CN105809181B - Method and apparatus for Logo detection - Google Patents
Method and apparatus for Logo detection Download PDFInfo
- Publication number
- CN105809181B CN105809181B CN201410855921.6A CN201410855921A CN105809181B CN 105809181 B CN105809181 B CN 105809181B CN 201410855921 A CN201410855921 A CN 201410855921A CN 105809181 B CN105809181 B CN 105809181B
- Authority
- CN
- China
- Prior art keywords
- logo
- image
- model
- detected
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Image Analysis (AREA)
Abstract
The application provides a kind of method and apparatus for Logo detection, the method and equipment utilization Ah's Da Busite algorithm combination supporting vector machine algorithm are trained Logo, obtain the model library of the Logo including Ah Da Busite model and supporting vector machine model, and the model library of corresponding Logo is selected according to the related category information of image to be detected, using in the model library of corresponding Logo Ah Da Busite model and supporting vector machine model to image to be detected.
Description
Technical field
This application involves communication and computer field more particularly to a kind of methods and apparatus for Logo detection.
Background technique
Logo is the foreign language abbreviation of logo or trade mark, is the abbreviation of LOGO type, plays the knowledge for possessing logo company
Not with the effect of popularization, consumer can be allowed to remember company's main body and brand culture by the Logo of image.In commodity picture
Logo identification is hot issue in a kind of information retrieval method and internet area, has very important application value.
Common application field includes face, license plate, cloud atlas identification in existing identification technology, wherein human face five-sense-organ is laid out base
This is all consistent, and license plate only has the variation of number and letter, and the background color of face, license plate, cloud atlas change substantially less, because
This needs one model of training.
Compared to above-mentioned common application field, the detection of Logo, identification major issue are many kinds of of Logo, application
Background is also relative complex.
Presently the most common Logo recognition methods is to be detected based on SIFT feature, is matched, and basic step is such as
Under:
(1) image to be matched and Logo image are obtained, the image to be matched and Logo image are separately converted to gray scale
Figure；
(2) SIFT feature of two gray level images is extracted respectively；
(3) according to the SIFT feature of image to be matched, KD-tree or other search trees are established；
(4) feature of Logo image is searched into similitude in KD-tree ((abbreviation of K-Dimensional tree)), such as
Fruit similitude is greater than specific threshold, then shows there is the Logo in image to be matched.
Other Logo recognition methods also have using shape feature, or increase the Logo of geometrical constraint on the basis of SIFT feature
Recognition methods, but realize from principle and step and be all very different with the method for this patent.
Described to be based on the directly matched method of SIFT feature, maximum problem is non-for fairly simple Logo discrimination
Often low, for example the Logo of certain brand is shape that only one is simply hooked, the case where identification, is very bad, and reason is mainly
SIFT is a kind of local feature for describing class angle point, for can mainly be mentioned by this kind of Logo that simple fair line item or color lump form
The SIFT feature got is natively seldom, i.e., can only extract 4 characteristic points for the Logo of the shape simply hooked, so that
With extremely difficult, slightly deformation, it is fuzzy the problems such as can all match less than.
Summary of the invention
The purpose of the application is to provide the detection method and equipment of a kind of detection Logo that speed is fast, false detection rate is low.
In view of this, the application provides a kind of method for Logo detection, wherein the described method includes:
Logo is trained using Ah's Da Busite algorithm combination supporting vector machine algorithm, to obtain the mould of the Logo
Type library, the model library of each Logo include the Ah Da Busite model and supporting vector machine model of the Logo；
Image to be detected is obtained, and selects the model of corresponding Logo according to the related category information of described image to be detected
Library；
Described image to be detected is detected using the Ah Da Busite model in the model library of corresponding Logo,
Candidate region is extracted if being unanimously if testing result；And
The candidate region is detected using the supporting vector machine model in the model library of corresponding Logo, with
Obtain corresponding testing result.
Preferably, Logo is trained using Ah's Da Busite algorithm combination supporting vector machine algorithm and includes:
Collect sample, the sample include several single images with the Logo positive sample and it is several without described in
The negative sample of the image of Logo；
The fisrt feature collection of the sample is extracted, and is instructed using Ah's Da Busite algorithm for the fisrt feature collection
Practice, to obtain Ah's Da Busite model；
The normal image with the Logo is collected, using described in the Ah Da Busite model inspection normal image
Logo, to obtain the candidate region with the Logo；And
The second feature collection of the candidate region is extracted, and is directed to the second feature using the algorithm of support vector machine
Collection is trained, to obtain supporting vector machine model.
Preferably, Logo is trained using Ah's Da Busite algorithm combination supporting vector machine algorithm further include:
Before the fisrt feature collection for extracting the sample, the collected sample is pre-processed, the pretreatment
Including carrying out gray processing processing and/or image size registration process to the sample.
Preferably, it is levied using Lis Hartel and calculates the fisrt feature collection for extracting the sample.
Preferably, Logo is trained using Ah's Da Busite algorithm combination supporting vector machine algorithm, further includes:
Before the second feature collection for extracting the candidate region, the candidate region of the acquisition is pre-processed, it should
Pretreatment includes:
The candidate region is cut into candidate image；
According to the candidate image whether there is corresponding Logo to carry out positive negative flag；And
Image size registration process is carried out to the candidate image.
Preferably, the second feature collection of the candidate region is extracted using LBP feature or HOG feature calculation.
Preferably, the method also includes:
Before described image to be detected is detected, gray processing processing and/or image are carried out to described image to be detected
Size registration process.
Preferably, the model library of each Logo all has several classification information labels.
Preferably, the model library for selecting corresponding Logo according to the related category information of described image to be detected includes:
Select to have the related category institute of described image to be detected all Logo of classification information label accordingly
Model library.
Preferably, include: after returning to corresponding testing result
If testing result be it is consistent, stopping continue to test；
If testing result be it is inconsistent, continue with the model library of the corresponding Logo and described image to be detected carried out
Detection.
The application also provides a kind of equipment for Logo detection, wherein the equipment includes:
First device, for being trained using Ah's Da Busite algorithm combination supporting vector machine algorithm to Logo, to obtain
The model library of the Logo is obtained, the model library of each Logo includes the Ah Da Busite model and support vector machines of the Logo
Model；
Second device selects phase for obtaining image to be detected, and according to the related category information of described image to be detected
The model library of the Logo answered；
3rd device, the Ah Da Busite model in model library for utilizing corresponding Logo is to described to be detected
Image is detected, and extracts candidate region if being unanimously if testing result；
4th device, the supporting vector machine model in model library for utilizing corresponding Logo is to the candidate regions
Domain is detected, to obtain corresponding testing result.
Preferably, wherein the first device includes:
First unit, for collecting sample, the sample includes the positive sample of several single images with the Logo
With the negative sample of several images without the Logo；
Second unit, for extracting the fisrt feature collection of the sample, and using Ah Da Busite algorithm for described the
One feature set is trained, to obtain Ah's Da Busite model；
Third unit, it is general using the Ah Da Busite model inspection for collecting the normal image with the Logo
Logo described in logical image, to obtain the candidate region with the Logo；And
Unit the 4th is directed to for extracting the second feature collection of the candidate region using the algorithm of support vector machine
The second feature collection is trained, to obtain supporting vector machine model.
Preferably, the first device further include:
Unit the 5th, for being located in advance before the fisrt feature collection for extracting the sample to the collected sample
Reason, the pretreatment include carrying out gray processing processing and/or image size registration process to the sample.
Preferably, the second unit calculates the fisrt feature collection for extracting the sample using Lis Hartel sign.
