CN102541954A - Method and system for searching trademarks - Google Patents
Method and system for searching trademarks Download PDFInfo
- Publication number
- CN102541954A CN102541954A CN2010106228051A CN201010622805A CN102541954A CN 102541954 A CN102541954 A CN 102541954A CN 2010106228051 A CN2010106228051 A CN 2010106228051A CN 201010622805 A CN201010622805 A CN 201010622805A CN 102541954 A CN102541954 A CN 102541954A
- Authority
- CN
- China
- Prior art keywords
- image
- trade mark
- point
- trademark
- retrieved
- 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.)
- Granted
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of image processing, in particular to a method and a system for searching trademarks by utilizing combined stable feature points. The method comprises the steps of: firstly, respectively performing combination extraction of stable feature points on to-be-searched images and trademark images, then performing feature description on the stable feature points extracted from the to-be-searched images and the original trademark images by using image information in adjacent regions to form feature vectors; and performing matching on all the feature points, screening out matched point sets by parameter voting, estimating trademark conversion models by using the matched point sets, and after obtaining the trademark conversion models, finding out the images on which the trademarks appear and the positions of the trademarks in the to-be-searched images. According to the method and the system provided by the invention, trademark searching can be performed aiming at complicated background images downloaded from the internet, and the method for searching the trademarks is not influenced by trademark scaling, illumination and rotation, can find out partially covered trademark patterns and has stronger adaptability for trademark deformation, color change and visual angle change as well.
Description
Technical field
The present invention relates to technical field of image processing, relate in particular to a kind of trade mark search method and system that utilizes the combinative stability unique point.
Background technology
Trade mark is not only the sign of an enterprise product, has also comprised many socio-cultural backgrounds.The management of enterprise trademark and protection are very important.Along with the professional rise of B2C, various shopping websites continue to bring out, and online merchandise news amount is huge, and the merchandise resources approach is numerous, also safeguards self brand to enterprise, and anti-infringement has brought new difficulty.If trade mark carried out the manual retrieval then need the labor manpower, the mass data of internet relatively, speed is very slow.And the retrieval of present internet, like Baidu, google etc. are main with key word still.Even more existing trade mark search method based on image, its application also mainly is in the trade mark storehouse, to search identical similar trade mark.Image in the trade mark storehouse is through standardized, only comprises trade mark, does not have other backgrounds.And trademark image is clear, rectify.And search in the situation of trade mark in the internet, trade mark reappears in image, tends to produce yardstick, the anglec of rotation, illumination, visual angle, shape variation.Internet images is more for the ease of transmitting general compression, and picture quality is not high yet.This is that the employed method in inquiry trade mark storehouse is insurmountable.
Present existing trade mark retrieval technique that is to say all from the trade mark global feature, all is trademark image in the image library to be retrieved, and every width of cloth image comprises and only comprise trademark image, does not have other backgrounds.Like one Chinese patent application 03152653.5 a kind of " trade mark search method " is provided, its technical scheme is that " a kind of trade mark search method comprises step: obtain trademark image; Filter out the noise in the pictorial trademark; Pre-service; Extract one group of characteristic the image after filtering with separability; Be stored in the characteristic that extracts in the characteristics dictionary; Dimensionality reduction; Utilize existing characteristic and characteristics dictionary to mate, calculate similarity and return one group of image the most similar as Query Result; Utilize relevant feedback to optimize result for retrieval.This method can be eliminated the heavy work of manual coding, reduces people's in the manual coding subjectivity influence, improves the efficient and the reliability of trade mark retrieval and authentication, shortens the authentication period of trade mark." the trade mark search method that occurs in recent years some periodicals basically all is this thinking.But this technical thought must be carried out feature extraction to whole trademark image, such as profile, distribution of color, texture, various moment characteristics etc.This just requires image to be checked must be complete and only comprise trademark image, can not be that trade mark appears on other backgrounds.Secondly, use global characteristics to carry out the trade mark retrieval, can't solve the problem that trade mark amplifies or dwindles, for the situation of blocking, the global characteristics error can be very big, thereby can't correctly retrieve.
Summary of the invention
The objective of the invention is to the defective to prior art, a kind of trade mark search method and system that utilizes the combinative stability unique point is provided, thereby can realize the accurate retrieval of the trademark image of trademark image that background is complicated or partial occlusion.
Technical scheme of the present invention is following: a kind of trade mark search method comprises the steps:
(S1) treat retrieving images and original trademark image respectively and carry out invariant feature point combination extraction, find out the invariant feature point that not influenced by illumination, color, yardstick, rotation change;
(S2) to the invariant feature point that extracts in image to be retrieved and the original trademark image, utilize the image information in its neighborhood to carry out feature description, form proper vector;
(S3) utilize range formula to calculate the distance between each unique point in unique point and the original trademark image in the image to be retrieved respectively, carry out Feature Points Matching;
(S4) through setting up brand-switching model,, can estimate its position if exist to confirm whether trade mark exists;
(S5) according to the transformation parameter of brand-switching model, calculate the coordinate transforming on four summits of trade mark boundary rectangle, and be linked in sequence, in image to be retrieved, find out position and scope that trade mark occurs, the sign trade mark.
Further; Aforesaid trade mark search method in step (S1) before, comprises that also treating retrieving images carries out pretreated step; Concrete grammar is: the change color situation in image to be retrieved, occurs for adapting to trade mark, calculate the error image of solid white image and original image; For adapting to the too small situation of trade mark in the image to be retrieved, treat retrieving images and carry out amplifieroperation.
