CN109460773A - A kind of cross-domain image sparse matching process based on depth convolutional network - Google Patents
A kind of cross-domain image sparse matching process based on depth convolutional network Download PDFInfo
- Publication number
- CN109460773A CN109460773A CN201810846985.8A CN201810846985A CN109460773A CN 109460773 A CN109460773 A CN 109460773A CN 201810846985 A CN201810846985 A CN 201810846985A CN 109460773 A CN109460773 A CN 109460773A
- Authority
- CN
- China
- Prior art keywords
- matching
- image
- tensor
- arest neighbors
- sub
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/757—Matching configurations of points or features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of cross-domain image sparse matching process based on depth convolutional network.It can be suitable for the image that external appearance characteristic changes greatly using the present invention, and robustness is preferable.The present invention utilizes the convolutional layer, pond layer of different levels in depth convolutional network or the tensor information of active coating, inverted pyramid formula gradually accurate Feature Points Matching is successively up carried out from bottom grade, it can be suitable for the image that external appearance characteristic changes greatly, and robustness is preferable.The not strong arest neighbors matching pair of semantic information is rejected to screening in addition, matching using tensor response to arest neighbors, so that the arest neighbors screened is matched to semantic dependency.
Description
Technical field
The present invention relates to technical field of computer vision, and in particular to a kind of cross-domain image based on depth convolutional network is dilute
Dredge matching process.
Background technique
Images match is not only a key problem in computer vision, and many graphics apply such as anamorphose
Basis.
In traditional algorithm, the process of searching images match point is divided into two steps, extracts the feature in each image first
Characteristic point, is then compared and is matched by point.It proposes many effective methods both at home and abroad to this, is broadly divided into angle point spy
The three categories such as sign detection, speckled characteristic detection and provincial characteristics detection.SIFT (the Scale- put forward by Lowe et al. in 1999
Invariant feature transform) algorithm, the point of interest based on image local appearance and with the size and rotation of image
It is unrelated, there is invariance (Lowe, David the G. " Object of translation, rotation, scale and part visual angle change
recognition from local scale-invariant features."Computer vision,1999.The
proceedings of the seventh IEEE international conference on.Vol.2.Ieee,
1999.)；SURF (the Speeded up robust features) algorithm put forward by Herbert for 2006 is SIFT algorithm
Improvement, greatly reduce time complexity (Bay, Herbert, Tinne Tuytelaars, and the Luc Van of algorithm
Gool."Surf:Speeded up robust features."European conference on computer
vision.Springer,Berlin,Heidelberg,2006.)。
However the feature for being based only on image itself, for the image that external appearance characteristic changes greatly, traditional algorithm can not be obtained
To preferable matching result.Further, since lacking semantic information, obtained match point does not have semantic dependency, Wu Fawei yet
More significant follow-up works provide guidance.Christopher B.Choy et al. proposes a kind of based on depth network
Learning method learns the matching between image, however their data for having the method for supervision to be only applicable to corresponding data collection, Shandong
Stick it is poor (Chandraker M, Choy C B, Savarese S.UNIVERSAL CORRESPONDENCE NETWORK:,
US20170124711[P].2017.)。
Summary of the invention
In view of this, the present invention provides a kind of cross-domain image sparse matching process based on depth convolutional network, it can
Suitable for the image that external appearance characteristic changes greatly, and robustness is preferable.
Cross-domain image sparse matching process based on depth convolutional network of the invention, includes the following steps:
Step 1, construct depth convolutional network, and it be trained, obtain capable of extracting characteristics of image, it is trained
Depth convolutional network；
Step 2, two images A, B to be matched are distinguished into the trained depth convolutional network of input step 1, extracts depth
The output of convolutional layer, pond layer in convolutional network in each level or active coating；Wherein, level extracts wherein one layer defeated
Out, each level can extract the output (such as extracting active coating) of same layer, can also extract the output of layer not of the same race；
Step 3, for the output of each layer of extraction, the search subregion of this layer is set separately；Since the bottom, under
It is supreme to carry out arest neighbors matching respectively in the corresponding search subregion of each layer；Wherein, arest neighbors matching process is as follows:
For l layers of n-th of search subregion, to each of l layers of n-th of search subregion of image A
Sub- tensorWith the sub- tensor of same position in l layers of image BCentered on matching area in find withDistance
Nearest sub- tensor, wherein matching area is less than search subregion；Likewise, to l layers of n-th of search sub-district of image B
The sub- tensor in each of domainWith the sub- tensor of same position in l layers of image ACentered on matching area in
Find withThe nearest sub- tensor of distance；If A, in n-th of search subregion of l layers of B there are two sub- tensors each other away from
From nearest, then the two sub- tensors are referred to as arest neighbors matching pair；
Wherein, l layers of n-th of search subregion is that n-th of matching of next layer (l+1 layers) is reflected in l layer
It penetrates；The search subregion of bottom l=L (total number that L is step 1 extract layer) is bottom whole region；
And so on, obtain the arest neighbors matching pair of top l=1；
Step 4, the matching using the arest neighbors matching of top l=1 to image A, B is carried out.