Preferably, the first device further includes Unit the 6th, in the second feature collection for extracting the candidate region
Before, the candidate region of the acquisition is pre-processed, Unit the 6th includes:
First subelement, for the candidate region to be cut into candidate image；
Second subelement, for according to the candidate image whether there is corresponding Logo to carry out positive negative flag；And
Third subelement carries out image size registration process to the candidate image.
Preferably, Unit the 4th extracts the second feature collection of the candidate region using LBP feature or HOG feature calculation,
Include:
Preferably, wherein the equipment further include:
5th device, for before being detected to described image to be detected, to image to be detected of the acquisition into
The processing of row gray processing and/or image size registration process.
Preferably, the model library of each Logo all has several classification information labels.
Preferably, the second device selects the mould of corresponding Logo according to the related category information of described image to be detected
Type library includes:
Select to have the related category institute of described image to be detected all Logo of classification information label accordingly
Model library.
Preferably, the equipment is after returning to corresponding testing result, further includes:
If testing result be it is consistent, stopping continue to test；
If testing result be it is inconsistent, continue with the model library of other corresponding Logo and described image to be detected carried out
Detection.
Compared with prior art, it is described herein for Logo detection using Ah Da Busite algorithm combine support to
Amount machine algorithm is trained Logo, obtains the model library of the Logo including Ah Da Busite model and supporting vector machine model,
And the model library of corresponding Logo is selected according to the related category information of image to be detected, in the model library using corresponding Logo
Ah Da Busite model and supporting vector machine model to image to be detected.
Further, after the model library for forming several Logo, the model library of Logo is managed, according to the phase of Logo
Closing classification information can basis before detecting to image to be detected for the model library increase classification information label of every Logo
The related category information of image to be detected chooses the corresponding all Logo model libraries of respective classes information labels.The mould of all Logo
Type library is successively detected, when the model library testing result of a Logo be it is inconsistent when, then select the model library of next Logo into
Row detection until testing result be it is consistent, if all testing results be it is inconsistent, do not find the image to be detected have phase
Close the Logo of classification.
Detailed description of the invention
By reading a detailed description of non-restrictive embodiments in the light of the attached drawings below, the application's is other
Feature, objects and advantages will become more apparent upon:
Fig. 1 shows the equipment schematic diagram for Logo detection on the one hand provided according to the application；
Fig. 2 shows the signals according to the first device for being used for the equipment that Logo is detected in one preferred embodiment of the application
Figure；
Fig. 3 shows the signal of the first device according to the equipment for being used for Logo detection in another preferred embodiment of the application
Figure；
Fig. 4 shows the signal of the first device according to the equipment for being used for Logo detection in the another preferred embodiment of the application
Figure；
Fig. 5 shows the 6th cell schematics of the first device according to one preferred embodiment of the application；
Fig. 6 shows the equipment schematic diagram for Logo detection provided according to one preferred embodiment of the application；
Fig. 7 shows the method flow diagram for realizing Logo detection according to the application one side；
Fig. 8, which is shown, utilizes Ah Da Busite algorithm combination supporting vector machine according to realization in one preferred embodiment of the application
The method flow diagram that algorithm is trained Logo；
Fig. 9, which is shown, utilizes Ah Da Busite algorithm combination supporting vector according to realization in another preferred embodiment of the application
The method flow diagram that machine algorithm is trained Logo；
Figure 10 show according in the another preferred embodiment of the application realize using Ah Da Busite algorithm combine support to
The method flow diagram that amount machine algorithm is trained Logo；
Figure 11 is shown according to the pretreated side of candidate region progress realized in one preferred embodiment of the application to acquisition
Method flow chart；
Figure 12, which is shown, utilizes Ah Da Busite algorithm combination supporting vector according to realization in one preferred embodiment of the application
The method flow diagram that machine algorithm is trained Logo；
Figure 13 is shown according to the process signal for utilizing LBP feature calculation second feature in one preferred embodiment of the application
Figure.
The same or similar appended drawing reference represents the same or similar component in attached drawing.
Specific embodiment
In a typical configuration of this application, terminal, the equipment of service network and trusted party include one or more
Processor (CPU), input/output interface, network interface and memory.
Memory may include the non-volatile memory in computer-readable medium, random access memory (RAM) and/or
The forms such as Nonvolatile memory, such as read-only memory (ROM) or flash memory (flashRAM).Memory is showing for computer-readable medium
Example.
Computer-readable medium includes permanent and non-permanent, removable and non-removable media can be by any method
Or technology come realize information store.Information can be computer readable instructions, data structure, the unit of program or other data.
The example of the storage medium of computer includes, but are not limited to phase change memory (PRAM), static random access memory (SRAM), moves
State random access memory (DRAM), other kinds of random access memory (RAM), read-only memory (ROM), electric erasable
Programmable read only memory (EEPROM), flash memory or other memory techniques, read-only disc read only memory (CD-ROM) (CD-ROM),
Digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or
Any other non-transmission medium, can be used for storage can be accessed by a computing device information.As defined in this article, computer
Readable medium does not include non-temporary computer readable media (transitory media), such as the data-signal and carrier wave of modulation.
The prior art handles Logo detection and mainly uses SIFT algorithm, and in view of the limitation of SIFT algorithm, the application is proposed
It is based on the Logo detection algorithm of Ah Da Busite algorithm (Adaboost algorithm) and algorithm of support vector machine (SVM algorithm), with
For the test pattern of shape, color-variable and Logo, detection accuracy is improved, detection speed is improved and reduces false detection rate.
Fig. 1 shows the equipment schematic diagram for Logo detection on the one hand provided according to the application, and in conjunction with Fig. 1, equipment 1 is wrapped
Include first device 11, second device 12,3rd device 13 and the 4th device 14.
Here, the equipment 1 can be including it is a kind of can according to the instruction for being previously set or storing, carry out numerical value automatically
It calculates and the electronic equipment of information processing, hardware includes but is not limited to microprocessor, specific integrated circuit (ASIC), may be programmed
Gate array (FPGA), digital processing unit (DSP), embedded device etc..Those skilled in the art will be understood that above equipment 1 is only
Citing, other first user equipmenies 1 that are existing or being likely to occur from now on are such as applicable to the application, should also be included in the application
Within protection scope, and it is incorporated herein by reference.
Wherein, first device 11 is for instructing Logo using Ah's Da Busite algorithm combination supporting vector machine algorithm
Practice, to obtain the model library of the Logo, the model library of each Logo includes the Ah Da Busite model and branch of the Logo
Hold vector machine model.Second device 12 is used to obtain image to be detected, and according to the related category information of described image to be detected
Select the model library of corresponding Logo；3rd device 13 is used for the Ah Da Busite in the model library using corresponding Logo
Model detects described image to be detected, extracts candidate region if being unanimously if testing result；4th device 14 is for benefit
The candidate region is detected with the supporting vector machine model in the model library of corresponding Logo, to obtain corresponding inspection
Survey result.
Here, the Ah Da Busite algorithm data mining algorithm, specifically a kind of Boosting method.Ah Da Busite
Algorithm is a kind of iterative algorithm, and core concept is the classifier (Weak Classifier) different for the training of the same training set, so
These weak classifier sets are got up afterwards, constitute a stronger final classification device (strong classifier).Its algorithm itself is to pass through
Change data distribution come what is realized, whether it is correct and last time total according to the classification of each sample among each training set
The accuracy rate of body classification, to determine the weight of each sample.The new data set for modifying weight is given to sub-classification device to carry out
Training finally finally merges the classifier that each training obtains, as last Decision Classfication device.Use Ah 's Buss
Special classifier can exclude some unnecessary training datas, and key is placed on above crucial training data.Ah the reaching
Buss spy's algorithm can establish in any sorting algorithm, can be decision tree, support vector machines etc..
