CN105809181A - Logo detection method and device - Google Patents

Abstract

The invention provides a Logo detection method and device. According to the method and the device, an Adabooster algorithm is combined with a support vector machine algorithm to train Logo; a Logo model library which comprises an Adabooster model and a support vector machine model is acquired; a corresponding Logo model library is selected according to the related category information of an image to be detected; and the Adabooster model and the support vector machine model in the corresponding Logo model library is used to detect the image to be detected.

Description

Method and apparatus for Logo detection

Technical field

The application relates to communication and computer realm, particularly relates to a kind of method and apparatus for Logo detection.

Background technology

Logo is the foreign language abbreviation of logo or trade mark, is the abbreviation of LOGOtype, plays and logo has the identification of company and the effect of popularization, consumer can be allowed to remember company's main body and brand culture by the Logo of image.Logo in commodity picture identifies it is a kind of information retrieval method, is also hot issue in internet arena, has very important using value.

In existing identification technology, common application includes face, car plate, cloud atlas identification, wherein, human face five-sense-organ layout is substantially consistent, the change of car plate only numeral and letter, and the background color of face, car plate, cloud atlas changes not quite substantially, therefore have only to one model of training.

Compared to above-mentioned common application, the detection of Logo, identification major issue are the of a great variety of Logo, its application background also relative complex.

Presently the most conventional Logo recognition methods is based on SIFT feature point and carries out detecting, mating, and basic step is as follows:

(1) obtain image to be matched and Logo image, described image to be matched and Logo image are separately converted to gray-scale map；

(2) SIFT feature of two gray level images is extracted respectively；

(3) SIFT feature according to image to be matched, sets up KD-tree or other search trees；

(4) feature of Logo image is searched in KD-tree ((abbreviation of K-Dimensional tree)) similitude, if similitude is more than specific threshold, then show image to be matched has this Logo.

Other Logo recognition methodss also utilize shape facility, or increase the Logo recognition methods of geometrical constraint on SIFT feature basis, but and the method for this patent realize there is a great difference from principle and step.

The described method directly mated based on SIFT feature, maximum problem is non-normally low for fairly simple Logo discrimination, the Logo of such as certain brand is the shape that only one of which simply hooks, situation about identifying is very bad, it is a kind of local feature describing class angle point that reason is mainly SIFT, natively little for this kind of Logo being mainly made up of simple optical slip bar or the color lump SIFT feature that can extract, namely the Logo of the shape for simply hooking can only extract 4 characteristic points, make coupling extremely difficult, slightly deformation, the problem such as fuzzy all can mate less than.

Summary of the invention

The purpose of the application is to provide a kind of detection method detecting the Logo that speed is fast, false drop rate is low and equipment.

In view of this, the application provides a kind of method for Logo detection, and wherein, described method includes:

Utilizing A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained, to obtain the model library of described Logo, the model library of each described Logo includes A Dabusite model and the supporting vector machine model of this Logo；

Obtain image to be detected the model library of the related category information corresponding Logo of selection according to described image to be detected；

Utilize the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extract candidate region；And

Utilize the supporting vector machine model in the model library of corresponding Logo that described candidate region is detected, to obtain corresponding testing result.

Preferably, A Dabusite algorithm combination supporting vector machine algorithm is utilized to be trained including to Logo:

Collecting sample, described sample includes the positive sample of some single images with described Logo and the negative sample of some images not having described Logo；

Extract the fisrt feature collection of described sample, and utilize A Dabusite algorithm to be trained for described fisrt feature collection, to obtain A Dabusite model；

Collect the normal image with described Logo, utilize Logo described in described A Dabusite model inspection normal image, to obtain the candidate region with described Logo；And

Extract the second feature collection of described candidate region, and utilize described algorithm of support vector machine to be trained for described second feature collection, to obtain supporting vector machine model.

Preferably, A Dabusite algorithm combination supporting vector machine algorithm is utilized to be trained Logo also including:

Before extracting the fisrt feature collection of described sample, the sample of described collection being carried out pretreatment, described pretreatment includes described sample carries out gray processing process and/or image size registration process.

Preferably, adopt Lis Hartel to levy and calculate the fisrt feature collection extracting described sample.

Preferably, utilize A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained, also include:

Before extracting the second feature collection of described candidate region, the candidate region of described acquisition being carried out pretreatment, this pretreatment includes:

Described candidate region is cut into candidate image；

According to described candidate image, whether there is corresponding Logo and carry out positive negative flag；And

Described candidate image is carried out image size registration process.

Preferably, LBP feature or HOG feature calculation is adopted to extract the second feature collection of described candidate region.

Preferably, described method also includes:

Before described image to be detected detects, described image to be detected is carried out gray processing process and/or image size registration process.

Preferably, the model library of each described Logo all has some classification information labels.

Preferably, the model library of corresponding Logo is selected to include according to the related category information of described image to be detected:

Select the model library with all described Logo of the related category institute accordingly classification information label of described image to be detected.

Preferably, include after returning corresponding testing result:

If testing result is consistent, then stop continuing detection；

If testing result is inconsistent, then described image to be detected is detected by the model library continuing with described corresponding Logo.

The application also provides for a kind of equipment for Logo detection, and wherein, described equipment includes:

First device, is used for utilizing A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained, and to obtain the model library of described Logo, the model library of each described Logo includes A Dabusite model and the supporting vector machine model of this Logo；

Second device, is used for obtaining image to be detected the model library of the related category information corresponding Logo of selection according to described image to be detected；

3rd device, is used for utilizing the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extracts candidate region；

4th device, is used for utilizing the supporting vector machine model in the model library of corresponding Logo that described candidate region is detected, to obtain corresponding testing result.

Preferably, wherein, described first device includes:

First module, is used for collecting sample, and described sample includes the positive sample of some single images with described Logo and the negative sample of some images not having described Logo；

Second unit, for extracting the fisrt feature collection of described sample, and utilizes A Dabusite algorithm to be trained for described fisrt feature collection, to obtain A Dabusite model；

Unit the 3rd, for collecting the normal image with described Logo, utilizes Logo described in described A Dabusite model inspection normal image, to obtain the candidate region with described Logo；And

Unit the 4th, for extracting the second feature collection of described candidate region, utilizes described algorithm of support vector machine to be trained for described second feature collection, to obtain supporting vector machine model.

Preferably, described first device also includes:

Unit the 5th, for, before extracting the fisrt feature collection of described sample, the sample of described collection being carried out pretreatment, described pretreatment includes described sample carries out gray processing process and/or image size registration process.

Preferably, described second unit adopts Lis Hartel to levy and calculates the fisrt feature collection extracting described sample.