Further, in the step 1, the depth convolutional network is the network extract with characteristics of image,
Such as image classification network.
Further, in the step 1, network training is carried out using public data collection such as ImageNet or is directlyed adopt
The existing disclosed good depth convolutional network of pre-training.
Further, in the step 2, the convolutional layer of preceding 4 or preceding 5 levels, pond layer or active coating are only extracted
Output executes subsequent step.
Further, in the step 3, the size of matching area is determined according to detection range and network structure.
Further, in the step 3, for l layers of n-th of search subregion ln, arest neighbors matching process is as follows:
The sub- tensor of each of search subregion ln to image AIn the sub- tensor of the same position of image B
Centered on matching area in find withThe nearest sub- tensor of distanceThen sub- tensor is calculatedIn the identical of image A
The sub- tensor of positionCentered on matching area inThe nearest sub- tensor of distance whether beIf so, thinkingWithFor the matching pair of a pair of of arest neighbors.
Further, in the step 3, for each layer, the tensor response of this layer of arest neighbors matching pair is calculated, chooses and rings
The final arest neighbors matching pair that this layer is paired into more than or equal to the arest neighbors of given threshold should be worth, using final nearest
Neighbour's matching is to execution subsequent step.
Further, it in the step 3, if source images and target image style gap are larger, is carrying out apart from meter
When calculation, the corresponding search subregion of image A, B is subjected to style conversion, is converted to unified public style, then carry out again away from
From calculating, arest neighbors matching pair is obtained.
Further, in the step 4, before the matching for carrying out image A, B, first respectively to the top l of step 3 acquisition
=1 arest neighbors matching is clustered to using unsupervised clustering method, then using each cluster centre to progress image A, B
Matching.
Further, the unsupervised clustering method is K-Means clustering procedure, DBSCAN clustering procedure or Mean-Shift
Clustering procedure.
The utility model has the advantages that
The present invention is using the convolutional layer, pond layer of different levels in depth convolutional network or the tensor information of active coating,
Successively up carry out gradually accurate Feature Points Matching from bottom grade to pyramid, can be suitable for external appearance characteristic variation compared with
Big image, and robustness is preferable.
Before search subregion carries out arest neighbors matching, search subregion is subjected to style unification, eliminates image in appearance
The influence of the stylistic differences such as the color in feature.
Arest neighbors is matched using tensor response and rejects the not strong arest neighbors matching pair of semantic information to screening,
So that the arest neighbors screened is matched to semantic dependency.
It is matched using arest neighbors of the unsupervised clustering method to top active coating to clustering, utilizes cluster centre
To images match is carried out, matching efficiency is improved；And it is possible to flexibly change match point quantity according to application demand.
Detailed description of the invention
Fig. 1 is of the invention by bottom active coating successively up gradually accurate matched exemplary diagram.
Fig. 2 is the exemplary diagram in the present invention by l layers of matching to the corresponding region for expanding to l-1 layers.
Fig. 3 is the exemplary diagram that style is converted in the present invention.
Fig. 4 is the arithmetic result of an example of the invention.
Specific embodiment
The present invention will now be described in detail with reference to the accompanying drawings and examples.
The present invention provides a kind of cross-domain image sparse matching process based on depth convolutional network carries out image A, B
Matching is exported using each level of depth convolutional network, gradually accurate matching is realized from the bottom to top, as shown in Figure 1, applicable
In the image that external appearance characteristic changes greatly, and robustness is preferable.Specifically comprise the following steps:
Step 1, construct depth convolutional network, and it be trained, obtain capable of extracting characteristics of image, it is trained
Depth convolutional network.