Fig. 2 shows the first device schematic diagrames that Logo detection is used for according to one preferred embodiment of the application.Further knot
Fig. 2 is closed, the first device 11 includes first unit 101, second unit 102, third unit 103 and the 4th unit 104.Its
In, first unit 101 includes the positive sample of several single images with the Logo for collecting sample, the sample 101
With the negative sample of several images without the Logo；Second unit 102 is used to extract the fisrt feature collection of the sample, and benefit
It is trained with Ah's Da Busite algorithm for the fisrt feature collection, to obtain Ah's Da Busite model；Third unit 103 is used
In collecting the normal image with the Logo, using Logo described in the Ah Da Busite model inspection normal image, with
Obtain the candidate region with the Logo；And the 4th unit 104 be used to extract the second feature collection of the candidate region, benefit
It is trained with the algorithm of support vector machine for the second feature collection, to obtain supporting vector machine model.
Specifically, first unit 101 collect sample include several single images with the Logo positive sample with
The negative sample of several images without the Logo.The positive sample of single image with Logo can be one and be somebody's turn to do through having
The sample of Logo image and solid background, without the image of the Logo negative sample can be it is various other without Logo
Any picture of image.The positive sample and the number of negative sample are not limited, and the number of positive sample and negative sample is more, thereafter
The training carried out to the image of Logo is more, then the Ah's Da Busite model accordingly obtained is then more accurate.
In the particular embodiment, the first training set of note is { X1i, i=1 ..., k1, each sample are to be denoted as X1i, in total
There are k1, each sample also has positive negative flag { Y1i, i=1 ..., k1, wherein the positive negative flag of positive sample: Y1i=1, it bears
The positive and negative label of sample are as follows: Y1i=0.
Those skilled in the art will be understood that above-mentioned collection sample and the mode for carrying out positive negative flag to sample are only for example,
Other existing or collection samples being likely to occur from now on and the mode for carrying out positive negative flag to sample are such as applicable to the application,
It should also be included within the application protection scope, and be incorporated herein by reference.
Fig. 3 shows the first device schematic diagram that Logo detection is used for according to another preferred embodiment of the application.In conjunction with figure
3, in preferred implement, the first device 11 includes first unit 101 ', second unit 102 ', third unit 103 ', the
Four units 104 ' and the 5th unit 105 ', wherein the 5th unit 105 ' is used for before the fisrt feature collection for extracting the sample,
The collected sample is pre-processed, the pretreatment includes big to sample progress gray processing processing and/or image
Small registration process.Using gray processing handle and/or image size registration process the sample is pre-processed can make it is subsequent
Detection process calculating is more easy, to improve detection processing speed.Here, the first unit 101 ' of first device 11, second
The first unit 101 of first device 11, second unit in unit 102 ', third unit 103 ' and the 4th unit 104 ' and Fig. 2
102, third unit 103 and 104 corresponding contents of Unit the 4th are identical or essentially identical, and for simplicity, therefore details are not described herein,
And it is incorporated herein by reference.
Specifically, the R of the 5th unit 105 ' pixel each for the image of each sample, G, B value can carry out as the following formula
Gray processing processing: Gray=0.299*R+0.587*G+0.144*B.
Certainly, the mode that those skilled in the art will be understood that above-mentioned 5th unit 105 ' carries out gray processing processing is only to lift
Example, the mode that other gray processings that are existing or being likely to occur from now on are handled such as is applicable to the application, for example, by using HSV sky
Between: GrayV=max (R, G, B), or using the simplest calculation method for taking intermediate value: Gray=(R+G+B)/3 also should include
Within the application protection scope, and it is incorporated herein by reference.
Then, the image that the 5th unit 105 ' obtains after handling for gray processing, in the particular embodiment, it is assumed that image
The size of image is respectively (w, h) and (wr, hr) before and after size registration process, then can do figure according to (i " * w/wr, j " * h/hr)
As size registration process, the X " after image size registration process1iEach pixel (i ", j ") and X '1iIn pixel
The color value of (i " * w/wr, j " * h/hr) is identical.
Certainly, those skilled in the art will be understood that above-mentioned 5th unit 105 ' carries out the side of image size alignmentization processing
Formula is only for example, and the mode that other image size alignmentization that are existing or being likely to occur from now on are handled such as is applicable to this Shen
Please, it should also be included within the application protection scope, and be incorporated herein by reference.
Then, second unit 102 extracts the fisrt feature collection of the sample, and using Ah Da Busite algorithm for described
Fisrt feature collection is trained, to obtain Ah's Da Busite model.In the preferred embodiment, the second unit 102 uses
Lis Hartel sign calculates the fisrt feature collection for extracting the sample.
Specifically, second unit 102 is as follows using Lis Hartel sign calculating:
Firstly, calculating the full figure pixel integral image of sample, formula S (i, j)=SUM0<ii<i,0<jj<j(ii,jj), that is, every point
Value be equal to abscissa and all smaller than current point the sum of all the points of ordinate.For the integrated value of rectangle any in figure, all may be used
To be obtained with the integral calculation of four angle points:
Srectangle=Sright-bottom-Sleft-bottom-Sright-top+Sleft-top；
Then, X " is calculated according to integrogram1iHarr feature, Harr feature has five category feature S1~S5, specific to calculate such as
Under:
S1=Stop-Sbottom；
S2=Sleft-Sright；
S3=Sleft+Sright-Smiddle；
S4=Stop+Sbottom-Smiddle；
S5=Sright-bottom+Sleft-top-Sleft-bottom-Sright-top。
The note Lis Hartel that wherein every sample obtains sign is X " '1i, then obtain Lis Hartel collection { X " '1iIt is exactly entire
The fisrt feature collection of sample set.
Certainly, those skilled in the art will be understood that above-mentioned second unit 102 extracts fisrt feature on the first training set
The mode of collection is only for example, other modes of extraction second feature collection V that are existing or being likely to occur from now on, such as HOG feature
Or LBP feature etc. should also be included within the application protection scope, and include by reference herein as being applicable to the application
In this.
Then, second unit 102 is directed to the fisrt feature collection { X " ' using Ah Da Busite algorithm1i, in conjunction with corresponding
Positive and negative label { the Y of sample1iBe trained, to obtain Ah's Da Busite model, steps are as follows:
Remember (xi,yi), wherein xiFor fisrt feature collection { X " '1iIn i-th of samples pictures feature vector, yiFor sample
Positive and negative label { Y1iIn the positive and negative label of i-th of samples pictures.
Firstly, initialization weight:
Work as yiWhen=0, weight W1,i=1/2M, works as yiWhen=1, weight W1,i=1/2L, wherein M is the total number of negative sample,
N is the total number of positive sample, M+N=k1, K1For the sum of sample.
Then, circulation executes step (a)~step S (f) T times, and (wherein T is artificial setting value, and T value range can be 1
~10 times, such as 2,5,8 times):
(a) normalization of weight: wt,i=wt,i/sumJ=1 ..., n(wt,j)；
(b) to each feature j, training generates classifier hj, and calculate error rate ej:
(c)ej=sumI=1 ..., k1(wi*|hj(xi)-yi|), wherein thetajTo manually set threshold value, according to training requirement
It is specifically set, is typically set to 0.5, PjThe sample labeling value that is positive or negative sample mark value, for example, 1 or -1.
(d) minimal error rate e is chosentCorresponding classifier ht；
(e) work as hj(xi)=yiOr wt+1,i=wt,i, then weight w is updatedt+1,i=wt,i*et/(1-et), in the case of other, then
Repetitive cycling step (a)~(d).