Further, aforesaid trade mark search method, in the step (S1), the concrete grammar that the combination of invariant feature point is extracted is following:
(S11) utilize sampling and this pyramid of Gaussian convolution structural map image height, it is made up of a plurality of frequency ranges, and the successive bands yardstick differs 50%, utilizes Gaussian convolution to construct a plurality of sublayers in each frequency range;
(S12) using various features to detect operator to each tomographic image handles;
(S13) in each frequency range; To each pixel in each sublayer, compare the result value of the feature detection operator in the neighborhood on the metric space, in its neighborhood if the end value on this pixel is a maximum value or minimum value; Just with it as the candidate feature point; Write down the frequency range that it occurs, sublayer sequence number, the coordinate information in the image;
(S14) remove the point that repeats in the candidate feature point, remove near the point of weak contrast and edge in the candidate feature point then, just obtained the invariant feature point.
Further, the method that aforesaid invariant feature point combination is extracted, in the step (S11), the graphical rule of described image gaussian pyramid reduces by frequency range from bottom to top one by one, and the yardstick of a last frequency range image is 50% of its next frequency range graphical rule; In each frequency range of image pyramid, utilize Gaussian convolution to generate 4~6 sub-layer; The pyramid bottom is original input picture, the Gaussian function that uses in each frequency range, and its standard deviation computing formula is following:
σ
in=σ
02
k
K=1+2 * in/IN wherein, σ
0=1.6, in is a sublayer sequence number in the frequency range, and IN is the sublayer sum in the frequency range;
If the image in the gaussian pyramid is GP
Ot, in, then
Wherein ot is the frequency range sequence number, and in is the sequence number of sublayer in the frequency range, and I is an original image, G (σ
In) be Gaussian function, on behalf of yardstick, symbol ↓ 2 dwindle 50% sample calculation.
Further, the method that aforesaid invariant feature point combination is extracted, in the step (S12), various features detects operator and comprises DoG, LoG, Hessian matrix determinant.
Further, aforesaid trade mark search method is in the step (S2); Described feature description is a histogram of gradients; The horizontal ordinate of histogram of gradients is the gradient angle, and ordinate is the weighted sum of gradient magnitude, and weighting coefficient is for being the center with the unique point, being the Gaussian function numerical value of standard deviation with the certain radius; Radius [4,6].In addition, feature description also can adopt the color intensity histogram, window Fourier transform characteristic, Image Description Methods such as gray level co-occurrence matrixes.
Further, aforesaid trade mark search method in the step (S2), when feature description is histogram of gradients, is carried out amplitude normalization to [0,1] with histogram of gradients, removes the bright dark influence of illumination; Confirm the principal direction of unique point according to the main peak of the Gradient distribution around the unique point, obtain the principal direction differential seat angle of unique point in image to be retrieved and the original trademark image, confirm the trade mark anglec of rotation; So-called principal direction angle, the i.e. corresponding angle of amplitude peak in histogram of gradients.
Further, aforesaid trade mark search method, in the step (S3), concrete characteristic point matching method is divided into two kinds of forms of forward coupling and reverse coupling.
Further, in the aforesaid characteristic point matching method, forward coupling step is following:
(a) with each the unique point f in the image to be retrieved
iAll unique point G={g in ∈ F and the original trademark image
j, j=1,2 ..., m} compares, and calculates bee-line d
0With inferior short distance d
1, establishing respectively, characteristic of correspondence point is g
pAnd g
q
(b) if d
0<(a * d
1), a is a constant, then thinks f
iWith bee-line characteristic of correspondence point g
pMate successfully.
Further, in the aforesaid characteristic point matching method, it is following oppositely to mate step:
1. with each the unique point g in the original trademark image
iAll unique point F={f among ∈ G and the figure to be retrieved
j, j=1,2 ..., n} compares, and calculates bee-line d
0With inferior short distance d
1, establish character pair point f respectively
rAnd f
s
If 2. d
0<(a * d
1), a is a constant, then thinks g
iWith bee-line characteristic of correspondence point f
rMate successfully;
If 3. d
0>=(a * d
1), then continue relatively d
1With the next one than short distance d
2Relation, if d
1<(a * d
2), then think and find match point, otherwise continue to seek successively, until d
K-1And d
kWherein the value of k is any positive integer greater than 1, and the adjusting of its value can be controlled final recall ratio; Span is [3,6] preferably.
Further, in the aforesaid characteristic point matching method, the span of constant a is [0.4,0.7]; Preferred a=0.5 during the forward coupling, preferred a=0.6 when oppositely mating.
Further, aforesaid trade mark search method, in the step (S3), described range formula comprises L
1, L
2, χ
2, KL divergence, cosine range formula etc.
Further, aforesaid trade mark search method is in the step (S3), when adopting χ
2During range formula, concrete computing method are following:
H wherein
1Be a certain characteristic neighborhood of a point histogram of gradients in the original trademark image, H
2Be a certain characteristic neighborhood of a point histogram of gradients in the image to be retrieved, i is the index of histogram array, and N is the length of histogram data, and θ is the anglec of rotation, and d θ is the angle quantized interval, mod
NBe remainder divided by N.
Further, aforesaid trade mark search method, in the step (S4), according to scale parameter and anglec of rotation conversion trademark image, horizontal stroke, the ordinate deviation of match point after the conversion of calculating trade mark; Find out from image to be retrieved according to deviation operation parameter voting method again and possibly be the feature point set of trade mark; The last method of RANSAC of in each feature point set, using is confirmed the trademark image transformation parameter; Described scale parameter is following:
sp=σ
02
layer
In the formula, layer=oct+in/IN, oct are the frequency range number that unique point occurs, and IN is the figure layer sum in the frequency range, and in is that the figure of unique point in a frequency range counts σ layer by layer
0Poor for image being carried out the level and smooth base standard of Gauss, get 1.6 usually.