Wherein it is possible to select the depth convolutional network for inherently having image characteristics extraction function, such as image classification network,
To reduce training burden, network complexity is reduced；In training, the large datas such as ImageNet collection can use to the net of building
Network carries out pre-training；Or the good image classification network of existing disclosed pre-training, such as VGG19 also can be directly used.
Step 2, using two images A, B to be matched as the input of network, the trained depth of input step 1 is rolled up respectively
Product network, the output of the convolutional layer of each level of registered depth convolutional network, pond layer or active coating as a result, be denoted as respectivelyL is total number of plies of record；Wherein, a level extracts wherein one layer of output, each level
The output (such as extracting active coating) that same layer can be extracted, can also extract the output of layer not of the same race；Below to mention
It takes and is illustrated for the active coating of each level.
Wherein, image A, B are scaled identical size using the image procossing library in language respectively, network is guaranteed with this
Each layer output is of similar shape.
To improve computational efficiency, the output of top n level can be only extracted as a result, can extract swashing for preceding 4 or 5 levels
The output of layer living is as a result, it is possible to which calculation amount is small in the case where realizing guarantee matching precision.
Step 3, the search subregion that this layer is respectively set in each active coating divides from the bottom to top since the active coating of bottom
Arest neighbors matching is not carried out in each layer, obtains the arest neighbors matching pair of each layer.
Wherein, the output of each active coating is all the tensor that a shape is (length × wide × channel), swashs from the bottom
Layer living starts, in the search subregion of this layer of (long × wide) plane sets of each active coating；For bottom active coating l=
L, search subregion are bottom active coating (long × wide) plane.
Closest matching is carried out in bottom active coating l=L to image A and image B first: a) constructing matching area,
Middle matching area is less than search subregion, and the size of matching area is flexibly determined according to detection range and network structure；Each activation
The size of the matching area of layer can be the same or different, but the matching area size in same layer needs unanimously；B) to image A
The bottom active coating outputTo the sub- tensor on its each position of tensor planeIn the bottom active coating of image B
OutputIn, with the sub- tensor of same positionCentered on matching areaIt is interior searching withThe nearest son of Euclidean distance
Amount；Likewise, to the bottom active coating of image BTo the sub- tensor on its each position of tensor plane?In with
The sub- tensor of same positionCentered on matching areaIt is interior searching withThe nearest sub- tensor of Euclidean distance；If
It is middle that there are one pair of sub- tensors each other all for apart from recently, then the two sub- tensors are referred to as a pair of of arest neighbors matching pair.
Above-mentioned is that the method in search subregion using traversal realizes arest neighbors matching, further, it is also possible to using following
Method carries out more efficiently arest neighbors matching:
For n-th of search subregion ln of active coating l, arest neighbors matching process is as follows:
The sub- tensor of each of search subregion ln to image AIn the sub- tensor of the same position of image B
Centered on matching area in find withThe nearest sub- tensor of distanceThen sub- tensor is calculatedIn the identical of image A
The sub- tensor of positionCentered on matching area inThe nearest sub- tensor of distance whether beIf so, thinkingWithFor the matching pair of a pair of of arest neighbors.
When carrying out the calculating of sub- tensor Euclidean distance, it is contemplated that the differences such as the color of image on external appearance characteristic, if figure
As A and image B style gap are larger, then can preferably first corresponding to image A, B search subregion carries out style conversion, unite
One becomes public style, then carries out Euclidean distance calculating again, obtains arest neighbors matching pair.
Wherein, style conversion can be obtained by following formula:
Wherein:ForIn PlValue in region；μA, μB, σA, σBRespectively represent the eigenmatrix in specified region and interior
Hold matrix；μm, σmThe public characteristic matrix for representing corresponding region, can be calculated by following formula:
Then the distance between tensor formula are as follows:
Further, it is also possible to screen out a part of tensor response lower than setting using the response value tag of active coating output tensor
The arest neighbors matching pair of threshold value is determined, so that the arest neighbors matching that the active coating is retained is to being rich in semantic information.
Wherein, the response of each sub- tensor is obtained by the following formula:
Wherein,Respectively p point, i point existThe output of layer, " | | | | " it is absolute value sign.
In all arest neighbors matching for obtaining bottom active coating to rear, according to network structure, arest neighbors is matched into mapping
It is mapped in one layer of active coating；Arest neighbors matching is search of L-1 layers of active coating to the mapping area in L-1 layers of active coating
Region；Then same method is used, arest neighbors matching is carried out respectively in each search subregion in L-1 layers of active coating, obtains
To the arest neighbors matching pair of L-1 layers of active coating；And so on, finally obtain the arest neighbors matching pair of top l=1 active coating.
Step 4, the matching using the arest neighbors matching of top active coating to image A, B is carried out.
Wherein, if the arest neighbors matching of top active coating is more to number, it can use unsupervised clustering method
The arest neighbors of top active coating is matched to clustering, respectively by matching all in image A, image B to be divided into 5,10,
15,20 or other are several classes of, obtain all kinds of cluster centres pair.Then the matching using each cluster centre to image A, B is carried out.
Wherein it is possible to which K-Means clustering procedure, DBSCAN clustering procedure or Mean-Shift clustering procedure etc. are clustered.
It is described in detail below with reference to a specific example.
Step 1: being separately input into after image A, B are scaled 224 × 224 respectively using the image procossing library in language
Trained VGG19 image classification network, extract network before 5 levels active coating output result:
Step 2: fromStart to carry out matched search.That the output result of layer 5 is 14 × 14 × 512
Amount, using the first peacekeeping second dimension as search plane, only one search subregion of layer 5, i.e. layer 5 searches for plane sheet
Body.For each sub- tensor of image A, Europe is found in the matching area in image B centered on the sub- tensor of same position
A nearest sub- tensor of formula distance；Equally, for each sub- tensor of image B, with the sub- tensor of same position in image A
Centered on matching area in find a nearest sub- tensor of Euclidean distance.One pair of sub- tensor is all most to connect each other if it exists
Closely, then they are a pair of of arest neighbors matchings pair.
During search, it is contemplated that the external appearance characteristics such as color of original image differ greatly, before distance calculates, to searching
Large rope region carries out style conversion.
Search for being averaged using further feature proposed in the article of the style convert reference Johnson in 2016 of subregion
The style conversion that value and variance are carried out as normalized parameter.The object of style conversion, is the corresponding search sub-district of image A, B
DomainIt is changed into public style using following formula:
Wherein:
Style switch target is a specific search subregion, therefore in search subregionIn, for two
The tensor p of specific position, the distance between q definition, also with the connection of search subregion together, formula is defined as:
The instance graph of style conversion can be found in Fig. 3.Search subregion after style conversion scans for, and obtains arest neighbors
Matching pair.
Then, by construction receptance function, the responsiveness of feature vector on each position is obtained, it can be by the position
Semantic abundant information be depicted come.Formula is as follows:
After the response for obtaining all arest neighbors matchings pair, the arest neighbors matching pair lower than given threshold is screened out, this
A little points mean do not have stronger semantic information.The arest neighbors matching pair positioned at tensor edge is screened out simultaneously, these points are to difficulty
To find corresponding region of search in upper layer activation layer.
Step 3: as shown in Fig. 2, low layer obtain arest neighbors matching pair position after, can use network structure, will be low
The matching of layer is extended in upper space coordinate, obtains the search subregion pair of upper space.In each search subregion
Centering repeats step 2, generates the arest neighbors matching pair of this layer.
Step 3 is repeated, layer extension is continued up, finally obtains the arest neighbors matching pair of the original image of top.
Step 4: tending to find more than a arest neighbors matchings pair up to a hundred in the original image of top.In order to allow arest neighbors
Pairing energy expresses more different semantic informations as far as possible, using K-Means clustering algorithm, with x, y of arest neighbors matching pair
Coordinate is clustered as the feature of point, obtains 5,10,30 etc. multi-class, takes each classification closest to the point at center as such
It Dai Biao not put and stay in the list of arest neighbors matching pair.Obtain the experimental result of Fig. 4.
Since then, the process by finding sparse match point in two images is just completed.
In conclusion the above is merely preferred embodiments of the present invention, being not intended to limit the scope of the present invention.
All within the spirits and principles of the present invention, any modification, equivalent replacement, improvement and so on should be included in of the invention
Within protection scope.
Claims (10)
1. a kind of cross-domain image sparse matching process based on depth convolutional network, which comprises the steps of:
Step 1, depth convolutional network is constructed, and it is trained, obtains capable of extracting characteristics of image, trained depth
Convolutional network；
Step 2, two images A, B to be matched are distinguished into the trained depth convolutional network of input step 1, extracts depth convolution
The output of convolutional layer, pond layer in network in each level or active coating；
Step 3, for the output of each layer of extraction, the search subregion of this layer is set separately；Since the bottom, from the bottom to top
Arest neighbors matching is carried out respectively in the corresponding search subregion of each layer；Wherein, arest neighbors matching process is as follows:
For l layers of n-th of search subregion, to each of l layers n-th of search subregion of image A
AmountWith the sub- tensor of same position in l layers of image BCentered on matching area in find withDistance is most
Close sub- tensor, wherein matching area is less than search subregion；Likewise, to l layers of n-th of search subregion of image B
Each of sub- tensorWith the sub- tensor of same position in l layers of image ACentered on matching area in seek
Look forThe nearest sub- tensor of distance；If A, there are two sub- tensors distances each other in n-th of search subregion of l layers of B
Recently, then the two sub- tensors are referred to as arest neighbors matching pair；
Wherein, n-th of l layers search subregion is that n-th of l+1 layer is matched in l layers of mapping；The search of the bottom
Subregion is bottom whole region；
And so on, obtain the arest neighbors matching pair of top；
Step 4, the matching using the arest neighbors matching of top to image A, B is carried out.
2. the cross-domain image sparse matching process based on depth convolutional network as described in claim 1, which is characterized in that described
In step 1, the depth convolutional network is image classification network.
3. the cross-domain image sparse matching process as stated in claim 1 or 2 based on depth convolutional network, feature exist
In carrying out network training using public data collection or directly adopt the existing disclosed good depth of pre-training in the step 1
Spend convolutional network.
4. the cross-domain image sparse matching process based on depth convolutional network as described in claim 1, which is characterized in that described
In step 2, the output execution subsequent step of the convolutional layer of preceding 4 or preceding 5 levels, pond layer or active coating is only extracted.
5. the cross-domain image sparse matching process based on depth convolutional network as described in claim 1, which is characterized in that described
In step 3, the size of matching area is determined according to detection range and network structure.
6. the cross-domain image sparse matching process based on depth convolutional network as described in claim 1, which is characterized in that described
In step 3, for l layers of n-th of search subregion ln, arest neighbors matching process is as follows:
To the search subregion l of image AnEach of sub- tensorIn the sub- tensor of the same position of image BFor in
Found in the matching area of the heart withThe nearest sub- tensor of distanceThen sub- tensor is calculatedIn the same position of image A
Sub- tensorCentered on matching area inThe nearest sub- tensor of distance whether beIf so, thinkingWithFor the matching pair of a pair of of arest neighbors.
7. the cross-domain image sparse matching process based on depth convolutional network as described in claim 1, which is characterized in that described
In step 3, for each layer, the tensor response of each arest neighbors matching pair of this layer is calculated, selection response, which is greater than or equal to, to be set
The arest neighbors for determining threshold value is paired into the final arest neighbors matching pair of this layer, subsequent to executing using final arest neighbors matching
Step.
8. the cross-domain image sparse matching process based on depth convolutional network as described in claim 1, which is characterized in that described
In step 3, when carrying out distance calculating, the corresponding search subregion of image A, B is subjected to style conversion, is converted to unified public affairs
Then style altogether carries out obtaining arest neighbors matching pair apart from calculating again.
9. the cross-domain image sparse matching process based on depth convolutional network as described in claim 1, which is characterized in that described
In step 4, before the matching for carrying out image A, B, the arest neighbors of the top first obtained respectively to step 3 is matched to using no prison
The clustering method superintended and directed is clustered, then the matching using each cluster centre to image A, B is carried out.
10. the cross-domain image sparse matching process based on depth convolutional network as claimed in claim 9, which is characterized in that institute
Stating unsupervised clustering method is K-Means clustering procedure, DBSCAN clustering procedure or Mean-Shift clustering procedure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810846985.8A CN109460773A (en) | 2018-07-27 | 2018-07-27 | A kind of cross-domain image sparse matching process based on depth convolutional network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810846985.8A CN109460773A (en) | 2018-07-27 | 2018-07-27 | A kind of cross-domain image sparse matching process based on depth convolutional network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109460773A true CN109460773A (en) | 2019-03-12 |
Family
ID=65606321
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810846985.8A Pending CN109460773A (en) | 2018-07-27 | 2018-07-27 | A kind of cross-domain image sparse matching process based on depth convolutional network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109460773A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909778A (en) * | 2019-11-12 | 2020-03-24 | 北京航空航天大学 | Image semantic feature matching method based on geometric consistency |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103456022A (en) * | 2013-09-24 | 2013-12-18 | 中国科学院自动化研究所 | High-resolution remote sensing image feature matching method |
| CN106203533A (en) * | 2016-07-26 | 2016-12-07 | 厦门大学 | The degree of depth based on combined training study face verification method |
| CN106227851A (en) * | 2016-07-29 | 2016-12-14 | 汤平 | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end |
| US20170083792A1 (en) * | 2015-09-22 | 2017-03-23 | Xerox Corporation | Similarity-based detection of prominent objects using deep cnn pooling layers as features |
-
2018
- 2018-07-27 CN CN201810846985.8A patent/CN109460773A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103456022A (en) * | 2013-09-24 | 2013-12-18 | 中国科学院自动化研究所 | High-resolution remote sensing image feature matching method |
| US20170083792A1 (en) * | 2015-09-22 | 2017-03-23 | Xerox Corporation | Similarity-based detection of prominent objects using deep cnn pooling layers as features |
| CN106203533A (en) * | 2016-07-26 | 2016-12-07 | 厦门大学 | The degree of depth based on combined training study face verification method |
| CN106227851A (en) * | 2016-07-29 | 2016-12-14 | 汤平 | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end |
Non-Patent Citations (1)
| Title |
|---|
| 周文罡等: "图像检索技术研究进展", 《南京信息工程大学学报》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909778A (en) * | 2019-11-12 | 2020-03-24 | 北京航空航天大学 | Image semantic feature matching method based on geometric consistency |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108764065B (en) | Pedestrian re-recognition feature fusion aided learning method | |
| Lin et al. | RSCM: Region selection and concurrency model for multi-class weather recognition | |
| CN106649487B (en) | Image retrieval method based on interest target | |
| CN105869173B (en) | A kind of stereoscopic vision conspicuousness detection method | |
| Peng et al. | Automatic image segmentation by dynamic region merging | |
| CN102314614B (en) | Image semantics classification method based on class-shared multiple kernel learning (MKL) | |
| CN108537121B (en) | Self-adaptive remote sensing scene classification method based on meteorological environment parameter and image information fusion | |
| CN107992850B (en) | Outdoor scene three-dimensional color point cloud classification method | |
| CN111339343A (en) | Image retrieval method, device, storage medium and equipment | |
| CN110516533B (en) | Pedestrian re-identification method based on depth measurement | |
| Li et al. | Real-time crop recognition in transplanted fields with prominent weed growth: a visual-attention-based approach | |
| CN109840518B (en) | Visual tracking method combining classification and domain adaptation | |
| Quek et al. | Structural image classification with graph neural networks | |
| CN109461115A (en) | A kind of automatic Image Registration Method based on depth convolutional network | |
| Zhang et al. | Category modeling from just a single labeling: Use depth information to guide the learning of 2d models | |
| Karaman et al. | Multi-layer local graph words for object recognition | |
| CN109460773A (en) | A kind of cross-domain image sparse matching process based on depth convolutional network | |
| JP6486084B2 (en) | Image processing method, image processing apparatus, and program | |
| Lu et al. | Closing the loop for edge detection and object proposals | |
| JP2017117025A (en) | Pattern identification method, device thereof, and program thereof | |
| Zhan et al. | Gaussian mixture model on tensor field for visual tracking | |
| Cui et al. | Weighted particle swarm clustering algorithm for self-organizing maps | |
| Xia et al. | Lazy texture selection based on active learning | |
| CN114494819A (en) | Anti-interference infrared target identification method based on dynamic Bayesian network | |
| CN108509838B (en) | Method for analyzing group dressing under joint condition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190312 |
|
| RJ01 | Rejection of invention patent application after publication |