(f) finally, available classifier H (X):
Finally, H (X) result of generation is final judge mark, and wherein five features of Lis Hartel sign need all full
Sufficient condition, could generate final judge mark obtains Ah Da Busite model as a result, training to be formed.
Then, third unit 103 collects the normal image with the Logo, utilizes the Ah Da Busite model inspection
Logo described in normal image collects normal image as the second training set to obtain the candidate region with the Logo.
Fig. 4 shows the signal of the first device according to the equipment for being used for Logo detection in the another preferred embodiment of the application
Figure.In conjunction with Fig. 4,11 first unit 101 " of first device, second unit 102 ", third unit 103 ", the 4th unit 104 "
With the 5th unit 105 " and the 6th unit 106 ", the 6th unit 106 " is for pre-processing the candidate region.Here, the
First device in the first unit 101 " of one device 11, second unit 102 ", third unit 103 " and the 4th unit 104 " and Fig. 2
11 first unit 101, second unit 102, third unit 103 and 104 corresponding contents of Unit the 4th are identical or essentially identical, and
The 5th unit 105 " and the 105 ' content of Unit the 5th described in Fig. 3 of one device 11 are identical or essentially identical, for simplicity, therefore
Details are not described herein, and is incorporated herein by reference.
Fig. 5 shows the 6th cell schematics of the first device according to one preferred embodiment of the application, in conjunction with figure
5, the 6th unit 106 " includes the first subelement 601, the second subelement 602 and third subelement 603.Wherein the first son is single
Member 601 is for being cut into candidate image for the candidate region；Whether the second subelement 602 according to the candidate image for having
There is corresponding Logo to carry out positive negative flag；Third subelement 603 is used to carry out image size registration process to the candidate image.
6th unit 106 ", which pre-processes the candidate region, can make the calculating of subsequent detection process more easy, to improve inspection
Survey processing speed.
Specifically, these candidate regions are all cut into as individual candidate image by first subelement 601
Classify, manual type or automated manner can be used in mode classification, and the second subelement 602 does positive negative sample after sorting
Positive negative flag, i.e., 0 or 1 label, then, third subelement 603 carry out image size registration process, scaling to the candidate image
To being uniformly sized for third training set.
Then, the 4th unit 104 extracts second feature collection on third training set, wherein second feature collection can be with
Choose with fisrt feature using Lis Hartel levy the different LBP (Local Binary Pattern, local binary patterns) of extraction or
HOG feature (Histogram of Oriented Gradient, HOG), can be further improved the accuracy of detection.
Here, it is identical or essentially identical using LBP feature calculation second feature and process content shown in Figure 13, it is concise
For the sake of, therefore details are not described herein, and is incorporated herein by reference.
In addition, including: using the process of HOG feature calculation second feature
(i) gradient for calculating each pixel of image, obtains size and direction；
(ii) block of cells is divided an image into, such as the block of cells of 6x6, the histogram of gradients in each block of cells is counted, obtains
To description of each block of cells；
(iii) as soon as 3x3 block of cells is formed a bulk, description of all block of cells of each bulk is together in series
Description of the bulk has been obtained, description of bulks all in image is together in series and has just obtained the HOG feature of the image.
This feature vector can be used to train SVM.
Certainly, those skilled in the art will be understood that above-mentioned 4th unit 104 extracts second feature on third training set
The mode of collection V is only for example, other modes of extraction second feature collection V that are existing or being likely to occur from now on, such as Lis Hartel
Sign is such as applicable to the application, should also be included within the application protection scope, and be incorporated herein by reference.
Then, the 4th unit 104 is using second feature collection { Vi } training, to obtain supporting vector machine model (SVM model
Library).Utilize second feature training SVM model process, comprising:
(1) LaGrange parameter is first calculated, (x is seti,yi) it is training data, wherein xiIt is that sample i obtains feature vector
V, yi=0 (negative sample) or yi=1 (positive sample):
Work as αi>=0, i=1 ..., n, andWhen,
(2) weight w and b are being calculated, weight w and b are two parameters in SVM model.
(3) final judge mark y is obtained:
Y=w*x+b.
Various Logo are trained using first device 11, to obtain the model library of numerous Logo, to numerous Logo's
Model library is managed collectively, and is sorted out by creating classification information label to the model library of Logo.Wherein, the classification letter
The contents such as businessman, product, service type that breath can be represented according to Logo be classified, such as the Logo of certain movement brand can be with
It can then when the subsequent corresponding Logo detection to image to be detected progress with the classification informations label such as clothes, trousers, shoes, ball
Identifying one of its relevant classification information, such as several classifications of image to be detected with the content with regard to image to be detected is clothes,
Then in subsequent detection, select class label information for the Logo model library of clothes.To solve, Logo is many kinds of to be caused
The big problem of detection difficulty, reduce detection difficulty, improve detection efficiency.
Then, second device 12 obtains image to be detected, and is selected according to the related category information of described image to be detected
It, can be according to the related category information of Logo, if the model library for each Logo increases when the model library of corresponding Logo
Different classification information labels is done, it, can be first according to image to be detected need when needing to detect image to be detected
The relevant information for the Logo to be detected determines the classification information label for needing to detect, and calls all tools further according to classification information label
There is the model library of all Logo of category information.
Then, 3rd device 13 is using the Ah Da Busite model in the model library of corresponding Logo to described to be checked
Altimetric image is detected, if testing result be it is consistent if extract candidate region, if testing result be it is inconsistent, illustrate to be detected
The Logo not responded in image is then directly returned.
When testing result is consistent, then the 4th device 14 utilizes the supporting vector in the model library of corresponding Logo
Machine model detects the candidate region, to obtain corresponding testing result.
Carrying out detection process to image to be detected using the model library of corresponding Logo can be the process of a circulation, benefit
If when being detected with the model library of a corresponding Logo to image to be detected testing result be it is inconsistent, using next corresponding
The model library of Logo image to be detected is detected, until the testing result obtained is consistent.
Fig. 6 shows the equipment schematic diagram for Logo detection provided according to one preferred embodiment of the application；Such as Fig. 6 institute
Show, the equipment 1 includes first device 11 ', second device 12 ', the 4th device 14 ' of 3rd device 13 ' and the 5th device 15 '.
Wherein, the 5th device 15 ' is used for before detecting to described image to be detected, is carried out to image to be detected of the acquisition
Gray processing processing and/or image size registration process.Here, the equipment 1 includes first device 11 ', second device 12 ', the
First device 11, second device 12, the 4th device 14 of 3rd device 13 of equipment 1 in the 4th device 14 ' and Fig. 1 of three devices 13 '
Corresponding contents are identical or essentially identical, and for simplicity, therefore details are not described herein, and is incorporated herein by reference.
Fig. 7 shows the method flow diagram for realizing Logo detection according to the application one side, in conjunction with Fig. 7, the method packet
Include step S01, step S02, step S03 and step S04.
Wherein, in step S01, Logo is trained using Ah's Da Busite algorithm combination supporting vector machine algorithm,
To obtain the model library of the Logo, the model library of each Logo include the Ah Da Busite model of the Logo and support to
Amount machine model；In step S02, image to be detected is obtained, and phase is selected according to the related category information of described image to be detected
The model library of the Logo answered；In step S03, using the Ah Da Busite model in the model library of corresponding Logo to institute
It states image to be detected to be detected, extracts candidate region if being unanimously if testing result；In step S04, using described corresponding
Logo model library in supporting vector machine model the candidate region is detected, to obtain corresponding testing result.