Further, aforesaid trade mark search method, in the step (S4), the computing method of the horizontal stroke of match point, ordinate deviation are following after the trade mark conversion:
The upper left corner with the original trademark image is initial point, sets up the XY coordinate system; The upper left corner with image to be retrieved is initial point, sets up the UV coordinate system, is (x for each coordinate in the original trademark image; Y) unique point, its coordinate that transforms to the UV coordinate system is (u ', v '); With the model approximation of trade mark conversion is affined transformation, and the affined transformation formula is following:
And in image to be retrieved with the original trademark image in unique point (x, y) actual coordinate of point of coupling be (u, v), then grid deviation is:
Wherein s is an enlargement factor, and θ is the anglec of rotation
γ
11=s·cosθ γ
12=s·sinθ。
γ
21=-s·sinθ γ
22=s·cosθ
A kind of trade mark searching system comprises:
Image preprocess apparatus is used for trademark image is carried out pre-service, makes trademark image adapt to change color and the too small situation of yardstick;
Invariant feature point combination extraction element carries out invariant feature point combination extraction to pretreated image to be retrieved and original trademark image respectively, finds out the invariant feature point that not influenced by illumination, color, yardstick, rotation change;
The unique point characterization device is used for treating the invariant feature point that retrieving images and original trademark image extract and carries out feature description, calculates the histogram of gradients in the neighborhood of the position on each invariant feature point place layer respectively;
The Feature Points Matching device utilizes histogram of gradients and range formula to calculate the distance between each unique point in unique point and the original trademark image in the image to be retrieved respectively, carries out Feature Points Matching;
The transformation model estimation unit is used for to confirm whether trade mark exists, can estimating its position if exist through setting up brand-switching model;
The trademark device is used for the transformation parameter according to brand-switching model, calculates the coordinate transforming on four summits of trade mark boundary rectangle, and is linked in sequence, and in image to be retrieved, finds out position and scope that trade mark occurs.
Beneficial effect of the present invention is following: method and system provided by the present invention utilizes the local message in series of stable unique point and this unique point neighborhood that trade mark is described; Utilize the point set that matees to estimate the model of trade mark conversion; Can carry out the trade mark retrieval to the background complicated image of downloading on the internet, and be not limited only to the image in the trade mark storehouse after the standardization.Its search method does not receive the influence of trade mark convergent-divergent, illumination, rotation, and can find out the pictorial trademark of partial occlusion, for trade mark deformation, look change, visual angle change stronger adaptability is arranged also.
Description of drawings
Fig. 1 utilizes the trade mark search method process flow diagram of combinative stability unique point for the present invention;
The concrete grammar process flow diagram that Fig. 2 extracts for the combination of invariant feature point;
Fig. 3 utilizes the structural representation of the trade mark searching system of combinative stability unique point for the present invention.
Embodiment
Be elaborated below in conjunction with the Figure of description specific embodiments of the invention.
As shown in Figure 3, the trade mark searching system of utilizing the combinative stability unique point provided by the present invention comprises:
Image preprocess apparatus 1 is used for trademark image is carried out pre-service, makes trademark image adapt to change color and the too small situation of yardstick;
Invariant feature point combination extraction element 2 carries out invariant feature point combination extraction to pretreated trademark image and image to be retrieved respectively, finds out the invariant feature point that not influenced by illumination, color, yardstick, rotation change;
Unique point characterization device 3 is used for treating the invariant feature point that retrieving images and original trademark image extract and carries out feature description, calculates the histogram of gradients in the neighborhood of the position on each invariant feature point place layer respectively;
Feature Points Matching device 4 utilizes histogram of gradients and range formula to calculate the distance between each unique point in unique point and the original trademark image in the image to be retrieved respectively, carries out Feature Points Matching;
Transformation model estimation unit 5 is used for to confirm whether trade mark exists, can estimating its position if exist through setting up brand-switching model;
The trade mark search method of utilizing the combinative stability unique point that said system realized is as shown in Figure 1, comprises the steps:
Step S1 treats retrieving images and original trademark image respectively and carries out invariant feature point combination extraction, finds out the invariant feature point that not influenced by illumination, color, yardstick, rotation change.This step is accomplished by invariant feature point combination extraction element 2.
Before the combination of invariant feature point is extracted; Can carry out pre-service through 1 pair of trademark image of image preprocess apparatus earlier; Pre-service generally comprises: for adapting to change color appears in trade mark in image to be retrieved situation; Calculate the error image of solid white image and original image, the trademark image of contrast is two like this, can retrieve and the opposite trade mark of target trademark image color; For adapting to the too small situation of trade mark in the image to be retrieved, image is carried out amplifieroperation.
The process that the combination of invariant feature point is extracted is as shown in Figure 2, and step is following:
(S11) utilize sampling and this pyramid of Gaussian convolution structural map image height, it is made up of a plurality of frequency ranges, and the successive bands yardstick differs 50%, utilizes Gaussian convolution to construct a plurality of sublayers in each frequency range.
The graphical rule of described image gaussian pyramid reduces by frequency range from bottom to top one by one, and the yardstick of a last frequency range image is 50% of its next frequency range graphical rule; In each frequency range of image pyramid, utilize Gaussian convolution to generate 4~6 sub-layer; The pyramid bottom is original input picture, the Gaussian function that uses in each frequency range, and its standard deviation computing formula is following:
σ
in=σ
02
k
K=1+2 * in/IN wherein, σ
0=1.6, in is a sublayer sequence number in the frequency range, and IN is the sublayer sum in the frequency range;
If the image in the gaussian pyramid is GP
Ot, in, then
Wherein ot is the frequency range sequence number, and in is the sequence number of sublayer in the frequency range, and I is an original image, G (σ
In) be Gaussian function, on behalf of yardstick, symbol ↓ 2 dwindle 50% sample calculation.