Preferably, described first device also includes Unit the 6th, and for, before extracting the second feature collection of described candidate region, the candidate region of described acquisition being carried out pretreatment, described Unit the 6th includes:

First subelement, for cutting into candidate image by described candidate region；

Second subelement, carries out positive negative flag for whether having corresponding Logo according to described candidate image；And

3rd subelement, carries out image size registration process to described candidate image.

Preferably, Unit the 4th adopts LBP feature or HOG feature calculation to extract the second feature collection of described candidate region, including:

Preferably, wherein, described equipment also includes:

5th device, for, before described image to be detected is detected, carrying out gray processing process and/or image size registration process to the image to be detected of described acquisition.

Preferably, the model library of each described Logo all has some classification information labels.

Preferably, described second device selects the model library of corresponding Logo to include according to the related category information of described image to be detected:

Select the model library with all described Logo of the related category institute accordingly classification information label of described image to be detected.

Preferably, described equipment, after returning corresponding testing result, also includes:

If testing result is consistent, then stop continuing detection；

If testing result is inconsistent, then described image to be detected is detected by the model library continuing with other corresponding Logo.

Compared with prior art, Logo is trained by the A Dabusite algorithm combination supporting vector machine algorithm that utilizes for Logo detection described herein, obtain the model library of the Logo including A Dabusite model and supporting vector machine model, and the model library of the related category information corresponding Logo of selection according to image to be detected, utilize the A Dabusite model in the model library of corresponding Logo and supporting vector machine model to image to be detected.

Further, after the model library forming some Logo, the model library of Logo is managed, the model library that related category information is each Logo according to Logo increases classification information label, before image to be detected is detected, it is possible to choose, according to the related category information of image to be detected, all Logo model libraries that respective classes information labels is corresponding.The model library of all Logo detects successively, when the model library testing result of a Logo is inconsistent, the model library then selecting next Logo carries out detecting until testing result is consistent, if all testing results are inconsistent, does not then find this image to be detected to have the Logo of related category.

Accompanying drawing explanation

By reading the detailed description that non-limiting example is made made with reference to the following drawings, other features, purpose and advantage will become more apparent upon:

Fig. 1 illustrates according to the equipment schematic diagram for Logo detection that the application provides on the one hand；

Fig. 2 illustrate according in the application one preferred embodiment for the schematic diagram of first device of the equipment of Logo detection；

Fig. 3 illustrate according in another preferred embodiment of the application for the schematic diagram of first device of the equipment of Logo detection；

Fig. 4 illustrate according in the another preferred embodiment of the application for the schematic diagram of first device of the equipment of Logo detection；

Fig. 5 illustrates the 6th cell schematics of first device according to the application one preferred embodiment；

Fig. 6 illustrates according to the equipment schematic diagram for Logo detection that the application one preferred embodiment provides；

Fig. 7 illustrates the method flow diagram realizing Logo detection according to the application one side；

Fig. 8 illustrates according to the method flow diagram that realization in the application one preferred embodiment utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained；

Fig. 9 illustrates according to the method flow diagram that realization in another preferred embodiment of the application utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained；

Figure 10 illustrates according to the method flow diagram that realization in the another preferred embodiment of the application utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained；

Figure 11 illustrates according to the method flow diagram realizing the candidate region obtained is carried out pretreatment in the application one preferred embodiment；

Figure 12 illustrates according to the method flow diagram that realization in the application one preferred embodiment utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained；

Figure 13 illustrates according to the process schematic utilizing LBP feature calculation second feature in the application one preferred embodiment.

In accompanying drawing, same or analogous accompanying drawing labelling represents same or analogous parts.

Detailed description of the invention

In one typical configuration of the application, terminal, the equipment of service network and trusted party all include one or more processor (CPU), input/output interface, network interface and internal memory.

Internal memory potentially includes the forms such as the volatile memory in computer-readable medium, random access memory (RAM) and/or Nonvolatile memory, such as read only memory (ROM) or flash memory (flashRAM).Internal memory is the example of computer-readable medium.

Computer-readable medium includes permanent and impermanency, removable and non-removable media can by any method or technology to realize information storage.Information can be computer-readable instruction, data structure, the unit of program or other data.The example of the storage medium of computer includes, but it is not limited to phase transition internal memory (PRAM), static RAM (SRAM), dynamic random access memory (DRAM), other kinds of random access memory (RAM), read only memory (ROM), Electrically Erasable Read Only Memory (EEPROM), fast flash memory bank or other memory techniques, read-only optical disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tape, magnetic disk storage or other magnetic storage apparatus or any other non-transmission medium, can be used for the information that storage can be accessed by a computing device.According to defining herein, computer-readable medium does not include non-temporary computer readable media (transitorymedia), such as data signal and the carrier wave of modulation.

Prior art processes Logo detection main employing SIFT algorithm, limitation in view of SIFT algorithm, present applicant proposes the Logo detection algorithm based on A Dabusite algorithm (Adaboost algorithm) and algorithm of support vector machine (SVM algorithm), with the test pattern for shape, color-variable and Logo, improve accuracy of detection, improve detection speed and reduce false drop rate.

Fig. 1 illustrates that in conjunction with Fig. 1, equipment 1 includes first device the 11, second device the 12, the 3rd device 13 and the 4th device 14 according to the equipment schematic diagram for Logo detection that the application provides on the one hand.

At this, described equipment 1 can be include a kind of can according to set in advance or storage instruction, automatically carry out the electronic equipment of numerical computations and information processing, its hardware includes but not limited to microprocessor, special IC (ASIC), programmable gate array (FPGA), digital processing unit (DSP), embedded device etc..Those skilled in the art will be understood that the said equipment 1 is only for example, and other existing or first user equipment 1 of being likely to occur from now on, as being applicable to the application, also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Wherein, first device 11 is used for utilizing A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained, and to obtain the model library of described Logo, the model library of each described Logo includes A Dabusite model and the supporting vector machine model of this Logo.Second device 12 is used for obtaining image to be detected the model library of the related category information corresponding Logo of selection according to described image to be detected；3rd device 13 is used for utilizing the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extracts candidate region；4th device 14 is used for utilizing the supporting vector machine model in the model library of corresponding Logo that described candidate region is detected, to obtain corresponding testing result.

At this, described A Dabusite algorithm data mining algorithm, specifically a kind of Boosting method.A Dabusite algorithm is a kind of iterative algorithm, its core concept is the grader (Weak Classifier) different for the training of same training set, then these weak classifier set are got up, constitute a higher final grader (strong classifier).Its algorithm itself realizes by changing data distribution, and whether it is correct according to the classification of sample each among each training set, and the accuracy rate of the general classification of last time, determines the weights of each sample.Give sub classification device by the new data set revising weights to be trained, finally the grader obtained will be trained finally to merge, as last Decision Classfication device every time.Use A Dabusite grader can get rid of some unnecessary training datas, and key is placed on above the training data of key.Described A Dabusite algorithm can be based upon in any sorting algorithm, it is possible to is decision tree, support vector machine etc..

