claimed in claim 11 is characterized in that, described shape canonical feature comprises connected component border roughness, empty number, and compactness and dutycycle, its formula is,

Connected component border roughness:

$\mathrm{Feature}\_\mathrm{ContourRoughness}=\frac{|\mathrm{CC}-\mathrm{open}(\mathrm{imfill}\left(\mathrm{CC}\right),2\&times;2)|}{\left|\mathrm{CC}\right|}$

The cavity number:

Feature_CCHoles＝|imholes(CC)|

Compactness:

$\mathrm{Feature}\_\mathrm{Compact}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{{\left[\mathrm{Perimeter}\left(\mathrm{CC}\right)\right]}^2}$

Dutycycle:

$\mathrm{Feature}\_\mathrm{OccupyRatio}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{\mathrm{Area}\left(\mathrm{BoundingBox}\left(\mathrm{CC}\right)\right)}$

In the above-mentioned formula, imfill (CC) is the operation of filling up the inner hole of connected component, 2 * 2nd, the structural element of morphology opening operation, open () is the morphology opening operation in the following formula, be that connected component is carried out level and smooth operation, imholes (CC) expression connected component cavity number, Area (CC) expression connected component area, the girth of Perimeter (CC) expression connected component, the area of Area (BoundingBox (CC)) expression connected component bounding box.

15. method as claimed in claim 11 is characterized in that, described stroke feature comprises stroke mean breadth and stroke width standard deviation, wherein,

The stroke mean breadth:

Feature_Stroke_Mean＝Mean(strokeWidth(skeleton(CC)))

The stroke width standard deviation:

$\mathrm{Feature}\_\mathrm{Stroke}\_\mathrm{std}=\frac{\mathrm{Deviation}\left(\mathrm{strokeWidth}\left(\mathrm{skeleton}\left(\mathrm{CC}\right)\right)\right)}{\mathrm{Mean}\left(\mathrm{strokeWidth}\left(\mathrm{skeleton}\left(\mathrm{CC}\right)\right)\right)}$

In the above-mentioned formula, skeleton is a morphology skeleton algorithm, connected component is taken out skeleton and obtained skeleton diagram, the stroke width of strokeWidth for obtaining for every bit on the described skeleton diagram, Mean is for averaging a little on the described skeleton diagram, obtain mean breadth, standard deviation is asked in Deviation () expression.

16. method as claimed in claim 11 is characterized in that, described Space Consistency feature comprises Space Consistency area ratio and Space Consistency boundary characteristic, wherein,

The Space Consistency area ratio:

$\mathrm{Feature}\_\mathrm{AreaRatio}\_S=\frac{\mathrm{Area}\left(\mathrm{imdilate}(\mathrm{CC},5\&times;5)\right)}{\mathrm{Area}\left(\mathrm{Picture}\right)}$

The Space Consistency boundary characteristic:

Feature_Boundary_S＝Bound(imdilate(CC，5×5))

In the above-mentioned formula, imdilate is the operation that connected component is expanded, 5 * 5 for the structural element of expansive working, Area (imdilate (CC, 5 * 5)) area of the connected component after the expansive working is carried out in expression, the border of the connected component after Bound (imdilate (CC, 5 * 5)) expression the carrying out expansive working.

17. a method of obtaining the image that will select from image is characterized in that, may further comprise the steps:

A. be connected component with picture breakdown;

B. described connected component is sent into the first order according to the cascade classifier of one group of different characteristic that method cascaded up of claim 1, described feature is relevant with the image that will select, each cascade classifier abandons non-selected connected component, and exports the connected component that will select to the next stage sorter; And

C. the connected component that will select with afterbody sorter output in the cascade classifier group is combined to form the image that will select.

18. method as claimed in claim 17 is characterized in that, steps A further may further comprise the steps:

A1. handle described image with non-linear Niblack thresholding method;

A2. be connected component with the picture breakdown after the described processing.

19. method as claimed in claim 18 is characterized in that, described non-linear Niblack thresholding method has respectively increased a statistics Order Statistic Filters in the background filter of standard Niblack method and prospect wave filter.