Fig. 8, which is shown, utilizes Ah Da Busite algorithm combination supporting vector machine according to realization in one preferred embodiment of the application
The method flow diagram that algorithm is trained Logo.Further combined with Fig. 8, step S101, step are specifically included in step S01
S102, step S103 and step S104.Wherein, in step s101, sample is collected, the sample 101 includes several with institute
State the positive sample of the single image of Logo and the negative sample of several images without the Logo；In step s 102, institute is extracted
The fisrt feature collection of sample is stated, and is trained using Ah's Da Busite algorithm for the fisrt feature collection, to obtain Ah reaching
Buss spy's model；In step s 103, the normal image with the Logo is collected, the Ah Da Busite model inspection is utilized
Logo described in normal image, to obtain the candidate region with the Logo；And in step S104, the candidate is extracted
The second feature collection in region is trained, to be supported for the second feature collection using the algorithm of support vector machine
Vector machine model.
Specifically, in step s101, the sample of collection includes the positive sample of several single images with the Logo
With the negative sample of several images without the Logo.The positive sample of single image with Logo can be one and be somebody's turn to do through having
The sample of Logo image and solid background, without the image of the Logo negative sample can be it is various other without Logo
Any picture of image.The positive sample and the number of negative sample are not limited, and the number of positive sample and negative sample is more, thereafter
The training carried out to the image of Logo is more, then the Ah's Da Busite model accordingly obtained is then more accurate.
In the particular embodiment, the first training set of note is { X1i, i=1 ..., k1, each sample are to be denoted as X1i, in total
There are k1, each sample also has positive negative flag { Y1i, i=1 ..., k1, wherein the positive negative flag of positive sample: Y1i=1, it bears
The positive and negative label of sample are as follows: Y1i=0.
Those skilled in the art will be understood that above-mentioned collection sample and the mode for carrying out positive negative flag to sample are only for example,
Other existing or collection samples being likely to occur from now on and the mode for carrying out positive negative flag to sample are such as applicable to the application,
It should also be included within the application protection scope, and be incorporated herein by reference.
Fig. 9, which is shown, utilizes Ah Da Busite algorithm combination supporting vector according to realization in another preferred embodiment of the application
The method flow diagram that machine algorithm is trained Logo.Step is specifically included in step S01 in preferred implement in conjunction with Fig. 3
S101 ', step S102 ', step S103 ', step S104 ' and step S105 ', wherein in step S105 ', to the collection
Sample pre-processed, the pretreatment includes carrying out gray processing processing and/or image size registration process to the sample.
Being pre-processed using gray processing processing and/or image size registration process to the sample can be such that subsequent detection process calculates
It is more easy, to improve detection processing speed.Here, step S101 ', step S102 ', step S103 ' and step S104 ' with
Step S101, step S102, step S103 and step S104 corresponding contents are identical or essentially identical in Fig. 8, for simplicity, therefore
Details are not described herein, and is incorporated herein by reference.
Specifically, in step S105 ', the R of pixel each for the image of each sample, G, B value can as the following formula into
Row gray processing processing: Gray=0.299*R+0.587*G+0.144*B.
Certainly, those skilled in the art will be understood that the above-mentioned mode that gray processing processing is carried out in step S105 ' is only
Citing, the mode that other gray processings that are existing or being likely to occur from now on are handled such as is applicable to the application, for example, by using HSV sky
Between: GrayV=max (R, G, B), or using the simplest calculation method for taking intermediate value: Gray=(R+G+B)/3 also should include
Within the application protection scope, and it is incorporated herein by reference.
Then, in step S105 ', for the image obtained after gray processing processing, in the particular embodiment, it is assumed that figure
As the size of image before and after size registration process is respectively (w, h) and (wr, hr), then can be done according to (i " * w/wr, j " * h/hr)
Image size registration process, the X " after image size registration process1iEach pixel (i ", j ") and X '1iIn pixel
The color value of (i " * w/wr, j " * h/hr) is identical.
Certainly, those skilled in the art will be understood that the above-mentioned progress image size alignmentization processing in step S105 '
Mode is only for example, and the mode that other image size alignmentization that are existing or being likely to occur from now on are handled such as is applicable to this Shen
Please, it should also be included within the application protection scope, and be incorporated herein by reference.
Then, in step s 102, the fisrt feature collection of the sample is extracted, and is directed to institute using Ah Da Busite algorithm
It states fisrt feature collection to be trained, to obtain Ah's Da Busite model.In the preferred embodiment, the second unit 102 is adopted
The fisrt feature collection for extracting the sample is calculated with Lis Hartel sign (Haar feature, moment characteristics).
Specifically, in step s 102, being calculated using Lis Hartel sign as follows:
Firstly, calculating the full figure pixel integral image of sample, formula S (i, j)=SUM0<ii<i,0<jj<j(ii,jj), that is, every point
Value be equal to abscissa and all smaller than current point the sum of all the points of ordinate.For the integrated value of rectangle any in figure, all may be used
To be obtained with the integral calculation of four angle points:
Srectangle=Sright-bottom-Sleft-bottom-Sright-top+Sleft-top；
Then, X " is calculated according to integrogram1iHarr feature, Harr feature has five category features, it is specific calculate it is as follows:
S1=Stop-Sbottom；
S2=Sleft-Sright；
S3=Sleft+Sright-Smiddle；
S4=Stop+Sbottom-Smiddle；
S5=Sright-bottom+Sleft-top-Sleft-bottom-Sright-top。
The note Lis Hartel that wherein every sample obtains sign is X " '1i, then obtain Lis Hartel collection { X " '1iIt is exactly entire
The fisrt feature collection of sample set.
Certainly, those skilled in the art, which will be understood that, above-mentioned extracts fisrt feature on the first training set in step s 102
The mode of collection is only for example, other modes of extraction second feature collection V that are existing or being likely to occur from now on, such as HOG feature
Or LBP feature etc. should also be included within the application protection scope, and include by reference herein as being applicable to the application
In this.
Then, in step s 102, the fisrt feature collection { X " ' is directed to using Ah's Da Busite algorithm1i, in conjunction with corresponding
Sample positive and negative label { Y1iBe trained, to obtain Ah's Da Busite model, steps are as follows:
Remember (xi,yi), wherein xiFor fisrt feature collection { X " '1iIn i-th of samples pictures feature vector, yiFor sample
Positive and negative label { Y1iIn the positive and negative label of i-th of samples pictures.
Firstly, initialization weight:
Work as yiWhen=0, weight W1,i=1/2M, works as yiWhen=1, weight W1,i=1/2L, wherein M is the total number of negative sample,
N is the total number of positive sample, and M+N=k1, K1 are the sum of sample.
Then, circulation executes step (a)~step S (f) T times, and (wherein T is artificial setting value, and T value can be 1~10
It is secondary, such as 2,5,8 times):
(a) normalization of weight: wt,i=wt,i/sumJ=1 ..., n(wt,j)；
(b) to each feature j, training generates classifier hj, and calculate error rate ej:
(c)ej=sumI=1 ..., k1(wi*|hj(xi)-yi|), wherein thetajTo manually set threshold value, according to training requirement
It is specifically set, is typically set to 0.5, PjThe sample labeling value that is positive or negative sample mark value, for example, 1 or -1.
(d) minimal error rate e is chosentCorresponding classifier ht；
(e) work as hj(xi)=yiOr wt+1,i=wt,i, then weight w is updatedt+1,i=wt,i*et/(1-et), in the case of other, then
Repetitive cycling step (a)~(d).
(f) finally, available classifier H (X):
Finally, H (X) result of generation is final judge mark, and wherein five features of Lis Hartel sign need all full
Sufficient condition, could generate final judge mark obtains Ah Da Busite model as a result, training to be formed.