Fig. 2 illustrates according to the application one preferred embodiment for the Logo first device schematic diagram detected.Further combined with Fig. 2, described first device 11 includes first module 101, second unit the 102, the 3rd unit 103 and the 4th unit 104.Wherein, first module 101 is used for collecting sample, and described sample 101 includes the positive sample of some single images with described Logo and the negative sample of some images not having described Logo；Second unit 102 is for extracting the fisrt feature collection of described sample, and utilizes A Dabusite algorithm to be trained for described fisrt feature collection, to obtain A Dabusite model；3rd unit 103, for collecting the normal image with described Logo, utilizes Logo described in described A Dabusite model inspection normal image, to obtain the candidate region with described Logo；And the 4th unit 104 for extracting the second feature collection of described candidate region, utilize described algorithm of support vector machine to be trained for described second feature collection, to obtain supporting vector machine model.

Concrete, the sample that first module 101 is collected includes the positive sample of some single images with described Logo and the negative sample of some images not having described Logo.The positive sample with the single image of Logo can be a sample through having this Logo image and solid background, it does not have the negative sample of the image of described Logo can be that various other do not have any picture of image of Logo.The number of described positive sample and negative sample is not limited, and the number of positive sample and negative sample is more many, and the training thereafter image of Logo carried out is more many, then the corresponding A Dabusite model obtained is then more accurate.

In the particular embodiment, remember that the first training set is for { X_1i, i=1 ..., k1, each sample is for being designated as X_1i, total total k1 opens, and each sample also has positive negative flag { Y_1i, i=1 ..., k1, wherein, the positive negative flag of positive sample: Y_1i=1, the positive and negative of negative sample is labeled as: Y_1i=0.

Those skilled in the art will be understood that above-mentioned collection sample and sample carries out the mode of positive negative flag are only for example; other existing or be likely to occur from now on collect sample and sample carried out the mode of positive negative flag as being applicable to the application; also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Fig. 3 illustrates according to another preferred embodiment of the application for the Logo first device schematic diagram detected.In conjunction with Fig. 3, in preferred enforcement, described first device 11 includes first module 101 ', second unit 102 ', the 3rd unit 103 ', the 4th unit 104 ' and the 5th unit 105 ', wherein, 5th unit 105 ' is for before extracting the fisrt feature collection of described sample, the sample of described collection is carried out pretreatment, and described pretreatment includes described sample carries out gray processing process and/or image size registration process.Adopt gray processing to process and/or described sample is carried out pretreatment and subsequent detection process computation can be made more easy by image size registration process, thus improving detection processing speed.At this, the first module 101 ' of first device 11, second unit 102 ', the 3rd unit 103 ' and the 4th unit 104 ' are identical or essentially identical with the first module 101 of first device in Fig. 2 11, second unit the 102, the 3rd unit 103 and the 4th unit 104 corresponding contents, for simplicity's sake, therefore do not repeat them here, and it is incorporated herein by reference.

Concrete, the 5th unit 105 ' can carry out gray processing process by following formula for R, G, the B value of each pixel of image of each sample: Gray=0.299*R+0.587*G+0.144*B.

Certainly, those skilled in the art will be understood that the mode that above-mentioned 5th unit 105 ' carries out gray processing process is only for example, and the mode that other gray processings that are existing or that be likely to occur from now on process is as being applicable to the application, for instance adopt HSV space: Gray_V=max (R, G, B), or adopt the simplest computational methods taking intermediate value: Gray=(R+G+B)/3, also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Then, the image that 5th unit 105 ' obtains after processing for gray processing, in the particular embodiment, assuming that the size respectively (w of image before and after image size registration process, h) and (wr, hr), then can according to (i " * w/wr; j " * h/hr) do image size registration process, the X after image size registration process "_1iEach pixel (i ", j ") and X '_1iIn the color value of pixel (i " * w/wr, j " * h/hr) identical.

Certainly; those skilled in the art will be understood that above-mentioned 5th unit 105 ' carry out image size alignmentization process mode be only for example; the mode that other image size alignmentization that are existing or that be likely to occur from now on process is as being applicable to the application; also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Then, second unit 102 extracts the fisrt feature collection of described sample, and utilizes A Dabusite algorithm to be trained for described fisrt feature collection, to obtain A Dabusite model.In preferred embodiment, described second unit 102 adopts Lis Hartel to levy and calculates the fisrt feature collection extracting described sample.

Concrete, second unit 102 adopts Lis Hartel to levy and is calculated as follows:

First, the full figure pixel integral image of sample, formula S (i, j)=SUM are calculated_{0<ii<i,0<jj<j(ii,jj)}, i.e. what the value of every was all little than current point equal to abscissa and vertical coordinate is had a sum.For the integrated value of rectangle any in figure, can obtain by the integral and calculating of four angle points:

S_rectangle=S_right-bottom-S_left-bottom-S_right-top+S_left-top；

Then, X is calculated according to integrogram "_1iHarr feature, Harr feature has five category feature S₁～S₅, specifically it is calculated as follows:

S₁=S_top-S_bottom；

S₂=S_left-S_right；

S₃=S_left+S_right-S_middle；

S₄=S_top+S_bottom-S_middle；

S₅=S_right-bottom+S_left-top-S_left-bottom-S_right-top。

The Lis Hartel that note wherein every sample obtains is levied as X " '_1i, then obtain Lis Hartel collection X " '_1iIt it is exactly the fisrt feature collection of whole sample set.

Certainly; those skilled in the art will be understood that above-mentioned second unit 102 extracts the mode of fisrt feature collection in the first training set and is only for example; other modes of extraction second feature collection V that are existing or that be likely to occur from now on; such as HOG feature or LBP feature etc. are as being applicable to the application; also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Then, second unit 102 utilize A Dabusite algorithm for described fisrt feature collection X " '_1i, in conjunction with the positive and negative label { Y of corresponding sample_1iBe trained, to obtain A Dabusite model, step is as follows:

Note (x_i,y_i), wherein x_iFor fisrt feature collection X " '_1iIn the characteristic vector of i-th samples pictures, y_iPositive and negative label { Y for sample_1iIn the positive and negative label of i-th samples pictures.

First, weights are initialized:

Work as y_iWhen=0, weights W_1,i=1/2M, works as y_iWhen=1, weights W_1,i=1/2L, wherein M is the total number of negative sample, and N is the total number of positive sample, M+N=k₁, K₁Sum for sample.