20. method as claimed in claim 19 is characterized in that, the formula of described non-linear Niblack thresholding method is:

$\mathrm{NLNiblack}(x,y)=\left\{\begin{array}{cc}1,& f(x,y)>T_{+}(x,y)\\ -1,& f(x,y)<T_{-}(x,y)\\ 0,& \mathrm{other}\end{array}\right.$

$T_{\&PlusMinus;}(x,y)={\widehat{\mathrm{\&mu;}}}_{p1}(x,y,W_B)\&PlusMinus;k\&CenterDot;{\widehat{\mathrm{\&sigma;}}}_{p2}(x,y,W_F)$

${\widehat{\mathrm{\&mu;}}}_{p1}=\mathrm{Order}[\mathrm{Mean}(f(x,y),W_B),p1,W_B]$

${\widehat{\mathrm{\&sigma;}}}_{p2}=\mathrm{Order}[\mathrm{Deviation}(f(x,y),W_F),p2,W_F]$

Wherein: k is set as 0.17-0.19 according to standard Niblack method;

(x y) is (x, y) the pixel intensity of position of input picture to f;

Mean (, be that window width is the mean filter of W W);

Deviation (, be that window width is the standard deviation wave filter of W W);

Order[, p, W] be to be number percent with p, W is the order statistics wave filter of width.

21. method as claimed in claim 20 is characterized in that, in background filter

In, filter width W _BBe made as 1/16 of picture traverse, number percent p1 is made as 50%; At the prospect wave filter

In, width W _FBe W _B1/5, p2 is made as 80%.

22. method as claimed in claim 21 is characterized in that, each grade sorter described in the step B in the cascade classifier of one group of different characteristic abandons non-selected connected component and exports the method for selecting connected component to the next stage sorter and may further comprise the steps:

B1. according to each grade sorter characteristic of correspondence, calculate the eigenwert of all connected components that receive;

B2. the eigenwert of all connected components is compared with the threshold interval of described feature respectively; And

B3. the connected component of eigenwert outside described threshold interval abandoned as non-selected connected component; The connected component of eigenwert in described threshold interval exported to the next stage sorter as the connected component that will select.

23. method as claimed in claim 22, it is characterized in that, the connected component of described cascade classifier group output is imported a strong classifier again, the sorter of described strong classifier for being trained by standard A daboost method, the feature of described strong classifier is identical with the feature of described cascade classifier group.

24. method as claimed in claim 23, it is characterized in that, described strong classifier carries out linear combination and judges whether this connected component is the connected component that will select all eigenwerts of the connected component of described cascade classifier group output, thereby abandon non-selected connected component, the connected component that output will be selected.

25. as each described method among the claim 17-24, it is characterized in that, the described image that will select is a text image, the described connected component that will select is the text connected component, described non-selected connected component is the connected component of non-text, and the feature relevant with text image comprises: geometric properties, contrast on border feature, shape canonical feature, stroke feature and Space Consistency feature.

26. method as claimed in claim 25 is characterized in that, described geometric properties comprises area ratio, length ratio and length breadth ratio, wherein,

Area ratio is the area of connected component and the ratio of image area, and its formula is:

$\mathrm{Feature}\_\mathrm{AreaRatio}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{\mathrm{Area}\left(\mathrm{Picture}\right)}$

The formula of length ratio is:

$\mathrm{Feature}\_\mathrm{MLengthRatio}=\frac{\max \{w,h\}}{\max \{\mathrm{PicW},\mathrm{PicH}\}}$

The formula of length breadth ratio is:

Feature_AspectRatio＝max{w/h，h/w}

In the above-mentioned formula, CC represents connected component, max{a, b} represents to get the higher value among a and the b, and w represents described connected component bounding box width value, and h represents the height value of described connected component bounding box, the width of PicW presentation video, and the height of PicH presentation video.