Then, in step s 103, the normal image with the Logo is collected, is examined using the Ah Da Busite model
Logo described in normal image is surveyed, to obtain the candidate region with the Logo, collects normal image as the second training set.
Figure 10 show according in the another preferred embodiment of the application realize using Ah Da Busite algorithm combine support to
The method flow diagram that amount machine algorithm is trained Logo.It is specifically included in step S01 in preferred implement in conjunction with Figure 10
Step S101 ", step S102 ", step S103 ", step S104 ", step S105 " and step S106 ", wherein in step S105 "
In, the collected sample is pre-processed, the pretreatment includes carrying out gray processing processing and/or image to the sample
Size registration process.It is handled using gray processing and/or image size registration process pre-processes the sample after capable of making
Continuous detection process calculating is more easy, to improve detection processing speed.Here, step S101 ", step S102 ", step
S103 " and step S104 " and step S101, step S102, step S103 and step S104 corresponding contents in Fig. 8 are identical or basic
Identical, step S105 ' corresponding contents are identical or essentially identical in step S105 " and Fig. 9, for simplicity, thus it is no longer superfluous herein
It states, and is incorporated herein by reference.
Figure 11 is shown according to the pretreated side of candidate region progress realized in one preferred embodiment of the application to acquisition
Method flow chart.In conjunction with Figure 11, step S106 " described in Figure 10 includes step S601, step S602 and step S603.Wherein, in step
In rapid S601, the candidate region is cut into candidate image；In step s 601, whether phase is had according to the candidate image
The Logo answered carries out positive negative flag；In step s 601, image size registration process is carried out to the candidate image.Step
S106 ", which pre-processes the candidate region, can make the calculating of subsequent detection process more easy, to improve detection processing
Speed.
Specifically, in step s 601, these candidate regions being all cut into and are carried out as individual candidate image
Classification, manual type or automated manner can be used in mode classification, in step S602, does the positive and negative of positive negative sample after sorting
Then label, i.e., 0 or 1 label in step S603, carry out image size registration process to the candidate image, zoom to system
One is sized for third training set.
Then, in conjunction with Fig. 8, in step S14, second feature collection is extracted on third training set, wherein second feature collection
It can choose from fisrt feature using the different LBP or HOG feature of Lis Hartel sign extraction, can be further improved the accurate of detection
Property.
Figure 13 is shown according to the process signal for utilizing LBP feature calculation second feature in one preferred embodiment of the application
Figure.As shown in figure 13, the calculating process of LBP feature is as follows: certain particular neighborhood of image, such as the region 3x3 5x5 are extracted,
Take its center pixel value be c, then by the value of the pixel of all the points in its neighborhood compared with c size, be then denoted as 1 if more than c,
It is then denoted as 0 less than c, a string 0,1 obtained strings are LBP feature.For example, being characterized in calculating LBP by the field 3x3 all in image
It is composed, is computed as the pixel map of Figure 13 (a) and obtains the pixel Distribution value as shown in Figure 13 (b), the pixel at center
Value be 83, the pixel in other neighborhoods is compared with the value of the pixel at its center, shown in comparison result such as Figure 13 (c), general
The result of acquisition is 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) gradient for calculating each pixel of image, obtains size and direction；
(ii) block of cells is divided an image into, such as the block of cells of 6x6, the histogram of gradients in each block of cells is counted, obtains
To description of each block of cells；
(iii) as soon as 3x3 block of cells is formed a bulk, description of all block of cells of each bulk is together in series
Description of the bulk has been obtained, description of bulks all in image is together in series and has just obtained the HOG feature of the image,
The feature vector finally obtained can be used to train SVM.
Certainly, those skilled in the art will be understood that above-mentioned in step S14, extract second feature on third training set
The mode of collection V is only for example, other modes of extraction second feature collection V that are existing or being likely to occur from now on, such as Lis Hartel
Sign is such as applicable to the application, should also be included within the application protection scope, and be incorporated herein by reference.
Then, the 4th unit 104 is using second feature collection { Vi } training, to obtain supporting vector machine model (SVM model
Library).Utilize second feature training SVM model process, comprising:
(1) LaGrange parameter is first calculated, (x is seti,yi) it is training data, wherein xiIt is that sample i obtains feature vector
V, yi=0 (negative sample) or yi=1 (positive sample):
Work as αi>=0, i=1 ..., n, andWhen,
(2) weight w and b are being calculated, weight w and b are two parameters in SVM model.
(3) final judge mark y is obtained:
Y=w*x+b.
Then, by being trained to various Logo, the model library of numerous Logo can be obtained, to the model of numerous Logo
Library is managed collectively, and is sorted out by creating classification information label to the model library of Logo.To various in step S01
Logo is trained, and to obtain the model library of numerous Logo, is managed collectively to the model library of numerous Logo, by creating class
Other information labels sort out the model library of Logo.Wherein, the classification information can according to Logo represent businessman, produce
The contents such as product, service type are classified, such as the Logo of certain movement brand can have the classes such as clothes, trousers, shoes, ball
Other information labels can identify it with regard to the content of image to be detected then when the subsequent corresponding Logo detection to image to be detected progress
One of relevant classification information, such as several classifications of image to be detected are clothes, then in subsequent detection, select class label
Information is the Logo model library of clothes.To solve the problems, such as that detection difficulty caused by Logo is many kinds of is big, inspection is reduced
Difficulty is surveyed, detection efficiency is improved.
For being carried out at gray processing to image to be detected of the acquisition before being detected to described image to be detected
Reason and/or image size registration process.
Then, in step S02, image to be detected is obtained, and select according to the related category information of described image to be detected
, can be according to the related category information of Logo when selecting the model library of corresponding Logo, the model library for being each Logo increases
Several different classification information labels can be first according to image to be detected when needing to detect image to be detected
The relevant information for the Logo for needing to detect determines the classification information label for needing to detect, and calls further according to classification information label all
The model library of all Logo with category information.
Then, in step S03, using the Ah Da Busite model in the model library of corresponding Logo to it is described to
Detection image is detected, if testing result be it is consistent if extract candidate region, if testing result be it is inconsistent, illustrate to be checked
The Logo not responded in altimetric image is then directly returned.
When testing result be it is consistent when, then in step S04, using the support in the model library of corresponding Logo to
Amount machine model detects the candidate region, to obtain corresponding testing result.
Carrying out detection process to image to be detected using the model library of corresponding Logo can be the process of a circulation, benefit
If when being detected with the model library of a corresponding Logo to image to be detected testing result be it is inconsistent, using next corresponding
The model library of Logo image to be detected is detected, until the testing result obtained is consistent.
Figure 12, which is shown, utilizes Ah Da Busite algorithm combination supporting vector according to realization in one preferred embodiment of the application
The method flow diagram that machine algorithm is trained Logo；As shown in figure 12, the method includes the steps S01 ', step S02 ', step
Rapid S03 ', step S04 ' and step S05 '.Wherein, step S05 ' is between step S01 ' and step S02 ', to the acquisition
Image to be detected carries out gray processing processing and/or image size registration process, so that the calculating of subsequent detection process is more easy, from
And improve detection processing speed.Here, the method includes the steps S01 ', step S02 ', step S03 ' and step S04 ' and Fig. 7
Middle step S01, step S02, step S03 and step S04 corresponding contents are identical or essentially identical, for simplicity, therefore herein not
It repeats, and is incorporated herein by reference again.