(S12) use various features to detect operator to each tomographic image and handle, various features detects operator and comprises DoG, LoG; Hessian matrix determinant; These all are feature detection operators of the prior art, are example with DoG, and the adjacent sublayers image that is about in the same frequency range subtracts computing.
(S13) in each frequency range; To each pixel in each sublayer, compare the result value of the feature detection operator in the neighborhood on the metric space, in its neighborhood if the end value on this pixel is a maximum value or minimum value; Just with it as the candidate feature point; Write down the frequency range that it occurs, sublayer sequence number, the coordinate information in the image;
(S14) owing to used various features to detect operator, so have repetition in the candidate feature point, need to remove the point of repetition in the candidate feature point, remove near the point of weak contrast and edge in the candidate feature point then, just obtained the invariant feature point.
Step S2 to the invariant feature point that extracts in image to be retrieved and the original trademark image, utilizes the image information in its neighborhood to carry out feature description, forms proper vector.This step is accomplished by unique point characterization device 3.
Described feature description can adopt Image Description Methods such as histogram of gradients, color intensity histogram, window Fourier transform characteristic, gray level co-occurrence matrixes, and these describing methods are the known technology of this area.When adopting histogram of gradients, the horizontal ordinate of histogram of gradients is the gradient angle, and ordinate is the weighted sum of gradient magnitude, and weighting coefficient is for being the center with the unique point, being the Gaussian function numerical value of standard deviation with the certain radius, radius [4,6].Histogram of gradients is carried out amplitude normalization to [0,1], remove the bright dark influence of illumination; Confirm the principal direction of unique point according to the main peak of the Gradient distribution around the unique point, obtain the principal direction differential seat angle of unique point in image to be retrieved and the original trademark image, confirm the trade mark anglec of rotation; So-called principal direction angle, the i.e. corresponding angle of amplitude peak in histogram of gradients.
Step S 3, utilize range formula to calculate the distance between each unique point in unique point and the original trademark image in the image to be retrieved respectively, carry out Feature Points Matching.This step is accomplished by Feature Points Matching device 4.
Concrete characteristic point matching method is divided into two kinds of forms of forward coupling and reverse coupling.
The step of described forward coupling is following:
(a) with each the unique point f in the image to be retrieved
iAll unique point G={g in ∈ F and the original trademark image
j, j=1,2 ..., m} compares, and calculates bee-line d
0With inferior short distance d
1, establish characteristic of correspondence point g respectively
pAnd g
q
(b) if d
0<(a * d
1), a is a constant, then thinks f
iWith bee-line characteristic of correspondence point g
pMate successfully.
The step of described reverse coupling is following:
1. with each the unique point g in the original trademark image
iAll unique point F={f among ∈ G and the figure to be retrieved
j, j=1,2 ..., n} compares, and calculates bee-line d
0With inferior short distance d
1, establish character pair point f respectively
rAnd f
s
If 2. d
0<(a * d
1), a is a constant, then thinks g
iWith bee-line characteristic of correspondence point f
rMate successfully;
If 3. d
0>=(a * d
1), then continue relatively d
1With the next one than short distance d
2Relation, if d
1<(a * d
2), then think and find match point, otherwise continue to seek successively, until d
K-1And d
kWherein the value of k is any positive integer greater than 1, and the adjusting of its value can be controlled final recall ratio; Span is [3,6] preferably.
In above-mentioned two kinds of matching process, the span of constant a is [0.4,0.7]; Preferred a=0.5 during the forward coupling, preferred a=0.6 when oppositely mating.
Calculate that the employed range formula of distance between each unique point comprises L in unique point and the original trademark image in the image to be retrieved
1, L
2, χ
2, KL divergence, cosine range formula.The calculating of above-mentioned range formula is the known technology of this area.
With χ
2Range formula is an example, when adopting χ
2During range formula, concrete computing method are following:
Wherein, H
1Be a certain characteristic neighborhood of a point histogram of gradients in the original trademark image, H
2Be a certain characteristic neighborhood of a point histogram of gradients in the image to be retrieved, i is the index of histogram array, and N is the length of histogram data, and θ is the anglec of rotation, and d θ is the angle quantized interval, mod
NBe remainder divided by N.
Step S4 through setting up brand-switching model, to confirm whether trade mark exists, can estimate its position if exist.This step is accomplished by model estimation unit 5.
According to scale parameter and principal direction angular transformation trademark image, horizontal stroke, the ordinate deviation of match point after the conversion of calculating trade mark; Find out from image to be retrieved according to deviation operation parameter voting method (or marking area method for distilling) again and possibly be the feature point set of trade mark; The last method of RANSAC of in each feature point set, using is confirmed the trademark image transformation parameter.Theoretical according to metric space, scale parameter does
sp=σ
02
layer
Wherein, layer=oct+in/IN.Oct is the frequency range number that unique point occurs, and IN is the figure layer sum in the frequency range, and in is that the figure of unique point in a frequency range counts layer by layer.σ
0Poor for image being carried out the level and smooth base standard of Gauss, get 1.6 usually.
The upper left corner with the original trademark image is initial point, sets up the XY coordinate system; The upper left corner with image to be retrieved is initial point, sets up the UV coordinate system.Each coordinate in the original trademark image is that (its coordinate that transforms to the UV coordinate system is (u ', v ') for x, unique point y).Can be affined transformation or perspective model with the model approximation of trade mark conversion, the affined transformation formula be following:
And in image to be retrieved with the original trademark image in unique point (x, y) actual coordinate of point of coupling be (u, v).Grid deviation then
Wherein s is an enlargement factor, and θ is the anglec of rotation,
γ
11=s·cosθ γ
12=s·sinθ
γ
21=-s·sinθ γ
22=s·cosθ
With grid deviation du and dv is that coordinate axis is set up parameter space, with certain pixel wide parameter space is divided.This width can be got arbitrarily the positive integer greater than 0, but its adjusting can influence final recall ratio, and step is comprehensively chosen before and after needing to cooperate, and gets 10 pixels here.Calculate all points that mated between grid deviation, and in parameter space, vote.To after all calculating and voting, search ballot value is greater than the parameter space division unit of nv to all points, and obtains one group of some pair set.Wherein the numerical value of nv is chosen relevant with the transformation model that adopts.Supposing to ask for the required smallest point of transformation parameter is nmin, then nv>nmin to quantity.Here adopt affined transformation, so nmin=3 gets nv=4.In each the some pair set that obtains, using the RANSAC method to carry out transformation model at last estimates, asks for transformation parameter γ
11, γ
12, γ
21, γ
22, du and dv.Wherein, the RANSAC method is computing method well known in the art.
Set up brand-switching model here three effects are arranged: (one), utilize Space Consistency to judge whether trade mark exists, because the coupling of the point of the front existence of illustrative trade mark not; (2), relax the restriction of point-to-point coupling, to improve adaptability to the distortion trade mark; (3), transformation model capable of using calculates the particular location of trade mark.
(S5) sign trade mark.According to the transformation parameter of brand-switching model, calculate the coordinate transforming on four summits of trade mark boundary rectangle, and be linked in sequence, in image to be retrieved, find out position and scope that trade mark occurs, the sign trade mark.This step is accomplished by trademark device 6.
If the height of original trademark image is ScrH, wide is ScrW, is that initial point is set up the XY coordinate system with the upper left corner of original trademark image.Upper left corner coordinate is (0,0), and upper right corner coordinate is (ScrW, 0), lower right corner coordinate be (ScrW, ScrH), lower left corner coordinate be (0, ScrH).The upper left corner with image to be retrieved is that initial point is set up the UV coordinate system, then utilizes the model transferring formula, like the affined transformation formula of front, can obtain the coordinate of original trademark image four angular coordinate in figure to be retrieved respectively.4 points that the order link calculates in figure to be retrieved have promptly identified trade mark position.
Obviously, those skilled in the art can carry out various changes and modification to the present invention and not break away from the spirit and scope of the present invention.Like this, belong within the scope of claim of the present invention and equivalent technologies thereof if of the present invention these are revised with modification, then the present invention also is intended to comprise these changes and modification interior.
Claims (17)
1. a trade mark search method comprises the steps:
(S1) treat retrieving images and original trademark image respectively and carry out invariant feature point combination extraction, find out the invariant feature point that not influenced by illumination, color, yardstick, rotation change;
(S2) to the invariant feature point that extracts in image to be retrieved and the original trademark image, utilize the image information in its neighborhood to carry out feature description, form proper vector;
(S3) utilize range formula to calculate the distance between each unique point in unique point and the original trademark image in the image to be retrieved respectively, carry out Feature Points Matching;
(S4) through setting up brand-switching model,, can estimate its position if exist to confirm whether trade mark exists;
(S5) according to the transformation parameter of brand-switching model, calculate the coordinate transforming on four summits of trade mark boundary rectangle, and be linked in sequence, in image to be retrieved, find out position and scope that trade mark occurs, the sign trade mark.
2. trade mark search method as claimed in claim 1; It is characterized in that: before in step (S1); Comprise that also treating retrieving images carries out pretreated step; Concrete grammar is: the change color situation in image to be retrieved, occurs for adapting to trade mark, calculate the error image of solid white image and original image; For adapting to the too small situation of trade mark in the image to be retrieved, treat retrieving images and carry out amplifieroperation.
3. according to claim 1 or claim 2 trade mark search method, it is characterized in that: in the step (S1), the concrete grammar that the combination of invariant feature point is extracted is following:
(S11) utilize sampling and this pyramid of Gaussian convolution structural map image height, it is made up of a plurality of frequency ranges, and the successive bands yardstick differs 50%, utilizes Gaussian convolution to construct a plurality of sublayers in each frequency range;
(S12) using various features to detect operator to each tomographic image handles;
(S13) in each frequency range; To each pixel in each sublayer, compare the result value of the feature detection operator in the neighborhood on the metric space, in its neighborhood if the end value on this pixel is a maximum value or minimum value; Just with it as the candidate feature point; Write down the frequency range that it occurs, sublayer sequence number, the coordinate information in the image;
(S14) remove the point that repeats in the candidate feature point, remove near the point of weak contrast and edge in the candidate feature point then, just obtained the invariant feature point.
4. trade mark search method as claimed in claim 3 is characterized in that: in the step (S11), the graphical rule of described image gaussian pyramid reduces by frequency range from bottom to top one by one, and the yardstick of a last frequency range image is 50% of its next frequency range graphical rule; In each frequency range of image pyramid, utilize Gaussian convolution to generate 4~6 sub-layer; The pyramid bottom is original input picture, the Gaussian function that uses in each frequency range, and its standard deviation computing formula is following:
σ
in=σ
02
k
K=1+2 * in/IN wherein, σ
0=1.6, in is a sublayer sequence number in the frequency range, and IN is the sublayer sum in the frequency range;
If the image in the gaussian pyramid is GP
Ot, in, then
Wherein ot is the frequency range sequence number, and in is the sequence number of sublayer in the frequency range, and I is an original image, G (σ
In) be Gaussian function, on behalf of yardstick, symbol ↓ 2 dwindle 50% sample calculation.
5. trade mark search method as claimed in claim 3 is characterized in that: in the step (S12), various features detects operator and comprises DoG, LoG, Hessian matrix determinant.
6. according to claim 1 or claim 2 trade mark search method; It is characterized in that: in the step (S2), described feature description is a histogram of gradients, and the horizontal ordinate of histogram of gradients is the gradient angle; Ordinate is the weighted sum of gradient magnitude; Weighting coefficient is for being the center with the unique point, being the Gaussian function numerical value of standard deviation with the certain radius, radius [4,6].
7. according to claim 1 or claim 2 trade mark search method, it is characterized in that: in the step (S2), described feature description also can combined be used the color intensity histogram, window Fourier transform characteristic or gray level co-occurrence matrixes.
8. trade mark search method as claimed in claim 6 is characterized in that: in the step (S2), histogram of gradients is carried out amplitude normalization to [0,1], remove the bright dark influence of illumination; Confirm the principal direction of unique point according to the main peak of the Gradient distribution around the unique point, obtain the principal direction differential seat angle of unique point in image to be retrieved and the original trademark image, confirm the trade mark anglec of rotation; So-called principal direction angle, the i.e. corresponding angle of amplitude peak in histogram of gradients.
9. trade mark search method as claimed in claim 1 is characterized in that: in the step (S3), concrete characteristic point matching method is divided into two kinds of forms of forward coupling and reverse coupling.
10. trade mark search method as claimed in claim 9 is characterized in that: in the step (S3), the step of described forward coupling is following:
(a) with each the unique point f in the image to be retrieved
iAll unique point G={g in ∈ F and the original trademark image
j, j=1,2 ..., m} compares, and calculates bee-line d
0With inferior short distance d
1, establishing respectively, characteristic of correspondence point is g
pAnd g
q
(b) if d
0<(a * d
1), a is a constant, then thinks f
iWith bee-line characteristic of correspondence point g
pMate successfully.
11. trade mark search method as claimed in claim 9 is characterized in that: in the step (S3), the step of described reverse coupling is following:
1. with each the unique point g in the original trademark image
iAll unique point F={f among ∈ G and the figure to be retrieved
j, j=1,2 ..., n} compares, and calculates bee-line d
0With inferior short distance d
1, establish character pair point f respectively
rAnd f
s
If 2. d
0<(a * d
1), a is a constant, then thinks g
iWith bee-line characteristic of correspondence point f
rMate successfully;
If 3. d
0>=(a * d
1), then continue relatively d
1With the next one than short distance d
2Relation, if d
1<(a * d
2), then think and find match point, otherwise continue to seek successively, until d
K-1And d
kWherein the value of k is any positive integer greater than 1, and the adjusting of its value can be controlled final recall ratio; Span is [3,6] preferably.
12. like claim 10 or 11 described trade mark search methods, it is characterized in that: in the step (S3), in characteristic point matching method, the span of constant a is [0.4,0.7]; Preferred a=0.5 during the forward coupling, preferred a=0.6 when oppositely mating.
13. trade mark search method as claimed in claim 1 is characterized in that: in the step (S3), described range formula comprises L
1, L
2, χ
2, KL divergence, cosine range formula.
14. trade mark search method as claimed in claim 13 is characterized in that: in the step (S3), when adopting χ
2During range formula, concrete computing method are following:
H wherein
1Be a certain characteristic neighborhood of a point histogram of gradients in the original trademark image, H
2Be a certain characteristic neighborhood of a point histogram of gradients in the image to be retrieved, i is the index of histogram array, and N is the length of histogram data, and θ is the anglec of rotation, and d θ is the angle quantized interval, mod
NBe remainder divided by N.
15. trade mark search method according to claim 1 or claim 2 is characterized in that: in the step (S4), according to scale parameter and anglec of rotation conversion trademark image, horizontal stroke, the ordinate deviation of match point after the conversion of calculating trade mark; Find out from image to be retrieved according to deviation operation parameter voting method again and possibly be the feature point set of trade mark; The last method of RANSAC of in each feature point set, using is confirmed the trademark image transformation parameter; Described scale parameter is following:
sp=σ
02
layer
In the formula, layer=oct+in/IN, oct are the frequency range number that unique point occurs, and IN is the figure layer sum in the frequency range, and in is that the figure of unique point in a frequency range counts σ layer by layer
0Poor for image being carried out the level and smooth base standard of Gauss, get 1.6 usually.
16. trade mark search method as claimed in claim 15 is characterized in that: in the step (S4), the computing method of the horizontal stroke of match point, ordinate deviation are following after the trade mark conversion:
The upper left corner with the original trademark image is initial point, sets up the XY coordinate system; The upper left corner with image to be retrieved is initial point, sets up the UV coordinate system, is (x for each coordinate in the original trademark image; Y) unique point, its coordinate that transforms to the UV coordinate system is (u ', v '); With the model approximation of trade mark conversion is affined transformation, and the affined transformation formula is following:
And in image to be retrieved with the original trademark image in unique point (x, y) actual coordinate of point of coupling be (u, v), then grid deviation is:
Wherein, s is an enlargement factor, and θ is the anglec of rotation,
γ
11=s·cosθ γ
12=s·sinθ。
γ
21=-s·sinθ γ
22=s·cosθ
17. a trade mark searching system comprises:
Image preprocess apparatus (1) is used for trademark image is carried out pre-service, makes trademark image adapt to change color and the too small situation of yardstick;
Invariant feature point combination extraction element (2) carries out invariant feature point combination extraction to pretreated image to be retrieved and original trademark image respectively, finds out the invariant feature point that not influenced by illumination, color, yardstick, rotation change;
Unique point characterization device (3) is used for treating the invariant feature point that retrieving images and original trademark image extract and carries out feature description, calculates the histogram of gradients in the neighborhood of the position on each invariant feature point place layer respectively;
Feature Points Matching device (4) utilizes histogram of gradients and range formula to calculate the distance between each unique point in unique point and the original trademark image in the image to be retrieved respectively, carries out Feature Points Matching;
Transformation model estimation unit (5) is used for to confirm whether trade mark exists, can estimating its position if exist through setting up brand-switching model;
Trademark device (6) is used for the transformation parameter according to brand-switching model, calculates the coordinate transforming on four summits of trade mark boundary rectangle, and is linked in sequence, and in image to be retrieved, finds out position and scope that trade mark occurs.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201010622805.1A CN102541954B (en) | 2010-12-29 | 2010-12-29 | Method and system for searching trademarks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201010622805.1A CN102541954B (en) | 2010-12-29 | 2010-12-29 | Method and system for searching trademarks |
Publications (2)
Publication Number | Publication Date |
---|---|
CN102541954A true CN102541954A (en) | 2012-07-04 |
CN102541954B CN102541954B (en) | 2014-11-05 |
Family
ID=46348869
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201010622805.1A Active CN102541954B (en) | 2010-12-29 | 2010-12-29 | Method and system for searching trademarks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN102541954B (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104462382A (en) * | 2014-12-11 | 2015-03-25 | 北京中细软移动互联科技有限公司 | Trademark image inquiry method |
CN104732225A (en) * | 2013-12-24 | 2015-06-24 | 中国科学院深圳先进技术研究院 | Image rotation processing method |
CN104809142A (en) * | 2014-01-29 | 2015-07-29 | 北京瑞天科技有限公司 | Trademark inquiring system and method |
CN105096308A (en) * | 2014-05-13 | 2015-11-25 | 斯耐尔有限公司 | Logo detection by edge matching |
CN105989043A (en) * | 2015-02-04 | 2016-10-05 | 阿里巴巴集团控股有限公司 | Method and device for automatically acquiring trademark in commodity image and searching trademark |
CN106530194A (en) * | 2015-09-09 | 2017-03-22 | 阿里巴巴集团控股有限公司 | Method and apparatus for detecting pictures of suspected infringing products |
WO2017101224A1 (en) * | 2015-12-15 | 2017-06-22 | 徐庆 | Trademark retrieval method, apparatus and system, and computer storage medium |
CN108664945A (en) * | 2018-05-18 | 2018-10-16 | 徐庆 | Image text and form-pronunciation-meaning characteristic recognition method and device |
CN108763380A (en) * | 2018-05-18 | 2018-11-06 | 徐庆 | Brand recognition search method, device, computer equipment and storage medium |
CN108804499A (en) * | 2018-04-03 | 2018-11-13 | 南昌奇眸科技有限公司 | A kind of trademark image retrieval method |
CN108845999A (en) * | 2018-04-03 | 2018-11-20 | 南昌奇眸科技有限公司 | A kind of trademark image retrieval method compared based on multiple dimensioned provincial characteristics |
CN108897747A (en) * | 2018-04-03 | 2018-11-27 | 南昌奇眸科技有限公司 | A kind of brand logo similarity comparison method |
US20180373960A1 (en) * | 2015-12-15 | 2018-12-27 | Qing Xu | Trademark graph element identification method, apparatus and system, and computer storage medium |
WO2019028594A1 (en) * | 2017-08-07 | 2019-02-14 | 深圳益强信息科技有限公司 | Big data-based word trademark registration possibility determining method and apparatus |
CN110929059A (en) * | 2018-08-30 | 2020-03-27 | 深圳市蓝灯鱼智能科技有限公司 | Picture processing method and device, storage medium and electronic device |
CN112926592A (en) * | 2021-04-13 | 2021-06-08 | 南京邮电大学 | Trademark retrieval method and device based on improved Fast algorithm |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5943442A (en) * | 1996-06-12 | 1999-08-24 | Nippon Telegraph And Telephone Corporation | Method of image processing using parametric template matching |
CN101714254A (en) * | 2009-11-16 | 2010-05-26 | 哈尔滨工业大学 | Registering control point extracting method combining multi-scale SIFT and area invariant moment features |
-
2010
- 2010-12-29 CN CN201010622805.1A patent/CN102541954B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5943442A (en) * | 1996-06-12 | 1999-08-24 | Nippon Telegraph And Telephone Corporation | Method of image processing using parametric template matching |
CN101714254A (en) * | 2009-11-16 | 2010-05-26 | 哈尔滨工业大学 | Registering control point extracting method combining multi-scale SIFT and area invariant moment features |
Non-Patent Citations (2)
Title |
---|
刘瑞等: "简化SIFT算法及其在商标图像检索中的应用", 《计算机应用研究》 * |
徐望明: "基于内容的图像检索技术研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (26)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104732225B (en) * | 2013-12-24 | 2018-12-18 | 中国科学院深圳先进技术研究院 | image rotation processing method |
CN104732225A (en) * | 2013-12-24 | 2015-06-24 | 中国科学院深圳先进技术研究院 | Image rotation processing method |
CN104809142A (en) * | 2014-01-29 | 2015-07-29 | 北京瑞天科技有限公司 | Trademark inquiring system and method |
CN104809142B (en) * | 2014-01-29 | 2018-03-23 | 北京瑞天科技有限公司 | Trade mark inquiry system and method |
CN105096308A (en) * | 2014-05-13 | 2015-11-25 | 斯耐尔有限公司 | Logo detection by edge matching |
CN104462382B (en) * | 2014-12-11 | 2018-03-20 | 中细软移动互联科技有限公司 | Trademark image querying method |
CN104462382A (en) * | 2014-12-11 | 2015-03-25 | 北京中细软移动互联科技有限公司 | Trademark image inquiry method |
CN105989043A (en) * | 2015-02-04 | 2016-10-05 | 阿里巴巴集团控股有限公司 | Method and device for automatically acquiring trademark in commodity image and searching trademark |
CN106530194A (en) * | 2015-09-09 | 2017-03-22 | 阿里巴巴集团控股有限公司 | Method and apparatus for detecting pictures of suspected infringing products |
CN106530194B (en) * | 2015-09-09 | 2020-02-07 | 阿里巴巴集团控股有限公司 | Method and device for detecting suspected infringement product picture |
US10430687B2 (en) * | 2015-12-15 | 2019-10-01 | Qing Xu | Trademark graph element identification method, apparatus and system, and computer storage medium |
US10152650B2 (en) | 2015-12-15 | 2018-12-11 | Qing Xu | Trademark retrieval method, apparatus and system, and computer storage medium |
WO2017101224A1 (en) * | 2015-12-15 | 2017-06-22 | 徐庆 | Trademark retrieval method, apparatus and system, and computer storage medium |
US20180373960A1 (en) * | 2015-12-15 | 2018-12-27 | Qing Xu | Trademark graph element identification method, apparatus and system, and computer storage medium |
WO2019028594A1 (en) * | 2017-08-07 | 2019-02-14 | 深圳益强信息科技有限公司 | Big data-based word trademark registration possibility determining method and apparatus |
CN108897747A (en) * | 2018-04-03 | 2018-11-27 | 南昌奇眸科技有限公司 | A kind of brand logo similarity comparison method |
CN108804499A (en) * | 2018-04-03 | 2018-11-13 | 南昌奇眸科技有限公司 | A kind of trademark image retrieval method |
CN108845999A (en) * | 2018-04-03 | 2018-11-20 | 南昌奇眸科技有限公司 | A kind of trademark image retrieval method compared based on multiple dimensioned provincial characteristics |
CN108845999B (en) * | 2018-04-03 | 2021-08-06 | 南昌奇眸科技有限公司 | Trademark image retrieval method based on multi-scale regional feature comparison |
CN108763380A (en) * | 2018-05-18 | 2018-11-06 | 徐庆 | Brand recognition search method, device, computer equipment and storage medium |
CN108664945A (en) * | 2018-05-18 | 2018-10-16 | 徐庆 | Image text and form-pronunciation-meaning characteristic recognition method and device |
CN108664945B (en) * | 2018-05-18 | 2021-08-10 | 徐庆 | Image text and shape-pronunciation feature recognition method and device |
CN108763380B (en) * | 2018-05-18 | 2022-03-08 | 徐庆 | Trademark identification retrieval method and device, computer equipment and storage medium |
CN110929059A (en) * | 2018-08-30 | 2020-03-27 | 深圳市蓝灯鱼智能科技有限公司 | Picture processing method and device, storage medium and electronic device |
CN112926592A (en) * | 2021-04-13 | 2021-06-08 | 南京邮电大学 | Trademark retrieval method and device based on improved Fast algorithm |
CN112926592B (en) * | 2021-04-13 | 2022-07-26 | 南京邮电大学 | Trademark retrieval method and device based on improved Fast algorithm |
Also Published As
Publication number | Publication date |
---|---|
CN102541954B (en) | 2014-11-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN102541954B (en) | Method and system for searching trademarks | |
US9665789B2 (en) | Device and method for analyzing the correlation between an image and another image or between an image and a video | |
Wu et al. | Blind image quality assessment based on multichannel feature fusion and label transfer | |
CN101609499B (en) | Rapid fingerprint identification method | |
CN103310453B (en) | A kind of fast image registration method based on subimage Corner Feature | |
Zhu et al. | Logo matching for document image retrieval | |
US20130101186A1 (en) | Systems and methods for ridge-based fingerprint analysis | |
CN110135438B (en) | Improved SURF algorithm based on gradient amplitude precomputation | |
Wendt et al. | Wavelet leader multifractal analysis for texture classification | |
Luo et al. | Fingerprint Classification Combining Curvelet Transform and Gray‐Level Cooccurrence Matrix | |
de Ves et al. | A statistical model for magnitudes and angles of wavelet frame coefficients and its application to texture retrieval | |
CN108133211B (en) | Power distribution cabinet detection method based on mobile terminal visual image | |
CN105631872A (en) | Remote sensing image registration method based on multiple feature points | |
Dwaich et al. | Signature texture features extraction using GLCM approach in android studio | |
US8842917B2 (en) | Local feature extraction apparatus, control method therefor, and computer-readable medium | |
CN109711420B (en) | Multi-affine target detection and identification method based on human visual attention mechanism | |
CN107679494A (en) | Based on the fingerprint image matching method selectively to extend | |
Rajesh et al. | Automatic tracing and extraction of text‐line and word segments directly in JPEG compressed document images | |
Kumar et al. | FPR using machine learning with multi‐feature method | |
Zhao | Motion track enhancement method of sports video image based on otsu algorithm | |
Surmacz et al. | An improved algorithm for feature extraction from a fingerprint fuzzy image | |
CN105868760A (en) | Pattern recognition method and system | |
CN104462111A (en) | Image retrieval database establishing method | |
Yu et al. | Image quality assessment via spatial‐transformed domains multi‐feature fusion | |
Zhang et al. | A new pansharpening method using objectness based saliency analysis and saliency guided deep residual network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant |