CN105989043A - Method and device for automatically acquiring trademark in commodity image and searching trademark - Google Patents
Method and device for automatically acquiring trademark in commodity image and searching trademark Download PDFInfo
- Publication number
- CN105989043A CN105989043A CN201510059267.2A CN201510059267A CN105989043A CN 105989043 A CN105989043 A CN 105989043A CN 201510059267 A CN201510059267 A CN 201510059267A CN 105989043 A CN105989043 A CN 105989043A
- Authority
- CN
- China
- Prior art keywords
- commodity image
- point
- weight
- trade mark
- candidate feature
- 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.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a searching technology and discloses a method and a device for automatically acquiring a trademark in a commodity image and searching the trademark. The method for automatically acquiring the trademark comprises the following steps: extracting local characteristics from the commodity image to obtain multiple characteristic points; executing the following steps on each commodity image: selecting candidate characteristic points from the characteristic points of each of the other commodity images, and calculating weight of the current characteristic point according to distance priority of the current characteristic point and whether the commodity image corresponding to the candidate characteristic points and the commodity image are in the same trademark sample category; and taking a region in which the weight of the characteristic point is larger than a preset threshold value as a trademark sample region. The method provided by the invention can automatically acquire the trademark sample regions of a lot of commodity images, so as to obtain a trademark sample image, and manual annotation does not need to be carried out. The trademark is searched according to the weight of the characteristic point the closest to the characteristic point of a to-be-searched commodity image in a characteristic space, and the trademark sample category of the to-be-searched commodity image is voted, so that searching accuracy is improved.
Description
Technical field
The present invention relates to searching field, obtain trade mark and retrieval trade mark in commodity image particularly to automatic
Method and device thereof.
Background technology
Trade mark retrieval and identification system have application widely, including the intellectual property protection of trade mark, brand
The analysis of trade mark exposure and the commercial articles searching based on trademark image etc..
The general flow of brand recognition in prior art: the collection of trade mark sample image, building of aspect indexing
The retrieval of vertical and input picture.This system is described in detail below, and it generallys include following step
Rapid:
1) collection of trade mark sample image: mainly extract trade mark institute from commodity image by manual type
In region, thus obtain.Assume that training set includes that the quantity of trade mark sample image is S, institute
The quantity belonging to trademark class is T;For a series of trade mark sample images in training set
{xi, i=1,2 ..., S}, its corresponding trademark class is respectively { yi, i=1,2 ..., S}.Trade mark sample image
Feature be that trade mark region occupies main body, do not comprise or less other of commodity in picture of comprising
Region.
2) local shape factor of sample image: general employing SIFT feature extraction algorithm.Here i-th is defined
J-th Based on Feature Points of width image zooming-out isUsed here as
NiRepresent the quantity of the characteristic point of the i-th width image zooming-out.
3) index construct: open in the feature space that sample image extracts the feature composition obtaining at S, use tree
Shape data structure (such as kd-tree etc.) mode sets up index, follow-up retrieves faster to facilitate
Characteristic point most like with input feature vector (Euclidean distance is nearest) in index.
4) retrieve: SIFT feature is extracted to input picture (commodity image);Each obtaining for extraction
SIFT feature point, with it apart from immediate K characteristic point in searching feature space, and by this K
The trade mark sample classification ballot of individual Feature point correspondence adds one.
5) result output: all carried out one according to all characteristic points of image to be retrieved to input for the above-mentioned steps
After secondary ballot, add up the final score result of each trade mark sample classification, and export highest scoring
Trade mark sample classification, as the matching result of input picture.
The shortcoming of technique scheme is:
1. the quantity of trade mark sample image, directly determines the performance of final system.And this parts of images
Collection needs to consume substantial amounts of artificial mark;And the brand that comprises in training set (and new business
Mark sample image) be continuously increased, maintenance cost also improves accordingly.
2., in traditional scheme, the contribution of such purpose of each feature point pairs ballot of feature space is identical
, this point is invalid in the middle of reality is applied.
Content of the invention
It is an object of the invention to provide the side of trade mark and retrieval trade mark in a kind of automatic acquisition commodity image
Method and device thereof, can obtain the trade mark sample areas in commodity image, it is not necessary to consume substantial amounts of automatically
Artificial mark;According to the weight of the different characteristic point in trade mark sample areas, the accurate of retrieval can be promoted
Degree.
For solving above-mentioned technical problem, embodiments of the present invention disclose a kind of acquisition commodity image automatically
Middle business's calibration method, comprises the following steps:
Local shape factor is carried out to the commodity image in training set, obtains multiple features of commodity image
Point;
Repeat following steps to each commodity image:
Obtain commodity image a characteristic point as current signature point, remaining every width commodity in training set
The characteristic point of image is respectively chosen one minimum with current signature point distance as candidate feature point, according to
Commodity corresponding to the distance-taxis of each candidate feature point and current signature point and each candidate feature point
Whether image is the positive sample belonging to blanket brand sample classification with commodity image, calculates current signature point and uses
In effective weight identifying trade mark in commodity image;
After the weight of all characteristic points of commodity image all calculates and finishes, choose weight in commodity image big
In the region corresponding to the characteristic point of predetermined threshold as the trade mark sample image in commodity image.
Embodiments of the present invention also disclose business's calibration method in a kind of retrieval commodity image, in training set
The weight of all commodity image is empty more than the characteristic point constitutive characteristic in the trade mark sample areas of predetermined threshold
Between, the method comprises the following steps:
Local shape factor is carried out to the commodity image to be retrieved inputting and obtains multiple characteristic point;
Each characteristic point obtaining extraction successively is as current signature point, and finds in feature space
Immediate K the characteristic point with the distance of current signature point, and the weight of K characteristic point is added separately to
In the ballot score of the trade mark sample classification belonging to each self-corresponding commodity image of K characteristic point;
Add up the ballot score of each trade mark sample classification, by the trade mark of the trade mark sample classification of highest scoring
Retrieval result as commodity image to be retrieved.
Embodiments of the present invention also disclose the device of trade mark in a kind of automatic acquisition commodity image, including
With lower module:
Characteristic extracting module, for carrying out local shape factor to the commodity image in training set, obtains business
Multiple characteristic points of product image;
Characteristic point weight computation module, for repeating following operation to each commodity image:
Obtain commodity image a characteristic point as current signature point, remaining every width commodity in training set
The characteristic point of image is respectively chosen one minimum with current signature point distance as candidate feature point, according to
Commodity corresponding to the distance-taxis of each candidate feature point and current signature point and each candidate feature point
Whether image is the positive sample belonging to blanket brand sample classification with commodity image, calculates current signature point and uses
In effective weight identifying trade mark in commodity image;
Module is chosen in trade mark region, for finishing when the weight of all characteristic points of commodity image all calculates
After, choose weight in commodity image and be more than the region corresponding to characteristic point of predetermined threshold as commodity image
In trade mark sample image.
Embodiments of the present invention also disclose a kind of device retrieving trade mark in commodity image, in training set
The weight of all commodity image is empty more than the characteristic point constitutive characteristic in the trade mark sample areas of predetermined threshold
Between, this device includes with lower module:
Characteristic extracting module, obtains many for carrying out local shape factor to the commodity image to be retrieved of input
Individual characteristic point;
Weight votes accumulator module, each characteristic point successively extraction being obtained as current signature point,
And find and immediate K the characteristic point of distance of current signature point in feature space, and by K spy
Levy weight a little and be added to the trade mark sample classification belonging to each self-corresponding commodity image of K characteristic point respectively
Ballot score in;
Module retrieved by trade mark, for adding up the ballot score of each trade mark sample classification, by highest scoring
The trade mark of trade mark sample classification is as the retrieval result of commodity image to be retrieved.
Compared with prior art, the main distinction and effect thereof are embodiment of the present invention:
Can automatically obtain the trade mark sample areas in shiploads of merchandise image, obtain trade mark sample image, no
Need to consume substantial amounts of artificial mark, and with the increase of trade mark quantity and commodity image, maintenance cost
Remain unchanged.
According in feature space with the characteristic point of commodity image to be retrieved apart from the power of immediate characteristic point
Weight, the trade mark sample classification belonging to commodity image to be retrieved is voted, and improves the accuracy of retrieval.
Further, if the characteristic point of commodity image is more likely to training set corresponding to positive sample
Candidate feature Point matching, then increase its weight, otherwise then reduce, and improves the accuracy rate of system and recalls
Rate.
Further, use employing local feature when extracting feature to go for image when extracting feature to mix
Fold and have situation about blocking.
Further, index is set up to feature space, can more quickly in search index with input feature vector
Most like characteristic point.
Brief description
Fig. 1 is the relation schematic diagram of commodity image and trade mark sample image;
Fig. 2 is the schematic diagram that trade mark sample classification comprises multiple trade mark sample images;
Fig. 3 is the stream of business's calibration method in a kind of automatic acquisition commodity image in first embodiment of the invention
Journey schematic diagram;
Fig. 4 is that in third embodiment of the invention, a kind of flow process retrieving business's calibration method in commodity image is shown
It is intended to;
Fig. 5 is the knot of the device of trade mark in a kind of automatic acquisition commodity image in four embodiment of the invention
Structure schematic diagram;
Fig. 6 is that in sixth embodiment of the invention, a kind of structure retrieving the device of trade mark in commodity image is shown
It is intended to.
Detailed description of the invention
In the following description, many technology are proposed in order to make reader be more fully understood that the application thin
Joint.But, even if it will be understood by those skilled in the art that do not have these ins and outs and based on
The many variations of following embodiment and modification, it is also possible to realize that each claim of the application is required and protect
The technical scheme protected.
Term is explained:
Trade mark sample image: refer in particular to not comprise or seldom comprise background in the present invention, only comprise trade mark
Image, typically obtains from commodity image by way of artificial mark.It is illustrated in fig. 1 shown below as commodity
Image and the relation schematic diagram of trade mark sample image, wherein full figure is commodity image, and rectangle frame chooses part
For trade mark sample image.
Trade mark sample classification: refer to that sample image or commodity image carry out group according to the mode of sample classification
Knit.Such as " Starbucks " is exactly a trade mark sample classification, and it may comprise under varying environment
Multiple trade mark sample images, are illustrated in figure 2 trade mark sample classification and comprise showing of multiple trade mark sample images
It is intended to.
Brand recognition (system): input an image to be retrieved (may comprise or not comprise trade mark),
Requirement system is capable of identify that and returns whether this image comprises the brand in trade mark sample classification, concrete product
Board information and the region at place.
The general flow of brand recognition: common flow process generally comprises the collection of sample image, aspect indexing
Foundation and three steps of retrieval of input picture (or image to be retrieved).
Local feature: global characteristics is used to describe the gross feature of whole image, such as color histogram.
The shortcoming of global characteristics is not to be suitable for image aliasing and has situation about blocking.Local feature typically wraps
Containing the segment space scope in image, a good local feature needs possess following character: repeatable,
Uniqueness, locality, quantitative, accuracy, high efficiency, is wherein most important with repeatability again.
Local feature mates: substantially can be attributed to and carry out similitude between high dimension vector by distance function
The problem of retrieval.Substantially having two class solutions, the first is by the method for exhaustion (linear scanning method),
Point in data set will enter row distance one by one with query point and compare;The second is to set up index to carry out quickly
Coupling, the kd tree such as commonly used and improved kd tree query mode (BBF, Best-Bin-First) etc..
For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with accompanying drawing to this
Bright embodiment is described in further detail.
First embodiment of the invention relates to business's calibration method in a kind of commodity image of acquisition automatically, and Fig. 3 is
This obtains the schematic flow sheet of business's calibration method in commodity image automatically.
Specifically, as it is shown on figure 3, this obtains business's calibration method in commodity image automatically includes following step
Rapid:
Step 101, carries out local shape factor to the commodity image in training set, obtains commodity image
Multiple characteristic points.
Repeat following steps to each commodity image:
Step 102, obtain commodity image a characteristic point as current signature point, in training set its
The characteristic point of remaining every width commodity image is chosen one minimum with current signature point distance as candidate spy
Levy a little, the distance-taxis according to each candidate feature point and current signature point and each candidate feature point institute
Whether corresponding commodity image is the positive sample belonging to blanket brand sample classification with commodity image, calculates and works as
Front characteristic point is for effective weight identifying trade mark in commodity image.
Can be determined by distance when being appreciated that and choose candidate feature point, such as Euclidean distance, remaining
String covariance distance.Additionally, not necessarily choose candidate feature point according to distance, other represent characteristic point phase
Also permissible like the method for degree.
Step 103, after the weight of all characteristic points of commodity image all calculates and finishes, chooses commodity figure
In Xiang, weight is more than the region corresponding to characteristic point of predetermined threshold as the trade mark sample graph in commodity image
Picture.
Present embodiment can obtain the trade mark sample areas in shiploads of merchandise image automatically, obtains trade mark sample
This image, it is not necessary to consume substantial amounts of artificial mark, and with the increase of trade mark quantity and commodity image,
Maintenance cost remains unchanged.
Second embodiment of the invention relates to business's calibration method in a kind of commodity image of acquisition automatically, and second is real
Mode of executing is improved on the basis of the first embodiment, mainly thes improvement is that: if commodity
The characteristic point of image is more likely to the candidate feature Point matching with training set corresponding to positive sample, then increase
Its weight, on the contrary then reduce, improve accuracy rate and the recall rate of system;Local is used when extracting feature
Feature goes for image aliasing and has situation about blocking.Specifically:
In above-mentioned steps 102, " distance-taxis according to each candidate feature point and current signature point with
And whether each commodity image corresponding to candidate feature point is to belong to blanket brand sample class with commodity image
The positive sample of purpose, calculates current signature point for effective weight identifying trade mark in commodity image " step
Including sub-step:
Step 1021, according to the distance size of each candidate feature point and current signature point, to each candidate
Characteristic point carries out ascending order arrangement;
Step 1022, judges that in ascending order arrangement, each commodity image corresponding to candidate feature point is whether successively
For positive sample;
Step 1023, whether the commodity image according to corresponding to candidate feature point is the judgement knot of positive sample
Really, and candidate feature point ascending order arrangement in position, calculate this feature point weight, wherein, institute
Corresponding commodity image is that the candidate feature point of positive sample is more forward in ascending order arrangement, then to current signature
The weight contribution value of point is bigger.
Wherein, after each candidate feature point carries out ascending order arrangement according to the distance size with current signature point,
Sort more forward i.e. less with the distance of current signature point.
Preferably, in above-mentioned steps 1023, including sub-step:
If the commodity image corresponding to candidate feature point is positive sample, then the weight by this candidate feature point
Contribution margin increases in the weight of current signature point;
If the commodity image corresponding to candidate feature point is negative sample, then do not increase the power of current signature point
Weight, but reduce the weight contribution value coming this candidate feature point candidate feature point thereafter.
Wherein negative sample refers to that the commodity image corresponding to candidate feature point is not belonging to same with commodity image
The commodity image of trade mark sample classification.
If the characteristic point of commodity image is more likely to the candidate feature with training set corresponding to positive sample
Point matching, then increase its weight, otherwise then reduce, and improves accuracy rate and the recall rate of system.
Furthermore, it is to be understood that in other embodiments of the present invention, it is also possible to take other to calculate and work as
The method of front characteristic point weight, and it is not limited to this.For example, if the commodity corresponding to candidate feature point
Image is positive sample, then increase to the weight contribution value of this candidate feature point in the weight of current signature point;
If the commodity image corresponding to candidate feature point is negative sample, then reduce the weight of current signature point accordingly,
Candidate feature point corresponding to negative sample does not affect the power coming this candidate feature point candidate feature point thereafter
Weight contribution margin.
Preferably, in above-mentioned steps 1023, sub-step is specifically included:
Initialize Qk=0, Pk=0;Wherein, k represents kth width commodity image in training set;
Ascending order arrangement for each candidate feature pointWherein,Represent kth width
In commodity image, the candidate feature point minimum with j-th characteristic point distance of commodity image i, S-1 represents
In training set in addition to commodity image i the number of remaining commodity image, if candidate feature pointCorresponding
Commodity image be positive sample, then use below equation update PkAnd Qk:
Pk=Pk-1+1
If negative sample, then below equation is used to update PkAnd Qk:
Qk=Qk-1
Pk=Pk-1
To current signature pointWeight use below equation be normalized:
Wherein,Represent the quantity for positive sample in each commodity image corresponding to candidate feature point.
Preferably, in a step 101, local feature is Scale invariant features transform feature.Extract feature
Shi Caiyong local feature goes for image aliasing and has situation about blocking.
Furthermore, it is to be understood that Scale invariant features transform feature (i.e. Scale-invariant feature
Transform, is called for short SIFT feature) generally include the X-coordinate of 1 dimension, the Y coordinate of 1 dimension, 1 dimension
Dimensional information, 1 dimension principal direction information and 128 dimension Feature Descriptor information.The present invention's
In other embodiments, it is also possible to extract local feature otherwise, such as extract SURF feature.
As the preference of present embodiment, the flow process obtaining trade mark sample image from commodity image is main
Including:
1. training set prepares.The sample included in training set compared with traditional scheme, required for this preference
Picture, not as traditional scheme requires it is the trade mark region marking from commodity picture and extracting, and permissible
It is the commodity picture itself comprising trade mark;Trade mark region is calculated automatically from by subsequent algorithm, nothing
Must manual intervention;And commodity picture can also be automatically obtained by technological means or simply by artificially collecting,
Such as utilize text search engine (Baidu, Taobao), obtained by the keyword batch of brand trademark, make
Training set for this preference.Assume that training set includes that S opens sample image altogether, be respectively belonging to T business
Mark classification: i.e. for a series of trade mark sample image { xi, i=1,2 ..., S}, its corresponding trade mark sample classification
It is respectively { yi, i=1,2 ..., S}.
2. local shape factor.The general SIFT algorithm using classics.J-th characteristic point of the i-th width image zooming-out
It is expressed asUsed here as NiRepresent the characteristic point of the i-th width image zooming-out
Quantity.
3. weight calculation.For each characteristic point in each width trade mark samples pictures, find this feature respectively
Point with in each characteristic point of other S-1 width images, Euclidean distance minimum characteristic point;Concrete steps describe
As follows:
1) assumeIt is j-th characteristic point in the i-th width image, calculate itself and other S-1 each feature of width image
The Euclidean distance of point.
2) choosing in above-mentioned calculated result, (S-1 width image altogether does not include the i-th width image originally to each figure
Body) characteristic point minimum with input feature vector point distance, it as candidate point, is designated asRepresent kth width figure
In Xiang, the characteristic point minimum with j-th characteristic point Euclidean distance in the i-th width image, useRepresent both
Between distance.
3) according toSize pairCarry out ascending order arrangement, form new sequence: trade mark samples pictures
{xk, k=1,2 ..., S-1} and corresponding trade mark sample classification { yk, k=1,2 ..., S-1}.
4) as k=0, Q is initializedk=0, Pk=0.
5) sequence for step 3 outputIn each characteristic point, if meet:
yi=yk
Picture belonging to i.e. k-th characteristic point, belongs to blanket brand sample classification with input picture, then uses
Below equation updates PkAnd Qk:
Pk=Pk-1+1
If be unsatisfactory for, then below equation is used to update:
Qk=Qk-1
Pk=Pk-1
The principle of above-mentioned formula is: the feature point pairs score value of the negative sample in matching sequence is not contributed, but
The weight (because denominator k adds) of the characteristic point of the positive sample at its rear for the ranking can be reduced.
6) sequence is worked asIn after all of characteristic point all calculated one time, use below equation to obtain final
'sWeight:
In above formulaRepresent that S opens in sample image altogether, with yiThe identical image of classification (positive sample, but
Do not comprise input picture itself) quantity.This formula is meant that, if positive sample ranking is more forward,
Then final score is higher;The interval of whole score is between 0~1;If sequenceIn positive sample standard deviation go out
Before present negative sample, then must be divided into 1.
7) result exporting in above formulaCan be as the weight of j-th feature of the i-th width image;This spy
The quantity levying in all positive samples coupling is more, and in all negative samples, the quantity of coupling is fewer, then its
Importance and weight are bigger;The high characteristic point of these weights has generally corresponded to the trade mark sample in commodity image
One's respective area.By setting a rational threshold θ, all characteristic points meeting following condition can be extracted
Validity feature point as brand recognition:
Third embodiment of the invention relates to business's calibration method in a kind of retrieval commodity image, and Fig. 4 is this inspection
The schematic flow sheet of business's calibration method in rope commodity image.
Specifically, in training set, the weight of all commodity image is more than the trade mark sample areas of predetermined threshold
Interior characteristic point constitutive characteristic space, as shown in Figure 4, in this retrieval commodity image, business's calibration method includes
Following steps:
Step 401, carries out local shape factor to the commodity image to be retrieved inputting and obtains multiple characteristic point.
Step 402, each characteristic point obtaining extraction successively is as current signature point, and in feature
Space is found immediate K the characteristic point of distance with current signature point, and by the power of K characteristic point
It is heavily added separately to the ballot score of trade mark sample classification belonging to each self-corresponding commodity image of K characteristic point
In.
Step 403, adds up the ballot score of each trade mark sample classification, by the trade mark sample of highest scoring
The trade mark of classification is as the retrieval result of commodity image to be retrieved.
Furthermore, it is to be understood that under normal circumstances, the part that in commodity image, trade mark region occupies is fewer,
If the characteristic point constitutive characteristic space do not extracted in advance in trade mark sample areas, and direct statistical nature point
Ballot score if, may non-trade mark provincial characteristics point add up the spy in branch's covering trade mark region
Levy a score, thus can not correctly retrieve the trade mark sample classification of commodity image.
And the weight of all commodity original images is generally corresponding more than the characteristic point of predetermined threshold in training set
The trade mark region of commodity original image, therefore, in present embodiment using the characteristic point in trade mark region as
Feature space, can solve the problem that the branch that obtains that above-mentioned non-trade mark provincial characteristics point is added up covers trade mark region
The situation of characteristic point score.
Preferably, " each characteristic point successively extraction being obtained as current signature point, and spy
Levy in space and find and immediate K the characteristic point of distance of current signature point " step in, including with
Lower sub-step:
Tree data structure is used to set up index to feature space.
Find the current signature point with commodity image to be retrieved by search index apart from each spy of immediate K
Levy a little.
Index is set up to feature space, spy that can be most like with input feature vector in search index more quickly
Levy a little.Furthermore, it is to be understood that tree sets up index can use such as kd tree and improved
Kd tree query mode (BBF, Best-Bin-First) etc..
Present embodiment according in feature space with the characteristic point of commodity image to be retrieved apart from immediate
The weight of characteristic point, the trade mark sample classification belonging to commodity image to be retrieved is voted, and improves inspection
The accuracy of rope.
As the preference of present embodiment, in retrieval commodity image, the flow process of trade mark specifically includes that
1. retrieve: SIFT feature is extracted to input picture (commodity image);For extract obtain each
Individual SIFT feature point, finds in feature space with it closest to K the feature of (Euclidean distance is minimum)
Point, and the trade mark classification ballot of this K Feature point correspondence is increased corresponding weight.
2. result output: all carried out according to all characteristic points of image to be retrieved to input for the above-mentioned steps
After single ballot, add up the final score result of each trade mark sample classification, and export the business of highest scoring
This classification of standard specimen, as the matching result of input picture.
The each method embodiment of the present invention all can realize in modes such as software, hardware, firmwares.No matter
The present invention is to realize with software, hardware or firmware mode, and instruction code may be stored in any class
In the addressable memory of computer of type (for example permanent or revisable, volatibility or non-
Volatibility, solid-state or non-solid, fixing or removable medium etc.).Equally,
Memory can e.g. programmable logic array (Programmable Array Logic, be called for short
" PAL "), random access memory (Random Access Memory, be called for short " RAM "),
Programmable read only memory (Programmable Read Only Memory is called for short " PROM "),
Read-only storage (Read-Only Memory is called for short " ROM "), electrically erasable are read-only
Memory (Electrically Erasable Programmable ROM is called for short " EEPROM "),
Disk, CD, digital versatile disc (Digital Versatile Disc is called for short " DVD ") etc..
Four embodiment of the invention relates to the device of trade mark in a kind of automatic acquisition commodity image, and Fig. 5 is
This obtains the structural representation of the device of trade mark in commodity image automatically.
Specifically, this device automatically obtaining trade mark in commodity image includes with lower module as shown in Figure 5:
Characteristic extracting module, for carrying out local shape factor to the commodity image in training set, obtains business
Multiple characteristic points of product image;
Characteristic point weight computation module, for repeating following operation to each commodity image:
Obtain commodity image a characteristic point as current signature point, remaining every width commodity in training set
The characteristic point of image is respectively chosen one minimum with current signature point distance as candidate feature point, according to
Commodity corresponding to the distance-taxis of each candidate feature point and current signature point and each candidate feature point
Whether image is the positive sample belonging to blanket brand sample classification with commodity image, calculates current signature point and uses
In effective weight identifying trade mark in commodity image;
Module is chosen in trade mark region, for finishing when the weight of all characteristic points of commodity image all calculates
After, choose weight in commodity image and be more than the region corresponding to characteristic point of predetermined threshold as commodity image
In trade mark sample image.
Present embodiment can obtain the trade mark sample areas in shiploads of merchandise image automatically, obtains trade mark sample
This image, it is not necessary to consume substantial amounts of artificial mark, and with the increase of trade mark quantity and commodity image,
Maintenance cost remains unchanged.
First embodiment is the method embodiment corresponding with present embodiment, and present embodiment can be with
First embodiment is worked in coordination enforcement.The relevant technical details mentioned in first embodiment is in this enforcement
In mode still effectively, in order to reduce repetition, repeat no more here.Correspondingly, present embodiment carries
To relevant technical details be also applicable in the first embodiment.
Fifth embodiment of the invention relates to the device of trade mark in a kind of automatic acquisition commodity image, and the 5th is real
Mode of executing is improved on the basis of four embodiments, mainly thes improvement is that: if commodity
The characteristic point of image is more likely to the candidate feature Point matching with training set corresponding to positive sample, then increase
Its weight, on the contrary then reduce, improve accuracy rate and the recall rate of system;Local is used when extracting feature
Feature goes for image aliasing and has situation about blocking.Specifically:
In features described above point weight computation module, including submodule:
Candidate feature point sorting sub-module, for the distance according to each candidate feature point and current signature point
Size, carries out ascending order arrangement to each candidate feature point;
Positive sample judges submodule, for judging ascending order in arranging corresponding to each candidate feature point successively
Whether commodity image is positive sample;
Whether weight calculation key submodule, be just for the commodity image according to corresponding to candidate feature point
The judged result of sample, and the position that candidate feature point is in ascending order arrangement, calculate the power of this feature point
Weight, wherein, corresponding commodity image is that the candidate feature point of positive sample is more forward in ascending order arrangement,
Then bigger to the weight contribution value of current signature point.
Preferably, in weight calculation key submodule, including submodule:
Positive sample process submodule, if being positive sample for the commodity image corresponding to candidate feature point,
Then the weight contribution value of this candidate feature point is increased in the weight of current signature point;
Negative sample processes submodule, if being negative sample for the commodity image corresponding to candidate feature point,
Then do not increase the weight of current signature point, but reduce and come this candidate feature point candidate feature point thereafter
Weight contribution value.
Preferably, in characteristic point weight computation module, submodule is also included:
Initialization submodule, is used for initializing Qk=0, Pk=0;Wherein, k represents kth width in training set
Commodity image;
Weight adds up submodule, for the ascending order arrangement for each candidate feature pointWherein,Represent in kth width commodity image, j-th with commodity image i
The minimum candidate feature point of characteristic point distance, S-1 represents in training set remaining commodity in addition to commodity image i
The number of image, if candidate feature pointCorresponding commodity image is positive sample, then use following public
Formula updates PkAnd Qk:
Pk=Pk-1+1
If negative sample, then below equation is used to update PkAnd Qk:
Qk=Qk-1
Pk=Pk-1
Weight normalizes submodule, for current signature pointWeight use below equation be normalized:
Wherein,Represent the quantity for positive sample in each commodity image corresponding to candidate feature point.
Preferably, in characteristic extracting module, local feature is Scale invariant features transform feature.
Second embodiment is the method embodiment corresponding with present embodiment, and present embodiment can be with
Second embodiment is worked in coordination enforcement.The relevant technical details mentioned in second embodiment is in this enforcement
In mode still effectively, in order to reduce repetition, repeat no more here.Correspondingly, present embodiment carries
To relevant technical details be also applicable in the second embodiment.
Sixth embodiment of the invention relates to a kind of device retrieving trade mark in commodity image, and Fig. 6 is this inspection
The structural representation of the device of trade mark in rope commodity image.
Specifically, in this retrieval commodity image trade mark device training set in the weight of all commodity image
More than the characteristic point constitutive characteristic space in the trade mark sample areas of predetermined threshold, as shown in Figure 6, this dress
Put and include with lower module:
Characteristic extracting module, obtains many for carrying out local shape factor to the commodity image to be retrieved of input
Individual characteristic point;
Weight votes accumulator module, each characteristic point successively extraction being obtained as current signature point,
And find and immediate K the characteristic point of distance of current signature point in feature space, and by K spy
Levy weight a little and be added to the trade mark sample classification belonging to each self-corresponding commodity image of K characteristic point respectively
Ballot score in;
Module retrieved by trade mark, for adding up the ballot score of each trade mark sample classification, by highest scoring
The trade mark of trade mark sample classification is as the retrieval result of commodity image to be retrieved.
Preferably, in weight votes accumulator module, following submodule is also included:
Submodule set up in index, is used for using tree data structure to set up index to feature space.
Indexed search submodule, for by search index find with the characteristic point of commodity image to be retrieved away from
From each characteristic point of immediate K.
Index is set up to feature space, spy that can be most like with input feature vector in search index more quickly
Levy a little.Furthermore, it is to be understood that tree sets up index can use such as kd tree and improved
Kd tree query mode (BBF, Best-Bin-First) etc..
Present embodiment according in feature space with the characteristic point of commodity image to be retrieved apart from immediate
The weight of characteristic point, the trade mark sample classification belonging to commodity image to be retrieved is voted, and improves inspection
The accuracy of rope.
3rd embodiment is the method embodiment corresponding with present embodiment, and present embodiment can be with
3rd embodiment is worked in coordination enforcement.The relevant technical details mentioned in 3rd embodiment is in this enforcement
In mode still effectively, in order to reduce repetition, repeat no more here.Correspondingly, present embodiment carries
To relevant technical details be also applicable in the 3rd embodiment.
The present invention proposes a kind of new scheme, by the trade mark region in automatic study and discovery picture
Characteristic point, thus solve the collection of commodity original image and the problem that mark needs consumption is artificial in a large number;
Meanwhile, the different weight of each characteristic point (depending on uniqueness and the robustness of this feature) is given, should
If the positive sample (belonging to identical classification with picture to be retrieved) that feature is more likely to be trained to concentrate
In characteristic matching, then increasing its weight, otherwise then reducing, the system that finally improves is in retrieving
Accuracy rate and recall rate.
It should be noted that each module mentioned in the present invention each equipment embodiment is all logic module,
Physically, a logic module can be a physical module, it is also possible to be the one of a physical module
Part, can also realize with the combination of multiple physical modules, the physics realization side of these logic modules itself
Formula is not most important, and the combination of the function that these logic modules are realized is only the solution present invention and is carried
The key of the technical problem going out.Additionally, for the innovative part highlighting the present invention, the present invention is above-mentioned respectively to be set
The module less close with solving technical problem relation proposed by the invention is not drawn by standby embodiment
Entering, this is not intended that the said equipment embodiment does not exist other module.
It should be noted that in the claim and specification of this patent, the first and second grades it
The relational terms of class is used merely to separate an entity or operation with another entity or operating space,
And not necessarily require or imply there is the relation of any this reality or suitable between these entities or operation
Sequence.And, term " includes ", "comprising" or its any other variant are intended to nonexcludability
Comprise, so that include that the process of a series of key element, method, article or equipment not only include that
A little key elements, but also include other key elements being not expressly set out, or also include for this process,
The intrinsic key element of method, article or equipment.In the case of there is no more restriction, by statement " bag
Include one " key element that limits, it is not excluded that at process, method, the article including described key element or set
Other identical element is there is also in Bei.
Although by referring to some of the preferred embodiment of the invention, the present invention has been shown and
Describe, but it will be understood by those skilled in the art that and can in the form and details it be made respectively
Plant and change, without departing from the spirit and scope of the present invention.
Claims (14)
1. business's calibration method in an automatic acquisition commodity image, it is characterised in that comprise the following steps:
Local shape factor is carried out to the commodity image in training set, obtains multiple features of commodity image
Point;
Repeat following steps to each commodity image:
Obtain a characteristic point of described commodity image as current signature point, remaining every width in training set
The characteristic point of commodity image is respectively chosen one minimum with current signature point distance as candidate feature point,
Distance-taxis according to each candidate feature point described and current signature point and each candidate feature point institute
Whether corresponding commodity image is the positive sample belonging to blanket brand sample classification with described commodity image, meter
Calculate current signature point for effective weight identifying trade mark in described commodity image;
After the weight of all characteristic points of described commodity image all calculates and finishes, choose described commodity image
Middle weight is more than the region corresponding to characteristic point of predetermined threshold as the trade mark sample in described commodity image
One's respective area.
2. business's calibration method in automatic acquisition commodity image according to claim 1, its feature exists
In in described " distance-taxis according to each candidate feature point described and current signature point and each time
Select whether the commodity image corresponding to characteristic point is to belong to blanket brand sample classification with described commodity image
Positive sample, calculate current signature point for effective weight identifying trade mark in described commodity image " step
In Zhou, including sub-step:
According to the distance size of each candidate feature point and current signature point, to each candidate feature point described
Carry out ascending order arrangement;
Judge in the arrangement of described ascending order, whether each commodity image corresponding to candidate feature point is just successively
Sample;
Whether the commodity image according to corresponding to candidate feature point is the judged result of positive sample, and candidate
Position in the arrangement of described ascending order for the characteristic point, calculates the weight of this feature point, wherein, corresponding business
Product image is that the candidate feature point of positive sample is more forward in the arrangement of described ascending order, then to described current signature
The weight contribution value of point is bigger.
3. business's calibration method in automatic acquisition commodity image according to claim 2, its feature exists
In, described " whether the commodity image according to corresponding to candidate feature point is the judged result of positive sample,
And candidate feature point described ascending order arrangement in position, calculate this feature point weight " step in,
Including sub-step:
If the commodity image corresponding to candidate feature point is positive sample, then the weight by this candidate feature point
Contribution margin increases in the weight of described current signature point;
If the commodity image corresponding to candidate feature point is negative sample, then do not increase described current signature point
Weight, but reduce and come the weight contribution value of this candidate feature point candidate feature point thereafter.
4. business's calibration method in automatic acquisition commodity image according to claim 2, its feature exists
In, described " whether the commodity image according to corresponding to candidate feature point is the judged result of positive sample,
And candidate feature point described ascending order arrangement in position, calculate this feature point weight " step in,
Including sub-step:
Initialize Qk=0, Pk=0;Wherein, k represents kth width commodity image in training set,Represent
In kth width commodity image, the candidate feature minimum with j-th characteristic point distance of described commodity image i
Point;
Ascending order arrangement for each candidate feature point describedWherein, S-1 represents
In training set in addition to commodity image i the number of remaining commodity image, if candidate feature pointCorresponding
Commodity image be positive sample, then use below equation update PkAnd Qk:
Pk=Pk-1+1
If negative sample, then below equation is used to update PkAnd Qk:
Qk=Qk-1
Pk=Pk-1
To described current signature pointWeight use below equation be normalized:
Wherein,Represent the quantity for positive sample in each commodity image corresponding to candidate feature point.
5. the side of trade mark in automatic acquisition commodity image according to any one of claim 1 to 4
Method, it is characterised in that " described local shape factor carried out to the commodity image in training set, obtains
In the step of multiple characteristic points of commodity image ", described local feature is Scale invariant features transform feature.
6. business's calibration method in a retrieval commodity image, it is characterised in that all commodity in training set
The weight of image is more than the characteristic point constitutive characteristic space in the trade mark sample areas of predetermined threshold, the method
Comprise the following steps:
Local shape factor is carried out to the commodity image to be retrieved inputting and obtains multiple characteristic point;
Each characteristic point obtaining described extraction successively is as current signature point and empty in described feature
Between middle find and immediate K the characteristic point of distance of current signature point, and by described K characteristic point
Weight is added separately to the ballot of the trade mark sample classification belonging to each self-corresponding commodity image of K characteristic point
In score;
Add up the ballot score of each trade mark sample classification, by the trade mark of the trade mark sample classification of highest scoring
Retrieval result as described commodity image to be retrieved.
7. business's calibration method in retrieval commodity image according to claim 6, it is characterised in that
Described " each characteristic point successively described extraction being obtained as current signature point, and described spy
Levy in space and find and immediate K the characteristic point of distance of current signature point " step in, including with
Lower sub-step:
Tree data structure is used to set up index to described feature space;
The current signature point distance found with described commodity image to be retrieved by the described index of retrieval is connect most
The each characteristic point of near K.
8. the device of trade mark in an automatic acquisition commodity image, it is characterised in that include with lower module:
Characteristic extracting module, for carrying out local shape factor to the commodity image in training set, obtains business
Multiple characteristic points of product image;
Characteristic point weight computation module, for repeating following operation to each commodity image:
Obtain a characteristic point of described commodity image as current signature point, remaining every width in training set
The characteristic point of commodity image is respectively chosen one minimum with current signature point distance as candidate feature point,
Distance-taxis according to each candidate feature point described and current signature point and each candidate feature point institute
Whether corresponding commodity image is the positive sample belonging to blanket brand sample classification with described commodity image, meter
Calculate current signature point for effective weight identifying trade mark in described commodity image;
Module is chosen in trade mark region, and the weight for all characteristic points when described commodity image has all calculated
Bi Hou, chooses weight in described commodity image and is more than the region corresponding to characteristic point of predetermined threshold as institute
State the trade mark sample areas in commodity image.
9. the device of trade mark in automatic acquisition commodity image according to claim 8, its feature exists
In, in described characteristic point weight computation module, including submodule:
Candidate feature point sorting sub-module, for the distance according to each candidate feature point and current signature point
Size, carries out ascending order arrangement to each candidate feature point described;
Positive sample judges submodule, for judging that in the arrangement of described ascending order, each candidate feature point institute is right successively
Whether the commodity image answered is positive sample;
Whether weight calculation key submodule, be just for the commodity image according to corresponding to candidate feature point
The judged result of sample, and the position that candidate feature point is in the arrangement of described ascending order, calculate this feature point
Weight, wherein, corresponding commodity image be positive sample candidate feature point described ascending order arrangement in
More forward, then bigger to the weight contribution value of described current signature point.
10. the device of trade mark in automatic acquisition commodity image according to claim 9, its feature
It is, in described weight calculation key submodule, including submodule:
Positive sample process submodule, if being positive sample for the commodity image corresponding to candidate feature point,
Then the weight contribution value of this candidate feature point is increased in the weight of described current signature point;
Negative sample processes submodule, if being negative sample for the commodity image corresponding to candidate feature point,
Then do not increase the weight of described current signature point, but reduce and come this candidate feature point candidate feature thereafter
The weight contribution value of point.
The device of trade mark, its feature in 11. automatic acquisition commodity image according to claim 9
It is, in described characteristic point weight computation module, also include submodule:
Initialization submodule, is used for initializing Qk=0, Pk=0;Wherein, k represents kth width in training set
Commodity image,Represent in kth width commodity image, j-th characteristic point distance with described commodity image i
Minimum candidate feature point;
Ascending order arrangement for each candidate feature point describedWherein, S-1 represents
In training set in addition to commodity image i the number of remaining commodity image, if candidate feature pointCorresponding
Commodity image be positive sample, then use below equation update PkAnd Qk:
Pk=Pk-1+1
If negative sample, then below equation is used to update PkAnd Qk:
Qk=Qk-1
Pk=Pk-1
To described current signature pointWeight use below equation be normalized:
Wherein,Represent the quantity for positive sample in each commodity image corresponding to candidate feature point.
12. according to Claim 8 to trade mark in the commodity image of acquisition automatically according to any one of 11
Device, it is characterised in that in described characteristic extracting module, described local feature is scale invariant feature
Transform characteristics.
13. 1 kinds of devices retrieving trade mark in commodity image, it is characterised in that all commodity in training set
The weight of image is more than the characteristic point constitutive characteristic space in the trade mark sample areas of predetermined threshold, this device
Including with lower module:
Characteristic extracting module, obtains many for carrying out local shape factor to the commodity image to be retrieved of input
Individual characteristic point;
Weight votes accumulator module, each characteristic point being used for obtaining described extraction successively is as currently
Characteristic point, and in described feature space, find immediate K the feature of distance with current signature point
Point, and the weight by described K characteristic point is added separately to each self-corresponding commodity image institute of K characteristic point
In the ballot score of the trade mark sample classification belonging to;
Module retrieved by trade mark, for adding up the ballot score of each trade mark sample classification, by highest scoring
The trade mark of trade mark sample classification is as the retrieval result of described commodity image to be retrieved.
The device of trade mark in 14. retrieval commodity image according to claim 13, it is characterised in that
In described weight votes accumulator module, also include following submodule:
Submodule set up in index, is used for using tree data structure to set up index to described feature space;
Indexed search submodule, for being found and described commodity image to be retrieved by the described index of retrieval
Characteristic point is apart from each characteristic point of immediate K.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510059267.2A CN105989043A (en) | 2015-02-04 | 2015-02-04 | Method and device for automatically acquiring trademark in commodity image and searching trademark |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510059267.2A CN105989043A (en) | 2015-02-04 | 2015-02-04 | Method and device for automatically acquiring trademark in commodity image and searching trademark |
Publications (1)
Publication Number | Publication Date |
---|---|
CN105989043A true CN105989043A (en) | 2016-10-05 |
Family
ID=57036064
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510059267.2A Pending CN105989043A (en) | 2015-02-04 | 2015-02-04 | Method and device for automatically acquiring trademark in commodity image and searching trademark |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105989043A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108038122A (en) * | 2017-11-03 | 2018-05-15 | 福建师范大学 | A kind of method of trademark image retrieval |
CN111125418A (en) * | 2020-01-15 | 2020-05-08 | 广东工业大学 | Trademark retrieval system |
CN111199439A (en) * | 2018-11-16 | 2020-05-26 | 阿里巴巴集团控股有限公司 | Commodity information processing method and device |
CN111241330A (en) * | 2020-01-13 | 2020-06-05 | 苏宁云计算有限公司 | Commodity picture auditing method and device |
CN112136151A (en) * | 2018-05-28 | 2020-12-25 | 株式会社理光 | Image search device, image search method, commodity catalog generation system, and recording medium |
WO2022083332A1 (en) * | 2020-10-23 | 2022-04-28 | 华为技术有限公司 | Commodity data management method and apparatus, and server |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102541954A (en) * | 2010-12-29 | 2012-07-04 | 北京大学 | Method and system for searching trademarks |
CN103049512A (en) * | 2012-12-14 | 2013-04-17 | 杭州淘淘搜科技有限公司 | Blocking, weighting and matching retrieval method based on commodity image saliency map |
CN104143088A (en) * | 2014-07-25 | 2014-11-12 | 电子科技大学 | Face identification method based on image retrieval and feature weight learning |
-
2015
- 2015-02-04 CN CN201510059267.2A patent/CN105989043A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102541954A (en) * | 2010-12-29 | 2012-07-04 | 北京大学 | Method and system for searching trademarks |
CN103049512A (en) * | 2012-12-14 | 2013-04-17 | 杭州淘淘搜科技有限公司 | Blocking, weighting and matching retrieval method based on commodity image saliency map |
CN104143088A (en) * | 2014-07-25 | 2014-11-12 | 电子科技大学 | Face identification method based on image retrieval and feature weight learning |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108038122A (en) * | 2017-11-03 | 2018-05-15 | 福建师范大学 | A kind of method of trademark image retrieval |
CN108038122B (en) * | 2017-11-03 | 2021-12-14 | 福建师范大学 | Trademark image retrieval method |
CN112136151A (en) * | 2018-05-28 | 2020-12-25 | 株式会社理光 | Image search device, image search method, commodity catalog generation system, and recording medium |
US11900423B2 (en) | 2018-05-28 | 2024-02-13 | Ricoh Company, Ltd. | Image retrieval apparatus image retrieval method, product catalog generation system, and recording medium |
CN111199439A (en) * | 2018-11-16 | 2020-05-26 | 阿里巴巴集团控股有限公司 | Commodity information processing method and device |
CN111199439B (en) * | 2018-11-16 | 2023-04-14 | 阿里巴巴集团控股有限公司 | Commodity information processing method and device |
CN111241330A (en) * | 2020-01-13 | 2020-06-05 | 苏宁云计算有限公司 | Commodity picture auditing method and device |
CN111241330B (en) * | 2020-01-13 | 2022-11-18 | 苏宁云计算有限公司 | Commodity picture auditing method and device |
CN111125418A (en) * | 2020-01-15 | 2020-05-08 | 广东工业大学 | Trademark retrieval system |
WO2022083332A1 (en) * | 2020-10-23 | 2022-04-28 | 华为技术有限公司 | Commodity data management method and apparatus, and server |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110084292B (en) | Target detection method based on DenseNet and multi-scale feature fusion | |
CN105989043A (en) | Method and device for automatically acquiring trademark in commodity image and searching trademark | |
Shalunts et al. | Architectural style classification of building facade windows | |
Dong et al. | Tablesense: Spreadsheet table detection with convolutional neural networks | |
Lee et al. | Classification of leaf images | |
CN103995889B (en) | Picture classification method and device | |
CN110717534B (en) | Target classification and positioning method based on network supervision | |
CN108920580A (en) | Image matching method, device, storage medium and terminal | |
CN109829467A (en) | Image labeling method, electronic device and non-transient computer-readable storage medium | |
CN107506703A (en) | A kind of pedestrian's recognition methods again for learning and reordering based on unsupervised Local Metric | |
CN108985360A (en) | Hyperspectral classification method based on expanding morphology and Active Learning | |
CN114937179B (en) | Junk image classification method and device, electronic equipment and storage medium | |
CN109886295A (en) | A kind of butterfly recognition methods neural network based and relevant device | |
CN113705570B (en) | Deep learning-based few-sample target detection method | |
CN108647595A (en) | Vehicle recognition methods again based on more attribute depth characteristics | |
CN110008899B (en) | Method for extracting and classifying candidate targets of visible light remote sensing image | |
CN109165666A (en) | Multi-tag image classification method, device, equipment and storage medium | |
CN103839078A (en) | Hyperspectral image classifying method based on active learning | |
CN102385592A (en) | Image concept detection method and device | |
CN106295498A (en) | Remote sensing image target area detection apparatus and method | |
CN111178196B (en) | Cell classification method, device and equipment | |
Lv et al. | Object detection for sweeping robots in home scenes (ODSR-IHS): a novel benchmark dataset | |
CN110083731B (en) | Image retrieval method, device, computer equipment and storage medium | |
Kanji | Unsupervised part-based scene modeling for visual robot localization | |
CN111553361B (en) | Pathological section label identification method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20161005 |