27. method as claimed in claim 25 is characterized in that, described contrast on border feature comprises contrast on border, and described contrast on border is the ratio on registration and the border of connected component of the edge image of the border of connected component and original image, and its formula is:

$\mathrm{EdgeContrast}=\frac{\mathrm{Border}\left(\mathrm{CC}\right)\&cap;\mathrm{Edge}\left(\mathrm{Picture}\right)}{\mathrm{Border}\left(\mathrm{CC}\right)}$

Wherein, Border (CC) is the border of connected component, and Edge (Picture) is the edge-detected image of original image, is the union of Canny operator and Sobel operator, and its formula is:

Edge(Picture)＝Canny(Picture)∪Sobel(Picture)

28. method as claimed in claim 25 is characterized in that, described shape canonical feature comprises connected component border roughness, empty number, and compactness and dutycycle, its formula is,

Connected component border roughness:

$\mathrm{Feature}\_\mathrm{ContourRoughness}=\frac{|\mathrm{CC}-\mathrm{open}(\mathrm{imfill}\left(\mathrm{CC}\right),2\&times;2)|}{\left|\mathrm{CC}\right|}$

The cavity number:

Feature_CCHoles=|imholes (CC) compactness:

$\mathrm{Feature}\_\mathrm{Compact}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{{\left[\mathrm{Perimeter}\left(\mathrm{CC}\right)\right]}^2}$

Dutycycle:

$\mathrm{Feature}\_\mathrm{OccupyRatio}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{\mathrm{Area}\left(\mathrm{BoundingBox}\left(\mathrm{CC}\right)\right)}$

In the above-mentioned formula, imfill (CC) is the operation of filling up the inner hole of connected component, 2 * 2nd, the structural element of morphology opening operation, open () is the morphology opening operation in the following formula, be that connected component is carried out level and smooth operation, imholes (CC) expression connected component cavity number, Area (CC) expression connected component area, the girth of Perimeter (CC) expression connected component, the area of Area (BoundingBox (CC)) expression connected component bounding box.

29. method as claimed in claim 25 is characterized in that, described stroke feature comprises stroke mean breadth and stroke width standard deviation, wherein,

The stroke mean breadth:

Feature_Stroke_Mean＝Mean(strokeWidth(skeleton(CC)))

The stroke width standard deviation:

$\mathrm{Feature}\_\mathrm{Stroke}\_\mathrm{std}=\frac{\mathrm{Deviation}\left(\mathrm{strokeWidth}\left(\mathrm{skeleton}\left(\mathrm{CC}\right)\right)\right)}{\mathrm{Mean}\left(\mathrm{strokeWidth}\left(\mathrm{skeleton}\left(\mathrm{CC}\right)\right)\right)}$

In the above-mentioned formula, skeleton is a morphology skeleton algorithm, connected component is taken out skeleton and obtained skeleton diagram, the stroke width of strokeWidth for obtaining for every bit on the described skeleton diagram, Mean is for averaging a little on the described skeleton diagram, obtain mean breadth, standard deviation is asked in Deviation () expression.

30. method as claimed in claim 25 is characterized in that, described Space Consistency feature comprises Space Consistency area ratio and Space Consistency boundary characteristic, wherein,

$\mathrm{Feature}\_\mathrm{AreaRatio}\_S=\frac{\mathrm{Area}\left(\mathrm{imdilate}(\mathrm{CC},5\&times;5)\right)}{\mathrm{Area}\left(\mathrm{Picture}\right)}$

Feature_Boundary_S＝Bound(imdilate(CC，5×5))

In the above-mentioned formula, imdilate is the operation that connected component is expanded, 5 * 5 for the structural element of expansive working, Area (imdilate (CC, 5 * 5)) area of the connected component after the expansive working is carried out in expression, the border of the connected component after Bound (imdilate (CC, 5 * 5)) expression the carrying out expansive working.

31. a device that obtains the image that will select from image is characterized in that, described device comprises:

Decomposer, being used for picture breakdown is connected component;

The cascade classifier of the one group of different characteristic that cascades up according to power 1 method, described connected component is imported the first order of the cascade classifier of described one group of different characteristic, each cascade classifier abandons non-selected connected component, and exports the connected component that will select to the next stage sorter; And

Image synthesizer is used for the connected component that will select of cascade classifier group afterbody sorter output is combined to form the image that will select.

32. device as claimed in claim 31 is characterized in that, described decomposer further comprises:

Treating apparatus is handled described image with non-linear Niblack thresholding method;

The picture breakdown device, being used for the picture breakdown after the described processing is connected component.

33. device as claimed in claim 32 is characterized in that, described non-linear Niblack thresholding method has respectively increased a statistics Order Statistic Filters in the background filter of standard Niblack method and prospect wave filter.

34. device as claimed in claim 33 is characterized in that, the formula of described non-linear Niblack thresholding method is:

$\mathrm{NLNiblack}(x,y)=\left\{\begin{array}{cc}1,& f(x,y)>T_{+}(x,y)\\ -1,& f(x,y)<T_{-}(x,y)\\ 0,& \mathrm{other}\end{array}\right.$

$T_{\&PlusMinus;}(x,y)={\widehat{\mathrm{\&mu;}}}_{p1}(x,y,W_B)\&PlusMinus;k\&CenterDot;{\widehat{\mathrm{\&sigma;}}}_{p2}(x,y,W_F)$

${\widehat{\mathrm{\&mu;}}}_{p1}=\mathrm{Order}[\mathrm{Mean}(f(x,y),W_B),p1,W_B]$

${\widehat{\mathrm{\&sigma;}}}_{p2}=\mathrm{Order}[\mathrm{Deviation}(f(x,y),W_F),p2,W_F]$

Wherein: k is set as 0.17-0.19 according to standard Niblack method;

(x y) is (x, y) the pixel intensity of position of input picture to f;

Mean (, be that window width is the mean filter of W W);

Deviation (, be that window width is the standard deviation wave filter of W W);

Order[, p, W] be to be number percent with p, W is the order statistics wave filter of width.

35. device as claimed in claim 34 is characterized in that, in background filter In, filter width W _BBe made as 1/16 of picture traverse, number percent p1 is made as 50%; At the prospect wave filter

In, width W _FBe W _B1/5, p2 is made as 80%.

36. device as claimed in claim 35 is characterized in that, each grade sorter in the described cascade classifier group comprises:

Calculation element is used for according to each grade sorter characteristic of correspondence, calculates the eigenwert of all connected components that receive;

Comparison means compares the eigenwert of all connected components respectively with the threshold interval of described feature, and

Output unit abandons the connected component of eigenwert outside described threshold interval as non-selected connected component; The connected component of eigenwert in described threshold interval exported to the next stage sorter as the connected component that will select.

37. device as claimed in claim 36, it is characterized in that, the connected component of described cascade classifier group output is further imported strong classifier, the sorter of described strong classifier for being trained by standard A daboost method, the feature of described strong classifier is identical with the feature of described cascade classifier group.

38. device as claimed in claim 37, it is characterized in that, described strong classifier carries out linear combination and judges whether this connected component is the connected component that will select all eigenwerts of the connected component of described cascade classifier group output, thereby abandon non-selected connected component, the connected component that output is selected.

39. as each described device of claim 31-38, it is characterized in that, the described image that will select is a text image, the described connected component that will select is the text connected component, described non-selected connected component is the connected component of non-text, and the feature relevant with text image comprises: geometric properties, contrast on border feature, shape canonical feature, stroke feature and Space Consistency feature.

40. device as claimed in claim 39 is characterized in that, described geometric properties comprises area ratio, length ratio and length breadth ratio, wherein,

Area ratio is the area of connected component and the ratio of image area, and its formula is:

$\mathrm{Feature}\_\mathrm{AreaRatio}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{\mathrm{Area}\left(\mathrm{Picture}\right)}$

The formula of length ratio is:

$\mathrm{Feature}\_\mathrm{MLengthRatio}=\frac{\max \{w,h\}}{\max \{\mathrm{PicW},\mathrm{PicH}\}}$

The formula of length breadth ratio is:

Feature_AspectRatio＝max{w/h，h/w}

In the above-mentioned formula, CC represents connected component, max{a, b} represents to get the higher value among a and the b, and w represents described connected component bounding box width value, and h represents the height value of described connected component bounding box, the width of PicW presentation video, and the height of PicH presentation video.

41. device as claimed in claim 39 is characterized in that, described contrast on border feature comprises contrast on border, and described contrast on border is the ratio on registration and the border of connected component of the edge image of the border of connected component and original image, and its formula is:

$\mathrm{EdgeContrast}=\frac{\mathrm{Border}\left(\mathrm{CC}\right)\&cap;\mathrm{Edge}\left(\mathrm{Picture}\right)}{\mathrm{Border}\left(\mathrm{CC}\right)}$

Wherein, Border (CC) is the border of connected component, and Edge (Picture) is the edge-detected image of original image, is the union of Canny operator and Sobel operator, and its formula is:

Edge(Picture)＝Canny(Picture)∪Sobel(Picture)

42. device as claimed in claim 39 is characterized in that, described shape canonical feature comprises connected component border roughness, empty number, and compactness and dutycycle, its formula is,

Connected component border roughness:

$\mathrm{Feature}\_\mathrm{ContourRoughness}=\frac{|\mathrm{CC}-\mathrm{open}(\mathrm{imfill}\left(\mathrm{CC}\right),2\&times;2)|}{\left|\mathrm{CC}\right|}$

The cavity number:

Feature_CCHoles=|imholes (CC) compactness:

$\mathrm{Feature}\_\mathrm{Compact}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{{\left[\mathrm{Perimeter}\left(\mathrm{CC}\right)\right]}^2}$

Dutycycle:

$\mathrm{Feature}\_\mathrm{OccupyRatio}=\frac{\mathrm{Area}\left(\mathrm{CC}\right)}{\mathrm{Area}\left(\mathrm{BoundingBox}\left(\mathrm{CC}\right)\right)}$

In the above-mentioned formula, imfill (CC) is the operation of filling up the inner hole of connected component, 2 * 2nd, the structural element of morphology opening operation, open () is the morphology opening operation in the following formula, be that connected component is carried out level and smooth operation, imholes (CC) expression connected component cavity number, Area (CC) expression connected component area, the girth of Perimeter (CC) expression connected component, the area of Area (BoundingBox (CC)) expression connected component bounding box.

43. device as claimed in claim 39 is characterized in that, described stroke feature comprises stroke mean breadth and stroke width standard deviation, wherein,

The stroke mean breadth:

Feature_Stroke_Mean＝Mean(strokeWidth(skeleton(CC)))

The stroke width standard deviation:

$\mathrm{Feature}\_\mathrm{Stroke}\_\mathrm{std}=\frac{\mathrm{Deviation}\left(\mathrm{strokeWidth}\left(\mathrm{skeleton}\left(\mathrm{CC}\right)\right)\right)}{\mathrm{Mean}\left(\mathrm{strokeWidth}\left(\mathrm{skeleton}\left(\mathrm{CC}\right)\right)\right)}$

In the above-mentioned formula, skeleton is a morphology skeleton algorithm, connected component is taken out skeleton and obtained skeleton diagram, the stroke width of strokeWidth for obtaining for every bit on the described skeleton diagram, Mean is for averaging a little on the described skeleton diagram, obtain mean breadth, standard deviation is asked in Deviation () expression.

44. device as claimed in claim 39 is characterized in that, described Space Consistency feature comprises Space Consistency area ratio and Space Consistency boundary characteristic, wherein,

The Space Consistency area ratio:

$\mathrm{Feature}\_\mathrm{AreaRatio}\_S=\frac{\mathrm{Area}\left(\mathrm{imdilate}(\mathrm{CC},5\&times;5)\right)}{\mathrm{Area}\left(\mathrm{Picture}\right)}$

The Space Consistency boundary characteristic:

Feature_Boundary_S＝Bound(imdilate(CC，5×5))

In the above-mentioned formula, imdilate is the operation that connected component is expanded, 5 * 5 for the structural element of expansive working, Area (imdilate (CC, 5 * 5)) area of the connected component after the expansive working is carried out in expression, the border of the connected component after Bound (imdilate (CC, 5 * 5)) expression the carrying out expansive working.

Images