Then, circulation performs step (a)～step S (f) T time, (wherein T is artificial setting value, and T span can be 1～10 time, for instance 2,5,8 times):

The normalization of (a) weights: w_t,i=w_t,i/sum_{J=1 ..., n}(w_t,j)；

B (), to each feature j, training generates grader h_j, and mistake in computation rate e_j:

(c)e_j=sum_{I=1 ..., k1}(w_i*|h_j(x_i)-y_i|), wherein theta_jFor manually setting threshold value, specifically set according to training requirement, be typically set to 0.5, P_jFor positive sample labeling value or negative sample mark value, for instance be 1 or-1.

D () chooses minimal error rate e_tCorresponding grader h_t；

E () works as h_j(x_i)=y_iOr w_t+1,i=w_t,i, then weight w is updated_t+1,i=w_t,i*e^t/(1-e_t), in other situations, then repetitive cycling step (a)～(d).

F () is last, it is possible to obtain grader H (X):

Finally, H (X) result of generation is final judge mark, and five features that wherein Lis Hartel is levied need all to satisfy condition, and could generate the result of final judge mark, and namely training is formed and obtains A Dabusite model.

Then, the 3rd unit 103 collects the normal image with described Logo, utilizes Logo described in described A Dabusite model inspection normal image, to obtain the candidate region with described Logo, collects normal image as the second training set.

Fig. 4 illustrate according in the another preferred embodiment of the application for the schematic diagram of first device of the equipment of Logo detection.In conjunction with Fig. 4, described first device 11 first module 101 ", second unit 102 ", the 3rd unit 103 ", the 4th unit 104 " and the 5th unit 105 " and the 6th unit 106 ", the 6th unit 106 " for described candidate region is carried out pretreatment.At this, the first module 101 of first device 11 ", second unit 102 ", the 3rd unit 103 " and the 4th unit 104 " identical or essentially identical with the first module 101 of first device in Fig. 2 11, second unit the 102, the 3rd unit 103 and the 4th unit 104 corresponding contents, 5th unit 105 of first device 11 " identical or essentially identical with the 5th unit 105 ' content described in Fig. 3; for simplicity's sake; therefore do not repeat them here, and be incorporated herein by reference.

Fig. 5 illustrates the 6th cell schematics of first device according to the application one preferred embodiment, in conjunction with Fig. 5, and described 6th unit 106 " include the first subelement the 601, second subelement 602 and the 3rd subelement 603.Wherein the first subelement 601 is for cutting into candidate image by described candidate region；Second subelement 602 carries out positive negative flag for whether having corresponding Logo according to described candidate image；3rd subelement 603 for carrying out image size registration process to described candidate image.6th unit 106 " described candidate region is carried out pretreatment subsequent detection process computation can be made more easy, thus improving detection processing speed.

Concrete, these candidate regions are all cut out by described first subelement 601 classifies as independent candidate image, mode classification can use manual type or automated manner, second subelement 602 does the positive negative flag of positive negative sample after sorting, i.e. 0 or 1 labelling, then, described candidate image is carried out image size registration process by the 3rd subelement 603, zooms to unification and is sized for the 3rd training set.

Then, described 4th unit 104 extracts second feature collection in the 3rd training set, wherein, second feature collection can be chosen and adopt Lis Hartel to levy the LBP (LocalBinaryPattern that extraction is different from fisrt feature, local binary patterns) or HOG feature (HistogramofOrientedGradient, HOG), it is possible to improve the accuracy of detection further.

At this, utilize LBP feature calculation second feature identical with the process content shown in Figure 13 or essentially identical, for simplicity's sake, therefore do not repeat them here, and be incorporated herein by reference.

Additionally, utilize the process of HOG feature calculation second feature to include:

I () calculates the gradient of each pixel of image, obtain size and direction；

(ii) divide an image into block of cells, such as the block of cells of 6x6, add up the histogram of gradients in each block of cells, obtain description of each block of cells；

(iii) 3x3 block of cells is formed a bulk, sub description just obtaining this bulk that is together in series of the description of all block of cells of each bulk, description of bulks all in image is together in series and just obtains the HOG feature of this image.This characteristic vector just can be used to train SVM.

Certainly; those skilled in the art will be understood that above-mentioned 4th unit 104 extracts the mode of second feature collection V in the 3rd training set and is only for example; other modes of extraction second feature collection V that are existing or that be likely to occur from now on; such as Lis Hartel is levied as being applicable to the application; also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Then, the 4th unit 104 utilize second feature collection Vi} train, to obtain supporting vector machine model (SVM model library).Utilize second feature training SVM model process, including:

(1) first calculate LaGrange parameter, set (x_i,y_i) it is training data, wherein, x_iIt is that sample i obtains characteristic vector V, y_i=0 (negative sample) or y_i=1 (positive sample):

Work as α_i>=0, i=1 ..., n, andTime,

$\underset{\mathrm{\&alpha;}}{\max }\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_i{x}_i^T{x}_j$

(2) calculate weight w and b, weight w and b be two parameters in SVM model.

$w=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{y}_i^{\left(i\right)}{x}^{\left(i\right)}$

$b*=-\frac{\stackrel{\max }{i:y^{\left(i\right)}=-1}{w}^{*T}{x}^{\left(i\right)}+\stackrel{\max }{i:y^{\left(i\right)}=1}{w}^{*T}{x}^{\left(i\right)}}{2}$

(3) final judge mark y is obtained:

Y=w*x+b.

Utilize first device 11 that various Logo are trained, to obtain the model library of numerous Logo, the model library of numerous Logo is carried out unified management, by creating classification information label, the model library of Logo is sorted out.Wherein, described classification information can be classified according to contents such as the businessman that Logo represents, product, COSs, the such as Logo of certain motion brand can have clothes, trousers, shoes, the classification information label such as ball, then follow-up image to be detected is carried out corresponding Logo detection time, its relevant classification information can be differentiated with regard to the content of image to be detected, one of some classifications of such as image to be detected are clothes, then when subsequent detection, and the Logo model library selecting class label information to be clothes.Thus solving the problem that the Logo detection difficulty caused of a great variety is big, reducing detection difficulty, improve detection efficiency.

Then, second device 12 obtains image to be detected, and during according to the model library of the related category information corresponding Logo of selection of described image to be detected, can according to the related category information of Logo, model library for each described Logo increases some different classification information labels, when needing image to be detected is detected, first the relevant information of the Logo detected can be needed to determine the classification information label needing detection according to image to be detected, call the model library of all Logo with category information further according to classification information label.

Then, 3rd device 13 utilizes the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extracts candidate region, if testing result is inconsistent, the Logo not responded in image to be detected is then described, then directly returns.

When testing result is consistent, then the 4th device 14 utilizes the supporting vector machine model in the model library of corresponding Logo that described candidate region is detected, to obtain corresponding testing result.

Image to be detected is carried out the process that detection process can be a circulation by the model library utilizing corresponding Logo, if testing result is inconsistent when image to be detected is detected by the model library utilizing a corresponding Logo, image to be detected is detected by the model library then utilizing next corresponding Logo, until the testing result obtained is consistent.

Fig. 6 illustrates according to the equipment schematic diagram for Logo detection that the application one preferred embodiment provides；As shown in Figure 6, described equipment 1 includes first device 11 ', the second device 12 ', the 3rd device 13 ' the 4th device 14 ' and the 5th device 15 '.Wherein, the 5th device 15 ' is for, before described image to be detected is detected, carrying out gray processing process and/or image size registration process to the image to be detected of described acquisition.At this, it is identical or essentially identical with first device the 11, second device the 12, the 3rd device 13 the 4th device 14 corresponding contents of equipment in Fig. 11 that described equipment 1 includes first device 11 ', the second device 12 ', the 3rd device 13 ' the 4th device 14 ', for simplicity's sake, therefore do not repeat them here, and it is incorporated herein by reference.

Fig. 7 illustrates the method flow diagram realizing Logo detection according to the application one side, and in conjunction with Fig. 7, described method includes step S01, step S02, step S03 and step S04.

Wherein, in step S01, utilizing A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained, to obtain the model library of described Logo, the model library of each described Logo includes A Dabusite model and the supporting vector machine model of this Logo；In step S02, obtain image to be detected the model library of the related category information corresponding Logo of selection according to described image to be detected；In step S03, utilize the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extract candidate region；In step S04, utilize the supporting vector machine model in the model library of corresponding Logo that described candidate region is detected, to obtain corresponding testing result.

Fig. 8 illustrates according to the method flow diagram that realization in the application one preferred embodiment utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained.Further combined with Fig. 8, step S01 specifically includes step S101, step S102, step S103 and step S104.Wherein, in step S101, collecting sample, described sample 101 includes the positive sample of some single images with described Logo and the negative sample of some images not having described Logo；In step s 102, extract the fisrt feature collection of described sample, and utilize A Dabusite algorithm to be trained for described fisrt feature collection, to obtain A Dabusite model；In step s 103, collect the normal image with described Logo, utilize Logo described in described A Dabusite model inspection normal image, to obtain the candidate region with described Logo；And in step S104, extract the second feature collection of described candidate region, utilize described algorithm of support vector machine to be trained for described second feature collection, to obtain supporting vector machine model.

Concrete, in step S101, the sample of collection includes the positive sample of some single images with described Logo and the negative sample of some images not having described Logo.The positive sample with the single image of Logo can be a sample through having this Logo image and solid background, it does not have the negative sample of the image of described Logo can be that various other do not have any picture of image of Logo.The number of described positive sample and negative sample is not limited, and the number of positive sample and negative sample is more many, and the training thereafter image of Logo carried out is more many, then the corresponding A Dabusite model obtained is then more accurate.

In the particular embodiment, remember that the first training set is for { X_1i, i=1 ..., k1, each sample is for being designated as X_1i, total total k1 opens, and each sample also has positive negative flag { Y_1i, i=1 ..., k1, wherein, the positive negative flag of positive sample: Y_1i=1, the positive and negative of negative sample is labeled as: Y_1i=0.

Those skilled in the art will be understood that above-mentioned collection sample and sample carries out the mode of positive negative flag are only for example; other existing or be likely to occur from now on collect sample and sample carried out the mode of positive negative flag as being applicable to the application; also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Fig. 9 illustrates according to the method flow diagram that realization in another preferred embodiment of the application utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained.In conjunction with Fig. 3, in preferred enforcement, step S01 specifically includes step S101 ', step S102 ', step S103 ', step S104 ' and step S105 ', wherein, in step S105 ', the sample of described collection is carried out pretreatment, and described pretreatment includes described sample carries out gray processing process and/or image size registration process.Adopt gray processing to process and/or described sample is carried out pretreatment and subsequent detection process computation can be made more easy by image size registration process, thus improving detection processing speed.At this, step S101 ', step S102 ', step S103 ' and step S104 ' are identical or essentially identical with step S101, step S102, step S103 and step S104 corresponding contents in Fig. 8, for simplicity's sake, therefore do not repeat them here, and be incorporated herein by reference.

Concrete, in step S105 ', R, G, the B value for each pixel of image of each sample can carry out gray processing process by following formula: Gray=0.299*R+0.587*G+0.144*B.

Certainly, those skilled in the art will be understood that the above-mentioned mode carrying out gray processing process in step S105 ' is only for example, and the mode that other gray processings that are existing or that be likely to occur from now on process is as being applicable to the application, for instance adopt HSV space: Gray_V=max (R, G, B), or adopt the simplest computational methods taking intermediate value: Gray=(R+G+B)/3, also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Then, in step S105 ', the image obtained after gray processing is processed, in the particular embodiment, it is assumed that before and after image size registration process, respectively (w, h) with (wr for the size of image, hr), then can do image size registration process according to (i " * w/wr, j " * h/hr), the X after image size registration process "_1iEach pixel (i ", j ") and X '_1iIn the color value of pixel (i " * w/wr, j " * h/hr) identical.

Certainly; those skilled in the art will be understood that the above-mentioned mode carrying out the process of image size alignmentization in step S105 ' is only for example; the mode that other image size alignmentization that are existing or that be likely to occur from now on process is as being applicable to the application; also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Then, in step s 102, extract the fisrt feature collection of described sample, and utilize A Dabusite algorithm to be trained for described fisrt feature collection, to obtain A Dabusite model.In preferred embodiment, described second unit 102 adopts Lis Hartel to levy (Haar feature, moment characteristics) and calculates the fisrt feature collection extracting described sample.

Concrete, in step s 102, adopt Lis Hartel to levy and be calculated as follows:

First, the full figure pixel integral image of sample, formula S (i, j)=SUM are calculated_{0<ii<i,0<jj<j(ii,jj)}, i.e. what the value of every was all little than current point equal to abscissa and vertical coordinate is had a sum.For the integrated value of rectangle any in figure, can obtain by the integral and calculating of four angle points:

S_rectangle=S_right-bottom-S_left-bottom-S_right-top+S_left-top；

Then, X is calculated according to integrogram "_1iHarr feature, Harr feature has five category features, is specifically calculated as follows:

S₁=S_top-S_bottom；

S₂=S_left-S_right；

S₃=S_left+S_right-S_middle；

S₄=S_top+S_bottom-S_middle；

S₅=S_right-bottom+S_left-top-S_left-bottom-S_right-top。

The Lis Hartel that note wherein every sample obtains is levied as X " '_1i, then obtain Lis Hartel collection X " '_1iIt it is exactly the fisrt feature collection of whole sample set.

Certainly; those skilled in the art will be understood that the above-mentioned mode extracting fisrt feature collection in step s 102 in the first training set is only for example; other modes of extraction second feature collection V that are existing or that be likely to occur from now on; such as HOG feature or LBP feature etc. are as being applicable to the application; also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Then, in step s 102, utilize A Dabusite algorithm for described fisrt feature collection X " '_1i, in conjunction with the positive and negative label { Y of corresponding sample_1iBe trained, to obtain A Dabusite model, step is as follows:

Note (x_i,y_i), wherein x_iFor fisrt feature collection X " '_1iIn the characteristic vector of i-th samples pictures, y_iPositive and negative label { Y for sample_1iIn the positive and negative label of i-th samples pictures.

First, weights are initialized:

Work as y_iWhen=0, weights W_1,i=1/2M, works as y_iWhen=1, weights W_1,i=1/2L, wherein M is the total number of negative sample, and N is the total number of positive sample, M+N=k1, and K1 is the sum of sample.

Then, circulation performs step (a)～step S (f) T time, (wherein T is artificial setting value, and T value can be 1～10 time, for instance 2,5,8 times):

The normalization of (a) weights: w_t,i=w_t,i/sum_{J=1 ..., n}(w_t,j)；

B (), to each feature j, training generates grader h_j, and mistake in computation rate e_j:

(c)e_j=sum_{I=1 ..., k1}(w_i*|h_j(x_i)-y_i|), wherein theta_jFor manually setting threshold value, specifically set according to training requirement, be typically set to 0.5, P_jFor positive sample labeling value or negative sample mark value, for instance be 1 or-1.

D () chooses minimal error rate e_tCorresponding grader h_t；

E () works as h_j(x_i)=y_iOr w_t+1,i=w_t,i, then weight w is updated_t+1,i=w_t,i*e^t/(1-e_t), in other situations, then repetitive cycling step (a)～(d).

F () is last, it is possible to obtain grader H (X):

Finally, H (X) result of generation is final judge mark, and five features that wherein Lis Hartel is levied need all to satisfy condition, and could generate the result of final judge mark, and namely training is formed and obtains A Dabusite model.

Then, in step s 103, collect the normal image with described Logo, utilize Logo described in described A Dabusite model inspection normal image, to obtain the candidate region with described Logo, collect normal image as the second training set.

Figure 10 illustrates according to the method flow diagram that realization in the another preferred embodiment of the application utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained.In conjunction with Figure 10, in preferred enforcement, step S01 specifically includes step S101 ", step S102 ", step S103 ", step S104 ", step S105 " and step S106 ", wherein, in step S105 " in; the sample of described collection is carried out pretreatment, and described pretreatment includes described sample carries out gray processing process and/or image size registration process.Adopt gray processing to process and/or described sample is carried out pretreatment and subsequent detection process computation can be made more easy by image size registration process, thus improving detection processing speed.At this, step S101 ", step S102 ", step S103 " and step S104 " identical or essentially identical with step S101, step S102, step S103 and step S104 corresponding contents in Fig. 8, step S105 " identical or essentially identical with step S105 ' corresponding contents in Fig. 9; for simplicity's sake; therefore do not repeat them here, and be incorporated herein by reference.

Figure 11 illustrates according to the method flow diagram realizing the candidate region obtained is carried out pretreatment in the application one preferred embodiment.In conjunction with step S106 described in Figure 11, Figure 10 " include step S601, step S602 and step S603.Wherein, in step s 601, described candidate region is cut into candidate image；In step s 601, according to described candidate image, whether there is corresponding Logo and carry out positive negative flag；In step s 601, described candidate image is carried out image size registration process.Step S106 " described candidate region is carried out pretreatment subsequent detection process computation can be made more easy, thus improving detection processing speed.

Concrete, in step s 601, all being cut out these candidate regions and classify as independent candidate image, mode classification can use manual type or automated manner, in step S602, do the positive negative flag of positive negative sample after sorting, i.e. 0 or 1 labelling, then, in step S603, described candidate image is carried out image size registration process, zooms to unification and be sized for the 3rd training set.

Then, in conjunction with Fig. 8, in step S14, extracting second feature collection in the 3rd training set, wherein, second feature collection can be chosen and adopt Lis Hartel to levy LBP or the HOG feature that extraction is different from fisrt feature, it is possible to improve the accuracy of detection further.

Figure 13 illustrates according to the process schematic utilizing LBP feature calculation second feature in the application one preferred embodiment.As shown in figure 13, the calculating process of LBP feature is as follows: extract certain particular neighborhood of image, such as 3x3 or 5x5 region, the value taking its center pixel is c, then by its neighborhood the value of pixel a little compare size with c, if more than c, be designated as 1, be then designated as 0 less than c, a string 0,1 string obtained is LBP feature.Such as, it is characterized by being calculated what LBP combined by all 3x3 fields in image, the pixel value distribution obtained as shown in Figure 13 (b) it is computed by the pixel map of Figure 13 (a), the value of the pixel at its center is 83, the value of the pixel at the pixel in other neighborhoods and its center compares, comparative result is such as shown in Figure 13 (c), and by the result of acquisition from the upper left corner by being recorded as 01111100 clockwise, then the value of feature is 01111100=124.

In another embodiment, the calculating process of HOG feature is as follows:

I () calculates the gradient of each pixel of image, obtain size and direction；

(ii) divide an image into block of cells, such as the block of cells of 6x6, add up the histogram of gradients in each block of cells, obtain description of each block of cells；

(iii) 3x3 block of cells is formed a bulk, sub description just obtaining this bulk that is together in series of the description of all block of cells of each bulk, sub being together in series that describe of bulks all in image is just obtained the HOG feature of this image, and the final characteristic vector obtained just can be used to train SVM.

Certainly; those skilled in the art will be understood that above-mentioned in step S14; the mode extracting second feature collection V in the 3rd training set is only for example; other modes of extraction second feature collection V that are existing or that be likely to occur from now on; such as Lis Hartel is levied as being applicable to the application; also should be included within the application protection domain, and be incorporated herein with way of reference at this.

Then, the 4th unit 104 utilize second feature collection Vi} train, to obtain supporting vector machine model (SVM model library).Utilize second feature training SVM model process, including:

(1) first calculate LaGrange parameter, set (x_i,y_i) it is training data, wherein, x_iIt is that sample i obtains characteristic vector V, y_i=0 (negative sample) or y_i=1 (positive sample):

Work as α_i>=0, i=1 ..., n, andTime,

$\underset{\mathrm{\&alpha;}}{\max }\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i,j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_i{x}_i^T{x}_j$

(2) calculate weight w and b, weight w and b be two parameters in SVM model.

$w=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{y}_i^{\left(i\right)}{x}^{\left(i\right)}$

$b*=-\frac{\stackrel{\max }{i:y^{\left(i\right)}=-1}{w}^{*T}{x}^{\left(i\right)}+\stackrel{\max }{i:y^{\left(i\right)}=1}{w}^{*T}{x}^{\left(i\right)}}{2}$

(3) final judge mark y is obtained:

Y=w*x+b.

Then, by various Logo are trained, it is possible to obtain the model library of numerous Logo, the model library of numerous Logo is carried out unified management, by creating classification information label, the model library of Logo is sorted out.Various Logo are trained by step S01, to obtain the model library of numerous Logo, the model library of numerous Logo are carried out unified management, by creating classification information label, the model library of Logo is sorted out.Wherein, described classification information can be classified according to contents such as the businessman that Logo represents, product, COSs, the such as Logo of certain motion brand can have clothes, trousers, shoes, the classification information label such as ball, then follow-up image to be detected is carried out corresponding Logo detection time, its relevant classification information can be differentiated with regard to the content of image to be detected, one of some classifications of such as image to be detected are clothes, then when subsequent detection, and the Logo model library selecting class label information to be clothes.Thus solving the problem that the Logo detection difficulty caused of a great variety is big, reducing detection difficulty, improve detection efficiency.

For, before described image to be detected is detected, the image to be detected of described acquisition being carried out gray processing process and/or image size registration process.

Then, in step S02, obtain image to be detected, and during according to the model library of the related category information corresponding Logo of selection of described image to be detected, can according to the related category information of Logo, model library for each described Logo increases some different classification information labels, when needing image to be detected is detected, first the relevant information of the Logo detected can be needed to determine the classification information label needing detection according to image to be detected, call the model library of all Logo with category information further according to classification information label.

Then, in step S03, utilize the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extract candidate region, if testing result is inconsistent, then the Logo not responded in image to be detected is described, then directly returns.

When testing result is consistent, then in step S04, utilize the supporting vector machine model in the model library of corresponding Logo that described candidate region is detected, to obtain corresponding testing result.

Image to be detected is carried out the process that detection process can be a circulation by the model library utilizing corresponding Logo, if testing result is inconsistent when image to be detected is detected by the model library utilizing a corresponding Logo, image to be detected is detected by the model library then utilizing next corresponding Logo, until the testing result obtained is consistent.

Figure 12 illustrates according to the method flow diagram that realization in the application one preferred embodiment utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained；As shown in figure 12, described method includes step S01 ', step S02 ', step S03 ', step S04 ' and step S05 '.Wherein, the image to be detected of described acquisition, between step S01 ' and step S02 ', is carried out gray processing process and/or image size registration process, so that subsequent detection process computation is more easy, thus improving detection processing speed by step S05 '.At this, it is identical or essentially identical with step S01, step S02, step S03 and step S04 corresponding contents in Fig. 7 that described method includes step S01 ', step S02 ', step S03 ' and step S04 ', for simplicity's sake, therefore do not repeat them here, and be incorporated herein by reference.

In the particular embodiment, the herein described method for Logo detection, first numerous sample is collected, the negative sample of positive sample that each sample includes having the image of Logo and the image without Logo, utilizing A Dabusite algorithm combination supporting vector machine algorithm that sample is trained obtaining the model library of various Logo, the model library of each described Logo includes A Dabusite model and the supporting vector machine model of this Logo.When image to be detected is detected by needs, first the model library of corresponding Logo is selected according to the related category information of described image to be detected, then utilize the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extract candidate region, described candidate region is detected by the supporting vector machine model recycled in the model library of corresponding Logo, to obtain corresponding testing result.The model library of all Logo detects successively, when the model library testing result of a Logo is inconsistent, the model library then selecting next Logo carries out detecting until testing result is consistent, if all testing results are inconsistent, this image to be detected is not then found to have the Logo of related category, so that the detection process of Logo completes fast and effectively, additionally, the detection that employing A Dabusite model and supporting vector machine model can make Logo is more accurate.

In sum, Logo is trained by the A Dabusite algorithm combination supporting vector machine algorithm that utilizes for Logo detection described herein, obtain the model library of the Logo including A Dabusite model and supporting vector machine model, and the model library of the related category information corresponding Logo of selection according to image to be detected, utilize the A Dabusite model in the model library of corresponding Logo and supporting vector machine model to image to be detected.

Further, after the model library forming some Logo, the model library of Logo is managed, the model library that related category information is each Logo according to Logo increases classification information label, before image to be detected is detected, it is possible to choose, according to the related category information of image to be detected, all Logo model libraries that respective classes information labels is corresponding.The model library of all Logo detects successively, when the model library testing result of a Logo is inconsistent, the model library then selecting next Logo carries out detecting until testing result is consistent, if all testing results are inconsistent, does not then find this image to be detected to have the Logo of related category.

Obviously, the application can be carried out various change and modification without deviating from spirit and scope by those skilled in the art.So, if these amendments of the application and modification belong within the scope of the application claim and equivalent technologies thereof, then the application is also intended to comprise these change and modification.

It should be noted that the application can be implemented in the assembly of software and/or software and hardware, for instance, special IC (ASIC), general purpose computer or any other similar hardware device can be adopted to realize.In one embodiment, the software program of the application can perform to realize steps described above or function by processor.Similarly, the software program of the application can be stored in computer readable recording medium storing program for performing (including the data structure being correlated with), for instance, RAM memory, magnetically or optically driver or floppy disc and similar devices.It addition, some steps of the application or function can employ hardware to realize, for instance, as coordinating with processor thus performing the circuit of each step or function.

It addition, the part of the application can be applied to computer program, for instance computer program instructions, when it is computer-executed, by the operation of this computer, it is possible to call or provide according to the present processes and/or technical scheme.And call the programmed instruction of the present processes, it is possibly stored in fixing or moveable record medium, and/or by broadcast or data stream in other signal bearing medias and be transmitted, and/or be stored in the working storage of the computer equipment run according to described programmed instruction.At this, an embodiment according to the application includes a device, this device includes the memorizer for storing computer program instructions and for performing the processor of programmed instruction, wherein, when this computer program instructions is performed by this processor, trigger this plant running based on the method for aforementioned multiple embodiments according to the application and/or technical scheme.

It is obvious to a person skilled in the art that the application is not limited to the details of above-mentioned one exemplary embodiment, and when without departing substantially from spirit herein or basic feature, it is possible to realize the application in other specific forms.Therefore, no matter from which point, embodiment all should be regarded as exemplary, and be nonrestrictive, scope of the present application is limited by claims rather than described above, it is intended that all changes in the implication of the equivalency dropping on claim and scope be included in the application.Any accompanying drawing labelling in claim should be considered as the claim that restriction is involved.Furthermore, it is to be understood that " including " word is not excluded for other unit or step, odd number is not excluded for plural number.Multiple unit or the device stated in device claim can also be realized by software or hardware by a unit or device.The first, the second word such as grade is used for representing title, and is not offered as any specific order.

Claims (20)

1., for a method for Logo detection, wherein, described method includes:

Utilizing A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained, to obtain the model library of described Logo, the model library of each described Logo includes A Dabusite model and the supporting vector machine model of this Logo；

Obtain image to be detected the model library of the related category information corresponding Logo of selection according to described image to be detected；

Utilize the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extract candidate region；And

Utilize the supporting vector machine model in the model library of corresponding Logo that described candidate region is detected, to obtain corresponding testing result.

2. method according to claim 1, wherein, utilizes A Dabusite algorithm combination supporting vector machine algorithm to be trained including to Logo:

Collecting sample, described sample includes the positive sample of some single images with described Logo and the negative sample of some images not having described Logo；

Extract the fisrt feature collection of described sample, and utilize A Dabusite algorithm to be trained for described fisrt feature collection, to obtain A Dabusite model；

Collect the normal image with described Logo, utilize Logo described in described A Dabusite model inspection normal image, to obtain the candidate region with described Logo；And

Extract the second feature collection of described candidate region, and utilize described algorithm of support vector machine to be trained for described second feature collection, to obtain supporting vector machine model.

3. method according to claim 2, wherein, utilizes A Dabusite algorithm combination supporting vector machine algorithm to be trained Logo also including:

Before extracting the fisrt feature collection of described sample, the sample of described collection being carried out pretreatment, described pretreatment includes described sample carries out gray processing process and/or image size registration process.

4. according to the method in claim 2 or 3, wherein, adopt Lis Hartel to levy and calculate the fisrt feature collection extracting described sample.

5. the method according to any one of claim 2 to 4, wherein, utilizes A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained, also includes:

Before extracting the second feature collection of described candidate region, the candidate region of described acquisition being carried out pretreatment, this pretreatment includes:

Described candidate region is cut into candidate image；

According to described candidate image, whether there is corresponding Logo and carry out positive negative flag；And

Described candidate image is carried out image size registration process.

6. the method according to any one of claim 2 to 5, wherein, adopts LBP feature or HOG feature calculation to extract the second feature collection of described candidate region.

7. method according to any one of claim 1 to 6, wherein, described method also includes:

Before described image to be detected detects, described image to be detected is carried out gray processing process and/or image size registration process.

8. method according to any one of claim 1 to 7, wherein, the model library of each described Logo all has some classification information labels.

9. method according to claim 8, wherein, selects the model library of corresponding Logo to include according to the related category information of described image to be detected:

Select the model library with all described Logo of the related category institute accordingly classification information label of described image to be detected.

10. method according to any one of claim 1 to 9, wherein, includes after returning corresponding testing result:

If testing result is consistent, then stop continuing detection；

If testing result is inconsistent, then described image to be detected is detected by the model library continuing with described corresponding Logo.

11. for an equipment for Logo detection, wherein, described equipment includes:

First device, is used for utilizing A Dabusite algorithm combination supporting vector machine algorithm that Logo is trained, and to obtain the model library of described Logo, the model library of each described Logo includes A Dabusite model and the supporting vector machine model of this Logo；

Second device, is used for obtaining image to be detected the model library of the related category information corresponding Logo of selection according to described image to be detected；

3rd device, is used for utilizing the A Dabusite model in the model library of corresponding Logo that described image to be detected is detected, if testing result is for consistent, extracts candidate region；

4th device, is used for utilizing the supporting vector machine model in the model library of corresponding Logo that described candidate region is detected, to obtain corresponding testing result.

12. equipment according to claim 11, wherein, described first device includes:

First module, is used for collecting sample, and described sample includes the positive sample of some single images with described Logo and the negative sample of some images not having described Logo；

Second unit, for extracting the fisrt feature collection of described sample, and utilizes A Dabusite algorithm to be trained for described fisrt feature collection, to obtain A Dabusite model；

Unit the 3rd, for collecting the normal image with described Logo, utilizes Logo described in described A Dabusite model inspection normal image, to obtain the candidate region with described Logo；And

Unit the 4th, for extracting the second feature collection of described candidate region, utilizes described algorithm of support vector machine to be trained for described second feature collection, to obtain supporting vector machine model.

13. equipment according to claim 12, wherein, described first device also includes:

Unit the 5th, for, before extracting the fisrt feature collection of described sample, the sample of described collection being carried out pretreatment, described pretreatment includes described sample carries out gray processing process and/or image size registration process.

14. the equipment according to claim 12 or 13, wherein, described second unit adopts Lis Hartel to levy and calculates the fisrt feature collection extracting described sample.

15. the equipment according to any one of claim 12 to 14, wherein, described first device also includes Unit the 6th, for before extracting the second feature collection of described candidate region, the candidate region of described acquisition is carried out pretreatment, and described Unit the 6th includes:

First subelement, for cutting into candidate image by described candidate region；

Second subelement, carries out positive negative flag for whether having corresponding Logo according to described candidate image；And

3rd subelement, carries out image size registration process to described candidate image.

16. the equipment according to any one of claim 11 to 15, wherein, Unit the 4th adopts LBP feature or HOG feature calculation to extract the second feature collection of described candidate region, including.

17. the equipment according to any one of claim 11 to 16, wherein, described equipment also includes:

5th device, for, before described image to be detected is detected, carrying out gray processing process and/or image size registration process to the image to be detected of described acquisition.

18. the equipment according to any one of claim 11 to 17, wherein, the model library of each described Logo all has some classification information labels.

19. equipment according to claim 18, wherein, described second device selects the model library of corresponding Logo to include according to the related category information of described image to be detected:

Select the model library with all described Logo of the related category institute accordingly classification information label of described image to be detected.

20. the equipment according to any one of claim 11 to 19, wherein, described equipment, after returning corresponding testing result, also includes:

If testing result is consistent, then stop continuing detection；

If testing result is inconsistent, then described image to be detected is detected by the model library continuing with other corresponding Logo.