In the particular embodiment, the herein described method for Logo detection, collects numerous samples first, per the same
Include the negative sample of the positive sample and image without Logo with the image of Logo in this, utilizes Ah's Da Busite algorithm knot
It closes algorithm of support vector machine and is trained the model library for obtaining various Logo to sample, the model library of each Logo includes should
The Ah Da Busite model and supporting vector machine model of Logo.When needing to detect image to be detected, first according to institute
The related category information for stating image to be detected selects the model library of corresponding Logo, then utilizes the model of corresponding Logo
Ah Da Busite model in library detects described image to be detected, extracts candidate region if being unanimously if testing result,
The supporting vector machine model in the model library of corresponding Logo is recycled to detect the candidate region, to obtain phase
Answer testing result.The model library of all Logo is successively detected, when the model library testing result of a Logo is inconsistent, then
Select the model library of next Logo carry out detection until testing result be it is consistent, if all testing results be it is inconsistent, not
The Logo that the image to be detected has related category is found, to make the detection process of Logo quickly and effectively complete, in addition, adopting
The detection of Logo can be made more accurate with Ah Da Busite model and supporting vector machine model.
In conclusion being calculated using Ah Da Busite algorithm combination supporting vector machine for Logo detection described herein
Method is trained Logo, and acquisition includes the model library of the Logo of Ah Da Busite model and supporting vector machine model, and according to
The related category information of image to be detected selects the model library of corresponding Logo, utilizes the A Da in the model library of corresponding Logo
Buss spy model and supporting vector machine model are to image to be detected.
Further, after the model library for forming several Logo, the model library of Logo is managed, according to the phase of Logo
Closing classification information can basis before detecting to image to be detected for the model library increase classification information label of every Logo
The related category information of image to be detected chooses the corresponding all Logo model libraries of respective classes information labels.The mould of all Logo
Type library is successively detected, when the model library testing result of a Logo be it is inconsistent when, then select the model library of next Logo into
Row detection until testing result be it is consistent, if all testing results be it is inconsistent, do not find the image to be detected have phase
Close the Logo of classification.
Obviously, those skilled in the art can carry out various modification and variations without departing from the essence of the application to the application
Mind and range.In this way, if these modifications and variations of the application belong to the range of the claim of this application and its equivalent technologies
Within, then the application is also intended to include these modifications and variations.
It should be noted that the application can be carried out in the assembly of software and/or software and hardware, for example, can adopt
With specific integrated circuit (ASIC), general purpose computer or any other realized similar to hardware device.In one embodiment
In, the software program of the application can be executed to implement the above steps or functions by processor.Similarly, the application
Software program (including relevant data structure) can be stored in computer readable recording medium, for example, RAM memory,
Magnetic or optical driver or floppy disc and similar devices.In addition, hardware can be used to realize in some steps or function of the application, example
Such as, as the circuit cooperated with processor thereby executing each step or function.
In addition, a part of the application can be applied to computer program product, such as computer program instructions, when its quilt
When computer executes, by the operation of the computer, it can call or provide according to the present processes and/or technical solution.
And the program instruction of the present processes is called, it is possibly stored in fixed or moveable recording medium, and/or pass through
Broadcast or the data flow in other signal-bearing mediums and transmitted, and/or be stored according to described program instruction operation
In the working storage of computer equipment.Here, including a device according to one embodiment of the application, which includes using
Memory in storage computer program instructions and processor for executing program instructions, wherein when the computer program refers to
When enabling by processor execution, method and/or skill of the device operation based on aforementioned multiple embodiments according to the application are triggered
Art scheme.
It is obvious to a person skilled in the art that the application is not limited to the details of above-mentioned exemplary embodiment, Er Qie
In the case where without departing substantially from spirit herein or essential characteristic, the application can be realized in other specific forms.Therefore, no matter
From the point of view of which point, the present embodiments are to be considered as illustrative and not restrictive, and scope of the present application is by appended power
Benefit requires rather than above description limits, it is intended that all by what is fallen within the meaning and scope of the equivalent elements of the claims
Variation is included in the application.Any reference signs in the claims should not be construed as limiting the involved claims.This
Outside, it is clear that one word of " comprising " does not exclude other units or steps, and odd number is not excluded for plural number.That states in device claim is multiple
Unit or device can also be implemented through software or hardware by a unit or device.The first, the second equal words are used to table
Show title, and does not indicate any particular order.
Claims (18)
1. a kind of method for Logo detection, wherein the described method includes:
Logo is trained using Ah's Da Busite algorithm combination supporting vector machine algorithm, to obtain the model of the Logo
Library, the model library of each Logo include the Ah Da Busite model and supporting vector machine model of the Logo；
Image to be detected is obtained, and selects the model library of corresponding Logo according to the related category information of described image to be detected；
Described image to be detected is detected using the Ah Da Busite model in the model library of corresponding Logo, if inspection
Surveying result is unanimously then to extract candidate region；And
The candidate region is detected using the supporting vector machine model in the model library of corresponding Logo, to obtain
Corresponding testing result；
Wherein, Logo is trained using Ah's Da Busite algorithm combination supporting vector machine algorithm and includes:
Collect sample, the sample include several single images with the Logo positive sample and it is several do not have the Logo
Image negative sample；
The fisrt feature collection of the sample is extracted, and is trained using Ah's Da Busite algorithm for the fisrt feature collection,
To obtain Ah's Da Busite model；
The normal image with the Logo is collected, using Logo described in the Ah Da Busite model inspection normal image,
To obtain the candidate region with the Logo；And
Extract the second feature collection of the candidate region, and using the algorithm of support vector machine for the second feature collection into
Row training, to obtain supporting vector machine model.
2. according to the method described in claim 1, wherein, using Ah's Da Busite algorithm combination supporting vector machine algorithm to Logo
It is trained further include:
Before the fisrt feature collection for extracting the sample, the collected sample is pre-processed, the pretreatment includes
Gray processing processing and/or image size registration process are carried out to the sample.
3. according to the method described in claim 1, wherein, being levied using Lis Hartel and calculating the fisrt feature collection for extracting the sample.
4. according to the method described in claim 1, wherein, using Ah's Da Busite algorithm combination supporting vector machine algorithm to Logo
It is trained, further includes:
Before the second feature collection for extracting the candidate region, the candidate region of the acquisition is pre-processed, the pre- place
Reason includes:
The candidate region is cut into candidate image；
According to the candidate image whether there is corresponding Logo to carry out positive negative flag；And
Image size registration process is carried out to the candidate image.
5. according to the method described in claim 1, wherein, extracting the candidate region using LBP feature or HOG feature calculation
Second feature collection.
6. according to the method described in claim 1, wherein, the method also includes:
Before described image to be detected is detected, gray processing processing and/or image size are carried out to described image to be detected
Registration process.
7. according to the method described in claim 1, wherein, the model library of each Logo all has several classification information marks
Label.
8. according to the method described in claim 7, wherein, being selected according to the related category information of described image to be detected corresponding
The model library of Logo includes:
Select to have the model of the related category institute of described image to be detected all Logo of classification information label accordingly
Library.
9. method according to any one of claim 1 to 8, wherein include: after returning to corresponding testing result
If testing result be it is consistent, stopping continue to test；
If testing result be it is inconsistent, the model library for continuing with the corresponding Logo examines described image to be detected
It surveys.
10. a kind of equipment for Logo detection, wherein the equipment includes:
First device, for being trained using Ah's Da Busite algorithm combination supporting vector machine algorithm to Logo, to obtain
The model library of Logo is stated, the model library of each Logo includes the Ah Da Busite model and support vector machines mould of the Logo
Type；
Second device selects accordingly for obtaining image to be detected, and according to the related category information of described image to be detected
The model library of Logo；
3rd device, the Ah Da Busite model in model library for utilizing corresponding Logo is to described image to be detected
It is detected, extracts candidate region if being unanimously if testing result；
4th device, for using corresponding Logo model library in supporting vector machine model to the candidate region into
Row detection, to obtain corresponding testing result；
Wherein, the first device includes:
First unit, for collecting sample, if the sample include several single images with the Logo positive sample and
Do the negative sample of the image of no Logo；
Second unit, for extracting the fisrt feature collection of the sample, and it is special for described first using Ah Da Busite algorithm
Collection is trained, to obtain Ah's Da Busite model；
Third unit is commonly schemed for collecting the normal image with the Logo using the Ah Da Busite model inspection
The Logo as described in, to obtain the candidate region with the Logo；And
Unit the 4th, for extracting the second feature collection of the candidate region, using the algorithm of support vector machine for described
Second feature collection is trained, to obtain supporting vector machine model.
11. equipment according to claim 10, wherein the first device further include:
Unit the 5th, for being pre-processed to the collected sample, institute before the fisrt feature collection for extracting the sample
Stating pretreatment includes carrying out gray processing processing and/or image size registration process to the sample.
12. equipment according to claim 10, wherein the second unit is calculated using Lis Hartel sign and extracts the sample
Fisrt feature collection.
13. equipment according to claim 10, wherein the first device further includes Unit the 6th, for extracting institute
Before the second feature collection for stating candidate region, the candidate region of the acquisition is pre-processed, Unit the 6th includes:
First subelement, for the candidate region to be cut into candidate image；
Second subelement, for according to the candidate image whether there is corresponding Logo to carry out positive negative flag；And
Third subelement carries out image size registration process to the candidate image.
14. equipment according to claim 10, wherein Unit the 4th is using described in LBP feature or the extraction of HOG feature calculation
The second feature collection of candidate region.
15. equipment according to claim 10, wherein the equipment further include:
5th device, for carrying out ash to image to be detected of the acquisition before detecting to described image to be detected
Degreeization processing and/or image size registration process.
16. equipment according to claim 10, wherein the model library of each Logo all has several classification information marks
Label.
17. equipment according to claim 16, wherein the second device is according to the related category of described image to be detected
Information selects the model library of corresponding Logo to include:
Select to have the model of the related category institute of described image to be detected all Logo of classification information label accordingly
Library.
18. equipment described in any one of 0 to 17 according to claim 1, wherein the equipment return corresponding testing result it
Afterwards, further includes:
If testing result be it is consistent, stopping continue to test；
If testing result be it is inconsistent, the model library for continuing with other corresponding Logo examines described image to be detected
It surveys.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410855921.6A CN105809181B (en) | 2014-12-31 | 2014-12-31 | Method and apparatus for Logo detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410855921.6A CN105809181B (en) | 2014-12-31 | 2014-12-31 | Method and apparatus for Logo detection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105809181A CN105809181A (en) | 2016-07-27 |
| CN105809181B true CN105809181B (en) | 2019-06-11 |
Family
ID=56465081
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410855921.6A Active CN105809181B (en) | 2014-12-31 | 2014-12-31 | Method and apparatus for Logo detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105809181B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108564631B (en) * | 2018-04-03 | 2021-07-09 | 上海理工大学 | Method and device for detecting light guide chromatic aberration of car lamp and computer readable storage medium |
| CN109800698B (en) * | 2019-01-11 | 2021-02-19 | 北京邮电大学 | Icon detection method based on deep learning, icon detection system and storage medium |
| JP2022091270A (en) * | 2020-12-09 | 2022-06-21 | ブラザー工業株式会社 | Method, system, and computer program |
| CN112989098B (en) * | 2021-05-08 | 2021-08-31 | 北京智源人工智能研究院 | Automatic retrieval method and device for image infringement entity and electronic equipment |
| CN113808116A (en) * | 2021-09-24 | 2021-12-17 | 无锡精质视觉科技有限公司 | Intelligent detection method and system based on image recognition and product detection system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102737255A (en) * | 2011-03-30 | 2012-10-17 | 索尼公司 | Target detection device and method |
| CN103077407A (en) * | 2013-01-21 | 2013-05-01 | 信帧电子技术（北京）有限公司 | Car logo positioning and recognition method and car logo positioning and recognition system |
| CN103646454A (en) * | 2013-12-24 | 2014-03-19 | 深圳市捷顺科技实业股份有限公司 | Parking lot management system and method |
| CN104182728A (en) * | 2014-07-26 | 2014-12-03 | 佳都新太科技股份有限公司 | Vehicle logo automatic location and recognition method based on pattern recognition |
-
2014
- 2014-12-31 CN CN201410855921.6A patent/CN105809181B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102737255A (en) * | 2011-03-30 | 2012-10-17 | 索尼公司 | Target detection device and method |
| CN103077407A (en) * | 2013-01-21 | 2013-05-01 | 信帧电子技术（北京）有限公司 | Car logo positioning and recognition method and car logo positioning and recognition system |
| CN103646454A (en) * | 2013-12-24 | 2014-03-19 | 深圳市捷顺科技实业股份有限公司 | Parking lot management system and method |
| CN104182728A (en) * | 2014-07-26 | 2014-12-03 | 佳都新太科技股份有限公司 | Vehicle logo automatic location and recognition method based on pattern recognition |
Non-Patent Citations (2)
| Title |
|---|
| "基于AdaBoost和SVM的图像检索研究";杨兴彤;《中国优秀硕士学位论文全文数据库》;20140215;全文 |
| "自然场景下交通标志的自动识别算法";何耀平 等;《微计算机信息》;20101231;图2 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105809181A (en) | 2016-07-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105809181B (en) | Method and apparatus for Logo detection | |
| Galleguillos et al. | Weakly supervised object localization with stable segmentations | |
| Lin et al. | Multiple instance ffeature for robust part-based object detection | |
| US9317781B2 (en) | Multiple cluster instance learning for image classification | |
| CN102414680B (en) | Utilize the semantic event detection of cross-domain knowledge | |
| CN104680144B (en) | Based on the lip reading recognition methods and device for projecting very fast learning machine | |
| CN103733224B (en) | Feature extraction device, feature extraction program, and image processing device | |
| CN107808126A (en) | Vehicle retrieval method and device | |
| CN110163033A (en) | Positive sample acquisition methods, pedestrian detection model generating method and pedestrian detection method | |
| CN105574550A (en) | Vehicle identification method and device | |
| CN105574063A (en) | Image retrieval method based on visual saliency | |
| CN102054178A (en) | Chinese painting image identifying method based on local semantic concept | |
| CN109886330A (en) | Method for text detection, device, computer readable storage medium and computer equipment | |
| CN109886295A (en) | A kind of butterfly recognition methods neural network based and relevant device | |
| CN109033944A (en) | A kind of all-sky aurora image classification and crucial partial structurtes localization method and system | |
| WO2012156774A1 (en) | Method and apparatus for detecting visual words which are representative of a specific image category | |
| CN102385592A (en) | Image concept detection method and device | |
| Zheng et al. | Fine-grained image classification based on the combination of artificial features and deep convolutional activation features | |
| Zhang et al. | An automatic detector for fungal spores in microscopic images based on deep learning | |
| Saitoh et al. | Feature Points based Fish Image Recognition. | |
| CN109886206A (en) | Three-dimensional object identification method and equipment | |
| Naik et al. | Flower Phenotype Recognition and Analysis using YoloV5 Models. | |
| CN105160333B (en) | A kind of model recognizing method and identification device | |
| BRPI0111017B1 (en) | METHOD AND APPARATUS FOR FINDING OBJECT IN AN IMAGE | |
| CN114399731A (en) | Target positioning method under single-coarse-point supervision |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |