CN103942276A - Novel logo detection technology - Google Patents
Novel logo detection technology Download PDFInfo
- Publication number
- CN103942276A CN103942276A CN201410125961.5A CN201410125961A CN103942276A CN 103942276 A CN103942276 A CN 103942276A CN 201410125961 A CN201410125961 A CN 201410125961A CN 103942276 A CN103942276 A CN 103942276A
- Authority
- CN
- China
- Prior art keywords
- key point
- sign
- tree
- model
- keypoints
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
Abstract
The invention discloses a novel logo detection technology used for detecting logos in an image. The method comprises the steps of (1) extracting multiple key points from a detected image; (2) judging whether the visual vocabulary serial number of each key point is identical with the visual vocabulary serial number of a certain root node in a logo model, (3) looking for one or more key point sets the visual vocabulary serial numbers of which are identical with the visual vocabulary serial numbers of leaf nodes of one or more tree structures for each key point with a positive judgment result, wherein the one or more key point sets meet tree structure constraint; (4) calculating the number of key point sets, corresponding to each tree structure in the one or more tree structures, found in the step (3), and determining that a logo corresponding to a certain tree structure in the one or more tree structures appears in the detected image if the number of key point sets corresponding to the tree structure is larger than a preset threshold.
Description
Technical field
The invention belongs to computer vision field, relate in particular to the extensive trade mark detection technique based on tree construction.
Background technology
The technical scheme of prior art one
First the detection architecture of prior art one proposes in document [1], has solved large-scale trade mark test problems.In the method, for each width input picture, first extract key point, and by feature, key point is described.By line between key point, and according to the direction of key point, determine the relative rotation angle of line, the feature using this as line.The line by merging with common vertex forms triangle, and using leg-of-mutton interior angle as triangle character, the method, by searching these leg-of-mutton matching relationships between band coupling picture and template picture, is carried out the coupling between final clear and definite key point.Meanwhile, the method has been introduced the method for multistage reverse indexing, can detect multiple trade mark simultaneously, therefore has very high detection efficiency.
The shortcoming of prior art one
1. the method has defined the position constraint relation between a plurality of key points, and this position constraint relation is not affine constant, and the method is not strict affine constant, can not process well the variation that the trade mark in real scene occurs;
2. the method involves a large amount of floating point arithmetics in testing process, and matching efficiency is lower;
3. the method has all been used random algorithm in training and matching process, has affected to a certain extent the stability of algorithm.
The technical scheme of prior art two
First the detection architecture of prior art two proposes in document [2], and the problem of solution is that the position relationship between key point is retrained, and improves by the coupling between a plurality of key points the precision detecting.In prior art two, for each width input picture, first extract key point and characteristic of correspondence, and set up multiple dimensioned Delaunay triangle (calling MSDT in the following text).In follow-up coupling, by the coupling between MSDT triangle, replace simple point to mate with point.
The shortcoming of prior art two
1. the method is not strict affine constant, can not process well the variation that the trade mark in real scene occurs;
2. the method need to be set up MSDT triangular net, and process of establishing efficiency is lower;
3. the method depends on the foundation of MSDT network, and in network, key point is if there is damaged, and by the formation of the final MSDT network of impact, robustness is bad;
4. the method can only detect a kind of trade mark at every turn, and when needs detect multiple trade mark, efficiency is lower.
List of references:
[1]Romberg,Stefan,et?al."Scalable?logo?recognition?in?real-world?images."Proceedings?of?the1st?ACM?International?Conference?on?Multimedia?Retrieval.ACM,2011.
[2]Kalantidis,Yannis,et?al."Scalable?triangulation-based?logo?recognition."Proceedings?of?the1st?ACM?International?Conference?on?Multimedia?Retrieval.ACM,2011.
Summary of the invention
The problem that the present invention solves is under the prerequisite of high measurement accuracy, and for a given picture, fast detecting goes out the trade mark occurring in picture.
For the deficiencies in the prior art of above introduction, the present invention mainly solves following problem:
(1) a kind of new shape descriptor is proposed, the local visual feature of while Description Image and the geometry site between key point, this descriptor has strict affine unchangeability, and the variation that two dimensional surface object is produced in real image has good robustness;
(2) in order to solve the too complicated problem of the model causing causing due to permutation and combination in many key points profile matching, this descriptor adopts tree construction to carry out two-stage arrangement to forming a plurality of key points of descriptor, and before having avoided, method need to be carried out to training pattern the process of sample reduction;
(3) according to the structure of shape descriptor, the present invention proposes a kind of new reverse indexing structure, make to detect a plurality of trade marks simultaneously, greatly improved detection efficiency, increased the practicality of this invention.
The means of technical solution problem (technical scheme)
Utilization scene of the present invention requires system to have higher efficiency and accuracy of detection simultaneously, also needs to multiple trade mark, to detect fast simultaneously.
From detection efficiency, consider, the present invention has adopted the method for key point coupling to detect object.In order further to raise the efficiency, the present invention can adopt based on
the method of model (referring to list of references 6) quantizes initial characteristics, makes can to mate with high spatiotemporal efficiency between feature.Simultaneously, the scene detecting in order to adapt to many trade marks, the present invention has adopted the method for setting up reverse indexing on the model basis training, key point feature for each input picture, the matching relationship between itself and the trade mark that trains can be found out fast, thereby multiple trade mark can be detected simultaneously.
From accuracy of detection, consider, first the present invention has proposed a kind of new shape descriptor, this descriptor has strict affine unchangeability, for two-dimentional trade mark in reality scene because the deformation that the variation of shooting angle produces can be carried out modeling, so this descriptor has good robustness.Because the coupling between descriptor is the coupling between a plurality of key points that meet certain positional relationship, therefore can reduce the generation of mistake coupling, the local feature of image and position are mated simultaneously, effectively improve the precision detecting.In addition, this descriptor, by introducing tree structure, has effectively been avoided in other methods based on vision phrase, due to the dimension disaster problem that permutation and combination problem causes, has reduced the complexity of model.More detailed method, referring to embodiment.
According to embodiments of the invention, a kind of label detection method that is used in image detecting sign is provided, comprise the following steps: step 1, from detected image, extract a plurality of key point k
1～k
i, the sum that wherein I is key point; Step 2, for described a plurality of key point k
1～k
iin each key point k
m, judge its visual vocabulary sequence number I (k
m) whether with sign model in certain root node k
avisual vocabulary sequence number 1 (k
a) identical, m ∈ 1～I wherein, wherein, described root node k
abe the root node of the one or more tree constructions in described sign model, wherein, each in described one or more tree constructions is corresponding from different signs; Step 3, for described each key point k that judgment result is that "Yes"
mother key point of extracting from detected image is traveled through, the visual vocabulary sequence number of finding visual vocabulary sequence number and described one or more tree constructions leaf node is separately identical and meet one or more set of keypoints that tree construction retrains respectively, wherein, each set of keypoints in described one or more set of keypoints is corresponding to a tree construction in described one or more tree constructions; And step 4, calculate the number of the set of keypoints corresponding to each tree construction in described one or more tree constructions find in step 3, if the number corresponding to the described set of keypoints of certain tree construction in described one or more tree constructions is greater than predetermined threshold, judge that identify in the present described detected image of the present invention beneficial effect corresponding with described certain tree construction is mainly the following aspects.
First, trade mark detects has purposes widely at numerous areas such as copyright detection, business survey, user behavior analysis.The problem that this invention will solve finds out fast the different trade mark kinds that every width picture occurs in mass picture, has good commercial value and social benefit.
Meanwhile, the method before comparing, this invention is all improved on precision and efficiency of detecting.These raisings will contribute to increase the practicality of this invention under plurality of application scenes:
(1) shape descriptor that this invention proposes has strict affine unchangeability.Because the trade mark in reality scene picture can be because illumination condition, the reasons such as shooting angle produce larger appearance change, but because most trade marks are two dimensional surface object, affine variation can change and carry out modeling the trade mark in most reality scenes, therefore by strict affine unchangeability, this invention has good robustness, can improve the precision that trade mark detects;
(2) shape descriptor that this invention proposes retrains the geometric position of key point by introducing tree structure, and the multilevel hierarchy feature by means of tree structure, can improve the efficiency of trade mark model training and detection, and reduces the complexity of model.In the training of invention and detection method, do not involve complicated geometric position change calculations and floating point arithmetic, greatly improved operational efficiency;
(3) in practical application, often require to detect multiple trade mark, and existing method is merely able to every kind of trade mark to detect successively conventionally, when the trade mark kind that detects when needs is more, these class methods will not have practicality.In this invention, by the trade mark model to training, set up direction index structure, can to multiple trade mark, detect simultaneously simultaneously, improved the practicality of efficiency and invention.
Accompanying drawing explanation
Fig. 1 illustrates the schematic diagram of tree structure and descriptor according to an embodiment of the invention.
Embodiment
Below, by reference to the accompanying drawings the enforcement of technical scheme is described in further detail.
In this manual, for being described, principle of the present invention presented some details that relevant trade mark detects, yet, those skilled in the art will appreciate that the invention is not restricted to trade mark detects, and also can be applicable to the label detection of other type.In addition, the technological means that wherein each step adopts is example and does not mean that the restriction of uniqueness, and the present invention can have various modification, and it all falls into the protection domain limiting as claims.
For this trade mark detection method of clearer explanation, below, by the position relationship structure between location point, training method and three aspects of detection method, the present invention will be described in detail.
In many detection methods, this invention belongs to the type of the method based on key point coupling, and by introducing
model improves detection efficiency.
As a kind of based on
the trade mark detection method of model, do not do specified otherwise, hereinafter for each width picture, first use hessian affine to extract key point as detecting son, and by sift feature, key point is described, after in the code book training, search the arest neighbors of this sift feature, and using its sequence number as the visual vocabulary that represents this key point local visual external appearance characteristic.Meanwhile, for each key point, add geological information to further describe this point.Thus, arbitrary key point k in picture can be expressed as k={P (k), S (k), R (k), I (k) }, wherein p (k) represents the position of key point, s (k) represents the yardstick of key point, (wherein, the yardstick of key point and response have related to detection of key point, and this is prior art in the response of R (k) expression key point.The sub-technology of existing multiple critical point detection, the present invention has used the Corner detector as described in list of references 3 in specific implementation, specifically refers to the 3rd part of this list of references 3), I (k) represents visual vocabulary sequence number corresponding to key point.
For arbitrary picture I
i, be expressed as a crucial point set I
i=k1 ..., k
n.Given two width picture I
1and I
2(aspect low-level image feature, the present invention adopts the framework based on key point to detect, so in follow-up test, every width picture is regarded as to the set of a key point, the expression of each key point as this section mentioned above,, different pictures has different crucial point sets), if k
l∈ I
1, k
2∈ I
2and I (k
1)=I (k
2), claim k
1and k
2coupling (two key point couplings in two width pictures).
The essence of the method is by introducing the mutual alignment relation between a plurality of key points, man-to-man matching problem between a key point is changed into the matching problem of multi-to-multi, thereby improves the precision of coupling.
The present invention proposes a kind of new tree-like position relationship and retrains four positions between key point, and then the implicit shape of having expressed trade mark, and proposes a kind of tree-like shape descriptor and describe this position relationship.This shape descriptor is called to the shape descriptor (tree-based shape descriptor, TSD) based on tree herein, in order to narrate conveniently, it is TSD for hereinafter referred.
The concrete logical relation of TSD structure is: for the concrete picture of a width, the present invention proposes a kind of tree-like position relationship several key points are retrained, and with structurized descriptor TSD, this structure is described.This structure has strict affine unchangeability, and the while, because tree structure adds, is achieved stratification reverse indexing.
Before describing TSD structure, the definition of given first tree structure tr, and represent one group of key point that meets certain positional relationship by tree structure, introduce how with TSD, to describe this tree structure afterwards.As shown in Figure 1, tree structure tr has four node A, B, C, D, and it can be expressed as tr={k
a, k
b, k
c, k
d.K wherein
afor root node, k
b, k
c, k
dfor leaf node, between them, meet following geometrical-restriction relation:
, ray P (k
a) P (k
d) at ∠ P (k
b) P (k
a) P (k
c) in, simultaneously the dimensional constraints of four key points within the specific limits, i.e. max (s (k
a) ..., s (k
d))-min (s (k
a) ..., s (kd)) < θ, can establish θ=5 in the present invention.
For tree structure tr, with TSD, it is described afterwards, i.e. tsd (tr)={ I (k
a), I (k
b), I (k
c), I (kd) }.It should be noted that TSD is an ordered set.And if only if tsd (tt
1)=tsd (tr
2), i.e. two tree structure tr
1, tr
2descriptor while equating, define this two tree structure tr
1, tr
2coupling.
According to above-mentioned definition, be not difficult to draw, the corresponding unique TSD descriptor tsd of each tree structure tr, a given tsd, can search the appearance that whether has corresponding tree structure tr in Given Graph picture simultaneously.
The training method of trade mark model is described below.
In training process, for each trade mark, we wish the trade mark model that finds simultaneously to have better affine unchangeability and distinguishing.Specifically, given training set (picture that comprises trade mark), supposes each trade mark L
icorresponding N
iwidth includes this trade mark and manually marks out the picture that position appears in trade mark, selects two width picture training to obtain the subset M of a model in N width figure at every turn, is finally included thus
The model of individual subset
?
For every two width, include the go forward side by side picture I of rower note of identic trade mark
a, I
b, the trained process of model is as follows.It should be noted that due to generally, it is more stable that the key point that response is large is compared the key point that response is little, so I
aand I
bkey point all according to the response descending sort of each key point.In method, adopted herein and be similar to NMS(non maximal suppress) greedy method, find out the tree structure of appearance and root node response large (the larger meaning is to arrange from big to small according to response, considers successively these nodes) in two width figure simultaneously.Each is taken turns in circulation, first at picture I
ain find out root node, select p candidate's leaf segment point set simultaneously, require the visual vocabulary that each leaf node and root node are corresponding not identical, simultaneously the visual vocabulary of leaf node is also different between two.Then, from P candidate's leaf node, select the tree structure that meets the described requirement of a upper joint.Finally, if candidate's tree structure has found coupling tree structure in picture B, be added in the model of training trade mark.
Finally, obtain the number of times that the TSD set element of model occurs in training set according to it and carry out descending sort, get front K as final trade mark model.In the present invention, found through experiments and select K=5,000 can obtain best effect.
The principle of above-mentioned training method can be described by following false code form:
Input:
picture I
ain crucial point set;
picture I
bin crucial point set;
The root node of the given tree construction of P:=, the selected quantity as leaf node candidate key point
Output:
M
aB:=tree node is described collection;
S
aB:=selected crucial point set;
Initialization: empty M
aBand S
aB;
Each k of for
i∈ K
ado
if
then
Selected root node v=k
i;
From K
ain select P and do not belong to S
aBand key point different between two, as candidate's leaf node of root node v;
From P node, carry out tlv triple and select to arrange, and will put as root node spanning tree structure collection T
v;
Each tr of for
a∈ T
vdo
Iftr
aat K
bin found coupling tr
bthen
By tr
ain the crucial point set that comprises join S
aBin, corresponding descriptor trd
ajoin M
aBin;
end?if
end?for
end?if
end?for
The trade mark detection method of utilizing the trade mark model training is described below.
At detection-phase, given one group of trade mark model training, target is in arbitrary picture to be detected, to judge whether this picture occurs the trade mark in given trade mark set.
First, need to set up reverse indexing to the trade mark model training, make the tsd for any input, can know which kind of trade mark this tsd belongs to according to index.Concrete process of establishing is as follows: the given N that comprises
lthe set of individual trade mark
first the model of corresponding trade mark training is carried out to cascade,
wherein
represent the trade mark L having trained
imodel (right, the description collection of model).Meanwhile, for any tsd
i, tsd
j∈ M, if tsd
i=tsd
j, in index, delete tsd simultaneously
iand tsd
j(tsd
iand tsd
jbe identical descriptor, the situation is here that different trade marks is enjoyed same descriptor, therefore so this descriptor do not there is good resolution and leave out), because such descriptor does not possess enough resolution.Can obtain like this one group of unduplicated TSD set M all between two, for any tsd ∈ M, use L (tsd) to represent that it belongs to any trade mark herein.
The concrete meaning of root node index: because TSD is tree structure, according to the level of tree construction, the TSD polymerization with same root node can be become to one group, such index structure is called to root node index.Equation expression be exactly this implication.
Then, set up the index of tree construction root node, represent the mapping of a key point to tsd set with f (), specific definition is f (k)={ tsd|tsd ∈ M, I (root node of tsd)=I (k) }.Under many circumstances, key point may be mapped to an empty set, and f (k)=NULL, this means that a k does not belong to the root node of any one trade mark model.
After index is set up, testing process is divided into following three steps, and detailed process is referring to method 2:
1) the crucial point set of given input picture
the root node index of setting up by previous step, for each k
m∈ M, can obtain the set T=f (k of a TSD
m).
2) for each, on root node index, can find the key point of corresponding TSD, continue to judge each tsd ∈ T whether can in find the tree construction of its coupling, if can find, the kind mark of its corresponding trade mark adds one, that is, kind mark is corresponding to the number of the tree construction corresponding with this trade mark occurring in picture in input;
3) traveling through all k
m∈ X
iafter, if the kind mark of certain trade mark is greater than default threshold value, judge that this trade mark occurs in input picture.
Can above-mentioned detection side's ratio juris be described by following false code form:
Input:
trade mark corresponding to each tsd ∈ M in M;
the crucial point set of input picture;
Output:
the detection score that each trade mark is final
Initialization: D
i=0, i=1 ..., N_L;
fork
mEK
Ido
By root node index f (k
m) calculating TSD set T
for?tSd
n∈T?do
If tsd
nat K
iin found the tree construction of coupling, then
end?if
end?for
end?for
List of references list:
[3]Bay,Herbert,Tinne?Tuytelaars,and?Luc?Van?Gool."Surf:Speeded?up?robust?features."Computer?Vision–ECCV2006.Springer?Berlin?Heidelberg,2006.404-417。
[4]S.Romberg,L.G.Pueyo,R.Lienhart,and?R.Van?Zwol,“Scalable?logo?recognition?in?real-world?images,”in?ICMR,2011
[5]Revaud?J,Douze?M,Schmid?C.“Correlation-based?burstiness?for?logo?retrieval”in?ACM?Multimedia.2012
[6]Sivic,Josef,and?Andrew?Zisserman."Video?Google:A?text?retrieval?approach?to?object?matching?in?videos."Computer?Vision,2003.Proceedings.Ninth?IEEE?International?Conference?on.IEEE,2003.
For fear of the description that makes this instructions, be limited to miscellaneous, in description in this manual, may carry out the processing such as omission, simplification, accommodation to the part ins and outs that can obtain in above-mentioned list of references or other prior art data, this is understandable for a person skilled in the art.At this, above-mentioned list of references is herein incorporated by reference of text.
In sum, those skilled in the art will appreciate that the above embodiment of the present invention can be made various modifications, modification and be replaced, it all falls into the protection scope of the present invention limiting as claims.
Claims (4)
1. a label detection method that is used for detecting sign in image, comprises the following steps:
Step 1, from detected image, extract a plurality of key point k
1～k
i, the sum that wherein I is key point;
Step 2, for described a plurality of key point k
1～k
iin each key point k
m, judge its visual vocabulary sequence number I (k
m) whether with sign model in certain root node k
avisual vocabulary sequence number I (k
a) identical, m ∈ 1～I wherein, wherein, described root node k
abe the root node of the one or more tree constructions in described sign model, wherein, each in described one or more tree constructions is corresponding from different signs;
Step 3, for described each key point k that judgment result is that "Yes"
mother key point of extracting from detected image is traveled through, the visual vocabulary sequence number of finding visual vocabulary sequence number and described one or more tree constructions leaf node is separately identical and meet one or more set of keypoints that tree construction retrains respectively, wherein, each set of keypoints in described one or more set of keypoints is corresponding to a tree construction in described one or more tree constructions; And
The number of the set of keypoints corresponding to each tree construction in described one or more tree constructions that step 4, calculating are found in step 3, if the number corresponding to the described set of keypoints of certain tree construction in described one or more tree constructions is greater than predetermined threshold, judge identify in present described detected image corresponding with described certain tree construction.
2. label detection method as claimed in claim 1, wherein, described tree construction is to comprise a root node and three leaf node k
b, k
c, k
dstructure, described tree construction meets following tree construction simultaneously and retrains:
1）
2) ray P (k
a) P (k
d) at ∠ P (k
b) P (k
a) P (k
c) in;
3）max(S(k
a),...,S(k
d))-min(S(k
a),...<S(k
d))＜θ，
Wherein, P () represents the position of node, and S () represents the yardstick of node, and θ is constant.
3. label detection method as claimed in claim 2, wherein, described sign model comprises separately corresponding to different identification L
ia plurality of sign models, each in described a plurality of sign models produces by following steps:
Step 11, described sign model
be initially sky, input N
iindividual image, described N
ieach image in individual image all comprises certain sign, and has marked the position being identified in described image, wherein, and N
i>2, L
irepresent the sign corresponding with this sign model;
Step 12, from described N
ieach image in individual image extracts key point;
Step 13, for described N
itwo images in individual image, by traveling through all key points, find all set of keypoints tr that can meet described tree construction constraint, and wherein, described set of keypoints tr comprises a root node k who forms described tree construction
awith three leaf node k
b, k
c, k
d,
Step 14, for each the set of keypoints tr in the described set of keypoints finding in described two images, calculate descriptor tsd (tr)={ I (k
a), I (k
b), I (k
c), I (k
d), wherein, I () represents the visual vocabulary sequence number that this key point is corresponding;
If belonging to respectively two set of keypoints of described two images, step 15 there is identical descriptor tsd (tr), described descriptor tsd (tr) is added to sign model subset, and record the corresponding relation of described descriptor tsd (tr) and the sign comprising in this image;
Step 16, for described N
iother every two images in individual image, repeat above-mentioned steps 13 to 15, finally obtain identifying model
, wherein, M1, M2 ...,
for described sign model subset.
4. label detection method as claimed in claim 3, wherein, will identify model
in all descriptors according to it at N
ithe total degree occurring in individual image carries out descending sort, and get front 5000 descriptors and form set, and, for all sign L
i, will identify model
cascade, the sign model after cascade is as the sign model using in step 2, in described cascade process, if according to the corresponding relation of record in step 15, find that same descriptor is corresponding to different identification, from sign model, deletes this descriptor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410125961.5A CN103942276A (en) | 2014-03-31 | 2014-03-31 | Novel logo detection technology |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410125961.5A CN103942276A (en) | 2014-03-31 | 2014-03-31 | Novel logo detection technology |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103942276A true CN103942276A (en) | 2014-07-23 |
Family
ID=51189944
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410125961.5A Pending CN103942276A (en) | 2014-03-31 | 2014-03-31 | Novel logo detection technology |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103942276A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104239531A (en) * | 2014-09-19 | 2014-12-24 | 上海依图网络科技有限公司 | Accurate comparison method based on local visual features |
| 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 |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101464948A (en) * | 2009-01-14 | 2009-06-24 | 北京航空航天大学 | Object identification method for affine constant moment based on key point |
| US20120106847A1 (en) * | 2010-10-29 | 2012-05-03 | Qualcomm Incorporated | Systems and methods to improve feature generation in object recognition |
-
2014
- 2014-03-31 CN CN201410125961.5A patent/CN103942276A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101464948A (en) * | 2009-01-14 | 2009-06-24 | 北京航空航天大学 | Object identification method for affine constant moment based on key point |
| US20120106847A1 (en) * | 2010-10-29 | 2012-05-03 | Qualcomm Incorporated | Systems and methods to improve feature generation in object recognition |
Non-Patent Citations (1)
| Title |
|---|
| CHENGDE WAN .ETC: ""Tree-based Shape Descriptor for scalable logo detection"", 《2013 VISUAL COMMUNICATIONS AND IMAGE PROCESSING（VCIP）》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104239531A (en) * | 2014-09-19 | 2014-12-24 | 上海依图网络科技有限公司 | Accurate comparison method based on local visual features |
| CN104239531B (en) * | 2014-09-19 | 2017-09-26 | 上海依图网络科技有限公司 | A kind of precise alignment method based on local visual feature |
| 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 |
| 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 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wei et al. | Toward automatic building footprint delineation from aerial images using CNN and regularization | |
| Workman et al. | Wide-area image geolocalization with aerial reference imagery | |
| Li et al. | Hierarchical line matching based on line–junction–line structure descriptor and local homography estimation | |
| CN104766084B (en) | A kind of nearly copy image detection method of multiple target matching | |
| CN109598290A (en) | A kind of image small target detecting method combined based on hierarchical detection | |
| Tau et al. | Dense correspondences across scenes and scales | |
| CN108921939A (en) | A kind of method for reconstructing three-dimensional scene based on picture | |
| CN104156965B (en) | A kind of automatic quick joining method of Mine Monitoring image | |
| CN107392215A (en) | A kind of multigraph detection method based on SIFT algorithms | |
| Al-Shahri et al. | Line matching in wide-baseline stereo: A top-down approach | |
| CN104182973A (en) | Image copying and pasting detection method based on circular description operator CSIFT (Colored scale invariant feature transform) | |
| CN111611861B (en) | Image change detection method based on multi-scale feature association | |
| CN105224937A (en) | Based on the semantic color pedestrian of the fine granularity heavily recognition methods of human part position constraint | |
| CN108154066B (en) | Three-dimensional target identification method based on curvature characteristic recurrent neural network | |
| CN105551022A (en) | Image error matching detection method based on shape interaction matrix | |
| CN104281572A (en) | Target matching method and system based on mutual information | |
| Li et al. | Line segment matching: A benchmark | |
| Pang et al. | Exploiting local linear geometric structure for identifying correct matches | |
| Guo et al. | Exploring GIS knowledge to improve building extraction and change detection from VHR imagery in urban areas | |
| CN103823887A (en) | Based on low-order overall situation geometry consistency check error match detection method | |
| CN105809678A (en) | Global matching method for line segment characteristics between two views under short baseline condition | |
| CN104835174A (en) | Robustness model fitting method based on supermap mode search | |
| Cao et al. | Parallel K nearest neighbor matching for 3D reconstruction | |
| CN103942276A (en) | Novel logo detection technology | |
| Bosilj et al. | Beyond MSER: Maximally stable regions using tree of shapes |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20140723 |
|
| WD01 | Invention patent application deemed withdrawn after publication |