CN106886573A - A kind of image search method and device - Google Patents
A kind of image search method and device Download PDFInfo
- Publication number
- CN106886573A CN106886573A CN201710038542.1A CN201710038542A CN106886573A CN 106886573 A CN106886573 A CN 106886573A CN 201710038542 A CN201710038542 A CN 201710038542A CN 106886573 A CN106886573 A CN 106886573A
- Authority
- CN
- China
- Prior art keywords
- image
- subset
- feature
- feature subset
- retrieved
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Library & Information Science (AREA)
- Databases & Information Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Processing Or Creating Images (AREA)
Abstract
The present invention discloses a kind of image search method and device, wherein, image search method comprises the following steps：Build multitask depth network structure model；Set up target image characteristics storehouse；It is input into the character subset of image to be retrieved；Calculate the character subset and every similarity distance of characteristics of image in characteristics of image storehouse of image to be retrieved；According to the image obtained after distance ordered arrangement from small to large with image distance to be retrieved minimum.The present invention is by setting up depth network structure model, the feature of multi-task learning target image completes image retrieval, the correlation of task can be made full use of to combine the precision for improving multiple attributive classifications, the minutia of the contact and image to image each attribute can be learnt, so and then during image retrieval, can preferably overcome weather, environment, illumination etc. influence.
Description
Technical field
The present invention relates to image processing field, and in particular to a kind of image search method and device.
Background technology
With the fast development of computer and internet, multimedia image information is promptly produced and passed just on network
Broadcast, so that the vision of abundant people, in face of various image informations, excavates the required image of people and have become one
The problem for receiving much concern, so setting up associated picture search method also turns into the focus of research and engineering practice.
The image retrieval mode of currently available technology typically directly applies to certain list using the deep learning of neutral net
One attribute completes the image retrieval of single task learning characteristic, because single task learning characteristic often only considers task information in itself
Amount, does not account for contacting between other tasks, therefore learning ability receives certain limitation, it is impossible to play depth god
Through the advantage function of network, therefore cannot be input into as supervisory signals by the use of multiple attributes, because without each attribute of calligraphy learning to image
Contact and image minutia, therefore retrieve image during, it is impossible to overcome weather, environment, illumination etc. influence, because
And the accurate feature of target image cannot be extracted, causing the retrieval precision of image reduces, and is retrieved the big image of similarity simultaneously
The target of undesired, image retrieval index is poor.
The content of the invention
Therefore, the embodiment of the present invention technical problem to be solved is that image retrieval mode of the prior art passes through certain
Single attribute complete single task learning characteristic image retrieval, it is impossible to learn each attribute of image contact and image it is thin
Section feature, during retrieval, it is impossible to overcome weather, environment, the influence of illumination, causing the retrieval progress of image reduces, schemes
As retrieval effectiveness is poor.
Therefore, the embodiment of the invention provides following technical scheme：
The present invention provides a kind of image search method, comprises the following steps：
Build multitask depth network structure model；
Set up target image characteristics storehouse；
It is input into the character subset of image to be retrieved；
Calculate the character subset of the image to be retrieved and every similarity of characteristics of image in described image feature database away from
From；
According to the image obtained after distance ordered arrangement from small to large with the image distance minimum to be retrieved.
Alternatively, described image search method, the structure multitask depth network structure model, including：
Determine the attribute classification and training image subset corresponding with the attribute classification of target image；
The multiple attribute task datas of input；
Share the related task data resource；
According to the multiple attribute task, corresponding loss function is determined；
According to 256 dimension depth characteristic descriptions, the fisrt feature subset of target image is obtained.
Alternatively, described image search method, it is described according to the multiple attribute task, determine corresponding loss function
Step；Including：
Formula Lt=- (1-t) * ㏒ (1-Pt)-t* ㏒ (pt)；
Wherein t represents whether the label that a certain learning tasks learning is arrived is equal with true tag, equal then t=1, not phase
Deng t=0；Pt represents that study is output as the probability of the true tag.
Alternatively, described image search method, it is described to set up target image characteristics storehouse, including：
Extract the second feature subset that sift describes image；
Extract the third feature subset that lbp describes image；
According to the second feature subset, third feature subset, the fourth feature subset of traditional description image is obtained；
Obtain the fifth feature subset of individual target image.
Alternatively, described image search method, the fourth feature subset includes second feature subset and third feature
Collection, fifth feature subset includes fisrt feature subset, second feature subset and third feature subset.
Alternatively, described image search method, the target feature library is comprising multiple to be retrieved with the described 5th
The target image of character subset.
Alternatively, described image search method, the character subset of the input image to be retrieved builds many using described
The step of business depth network structure model, extracts image.
Alternatively, described image search method, calculates the character subset and described image feature database of the image to be retrieved
In every similarity distance of image, including：
Formula Dis=0.75*d (d)+0.125*d (s)+0.125*d (l)；
Wherein, d (d) represents the character subset of the image to be retrieved and the distance of the fisrt feature subset；D (s) tables
Show the character subset of the image to be retrieved and the distance of the second feature subset；
D (l) represents the character subset of the image to be retrieved and the distance of the third feature subset；
Dis represents final similarity distance.
The present invention provides a kind of image retrieving apparatus, including such as lower unit：
Construction unit, for building multitask depth network structure model；
Unit is set up, for setting up target image characteristics storehouse；
Input block, the character subset for being input into image to be retrieved；
Computing unit, for calculating the character subset of the image to be retrieved and every image spy in described image feature database
The similarity distance levied；
Acquiring unit, it is minimum with the image distance to be retrieved according to being obtained after distance ordered arrangement from small to large
Image.
Alternatively, described image retrieval device, it is characterised in that the construction unit, including：
First determining module, attribute classification and training figure corresponding with the attribute classification for determining target image
As subset；
Input module, for being input into multiple attribute task datas；
Sharing module, for sharing the related task data resource；
Second determining module, for according to the multiple attribute task, determining corresponding loss function；
First acquisition module, for according to 256 dimension depth characteristic descriptions, obtaining the fisrt feature subset of target image.
Alternatively, described image retrieval device, second determining module；Including：
Formula Lt=- (1-t) * ㏒ (1-Pt)-t* ㏒ (pt)；
Wherein t represents whether the label that a certain learning tasks learning is arrived is equal with true tag, equal then t=1, not phase
Deng t=0；Pt represents that study is output as the probability of the true tag.
Alternatively, described image retrieval device, described to set up unit, including：
First extraction module, the second feature subset of image is described for extracting sift；
Second extraction module, the third feature subset of image is described for extracting lbp；
Second acquisition module, according to the second feature subset, third feature subset, obtains the 4th of traditional description image
Character subset；
3rd acquisition module, the fifth feature subset for obtaining individual target image.
Alternatively, described image retrieval device, the fourth feature subset includes second feature subset and third feature
Collection, fifth feature subset includes fisrt feature subset, second feature subset and third feature subset.
Alternatively, described image retrieval device, the target feature library set up in unit is to be retrieved comprising multiple
Target image with the fifth feature subset.
Alternatively, described image retrieval device, image is extracted in the input block using construction unit.
Alternatively, described image retrieval device, calculates the character subset and described image feature database of the image to be retrieved
In every similarity distance of image, including：
Formula Dis=0.75*d (d)+0.125*d (s)+0.125*d (l)；
Wherein, d (d) represents the character subset of the image to be retrieved and the distance of the fisrt feature subset；D (s) tables
Show the character subset of the image to be retrieved and the distance of the second feature subset；
D (l) represents the character subset of the image to be retrieved and the distance of the third feature subset；
Dis represents final similarity distance.
Embodiment of the present invention technical scheme, has the following advantages that：
The present invention discloses a kind of image search method and device, wherein, image search method comprises the following steps：Build
Multitask depth network structure model；Set up target image characteristics storehouse；It is input into the character subset of image to be retrieved；Calculate to be retrieved
The character subset of image and every similarity distance of characteristics of image in characteristics of image storehouse；Arranged according to distance sequence from small to large
The image minimum with image distance to be retrieved is obtained after row.The present invention is by setting up depth network structure model, multi-task learning
The feature of target image completes image retrieval, and the correlation of task can be made full use of to combine the essence for improving multiple attributive classifications
Degree, can learn the minutia of the contact and image to image each attribute, so and then during image retrieval,
Weather, environment, illumination etc. can preferably be overcome to be influenceed.
Brief description of the drawings
In order to illustrate more clearly of the specific embodiment of the invention or technical scheme of the prior art, below will be to specific
The accompanying drawing to be used needed for implementation method or description of the prior art is briefly described, it should be apparent that, in describing below
Accompanying drawing is some embodiments of the present invention, for those of ordinary skill in the art, before creative work is not paid
Put, other accompanying drawings can also be obtained according to these accompanying drawings.
Fig. 1 is the flow chart of image search method in the embodiment of the present invention 1；
Fig. 2 is the flow chart of structure model in image search method in the embodiment of the present invention 1；
Fig. 3 is multitask depth network architecture block diagram in image search method in the embodiment of the present invention 1；
Fig. 4 is the flow chart for setting up target image characteristics storehouse in the embodiment of the present invention 1 in image search method；
Fig. 5 is the structured flowchart of image retrieving apparatus in the embodiment of the present invention 2；
Fig. 6 is the structured flowchart of construction unit in image retrieving apparatus in the embodiment of the present invention 2；
Fig. 7 is the structured flowchart for setting up unit in the embodiment of the present invention 2 in image retrieving apparatus.
Specific embodiment
The technical scheme of the embodiment of the present invention is clearly and completely described below in conjunction with accompanying drawing, it is clear that described
Embodiment be a part of embodiment of the invention, rather than whole embodiments.Based on the embodiment in the present invention, this area is general
The every other embodiment that logical technical staff is obtained under the premise of creative work is not made, belongs to present invention protection
Scope.
, it is necessary to explanation in the description of the embodiment of the present invention, term " " center ", " on ", D score, "left", "right",
The orientation or position relationship of the instruction such as " vertical ", " level ", " interior ", " outward " be based on orientation shown in the drawings or position relationship,
It is for only for ease of the description embodiment of the present invention and simplifies description, must has rather than the device or element for indicating or imply meaning
Have specific orientation, with specific azimuth configuration and operation, therefore be not considered as limiting the invention.Additionally, term " the
One ", " second ", " the 3rd " are only used for describing purpose, and it is not intended that indicating or implying relative importance.
, it is necessary to explanation, unless otherwise clearly defined and limited, term " is pacified in the description of the embodiment of the present invention
Dress ", " connected ", " connection " should be interpreted broadly, for example, it may be fixedly connected, or be detachably connected, or integratedly
Connection；Can mechanically connect, or electrically connect；Can be joined directly together, it is also possible to be indirectly connected to by intermediary,
Two connections of element internal are can also be, can be wireless connection, or wired connection.For the common skill of this area
For art personnel, above-mentioned term concrete meaning in the present invention can be understood with concrete condition.
As long as additionally, technical characteristic involved in invention described below different embodiments non-structure each other
Can just be combined with each other into conflict.
Embodiment 1
The present embodiment provides a kind of image search method, as shown in figure 1, comprising the following steps：
S1, structure multitask depth network structure model；Multitask depth network structure model herein is to utilize depth
Neural network learning platform realizes the learning framework of multitask so that multitask deep learning has puts space to good use；
As a kind of implementation, the image search method in the present embodiment, step S1, including：
S11, the attribute classification for determining target image and training image subset corresponding with attribute classification；In retrieval image
During, because image species is rich and varied, need to divide various images in detail, could be easier to retrieve
Target image, for example, in the image attributes classification of people, smearing lipstick and wearing earrings has certain correlation, in the attribute point of car
In class, more typically, the Audi of black is in the majority for red faraday, and taxi is blue or green mostly, and truck is usually red
Color or blueness.The target image that will be cut out, such as vehicle, face, pedestrian carry out description text by multiple attribute classifications, such as
Picture name be attribute classification, respectively (A1, B2, C3 ... Fn), training image subset corresponding with attribute classification for (H1,
H2, H3 ... Hm), by taking vehicle as an example, license plate number identical car be same category, its sub- attribute for vehicle body color, logo, car
Type etc.；By taking face as an example,, used as a classification, its sub- attribute is angle, expression, the colour of skin, sex etc. for everyone；
S12, the multiple attribute task datas of input；It is many, it is necessary to carry out multi-task learning after above-mentioned multiple attributes are classified
Tasking learning is, using depth network structure model, to share and represent using the method study multiple tasks of training.Depth nerve net
The powerful expression ability of network causes that multi-task learning can then utilize task dependencies joint to improve the precision of multiple attributive classifications.
Individually when training, it is impossible to utilize these information.It is input into as supervisory signals by the use of multiple attributes, can be fully learnt to image
The contact of each attribute and the minutia of image, are characterized using multi-task learning to image, in answering for image retrieval
Use, can preferably overcome weather, environment, illumination etc. to influence.Multiple attribute task datas are input into depth network structure model,
As shown in figure 3, Input1, Input2, Input3, represent the input data of multiple different tasks, L is represented between different task
Shared layer, M represents specific layer between each task.
S13, shared relevant task data resource；Have by between the shared multiple tasks of the depth network structure model L for scheming X
Related data, excavate common characteristic, are easy to retrieval.
S14, according to multiple attribute tasks, determine corresponding loss function；In statistics, loss function refers to a kind of by one
An element of the individual event in a sample space be mapped to an expression financial cost related to its event or chance into
A kind of function on this real number, such as figure X represent different task corresponding loss function layer Task1, Task2, Task3, each
Task has its corresponding loss function.
As a kind of implementation, the image search method in the present embodiment, step S14, according to multiple attribute tasks, really
Fixed corresponding loss function；Including：
Formula Lt=- (1-t) * ㏒ (1-Pt)-t* ㏒ (pt)；
Wherein t represents whether the label that a certain learning tasks learning is arrived is equal with true tag, equal then t=1, not phase
Deng t=0；Pt represents that study is output as the probability of true tag.Image is fully learnt using depth learning technology, until
The error that the loss function of each attribute is represented is reduced to certain degree.
S15, according to 256 dimension depth characteristics description, obtain target image fisrt feature subset.By all image normalizations
Be the images of 224 × 224 unification sizes, according to set network (several shared convolutional layers+full articulamentum 256 ties up+it is defeated
Go out layer), learn 256 dimension datas, extract the most representative characteristic of image, using above-mentioned depth tasking learning train come
Network model, 256 dimension depth characteristic Fd are turned to by image abstraction, and the dramatic subset for obtaining image is fisrt feature subset D d
={ Fd }.
S2, set up target image characteristics storehouse；Target image characteristics storehouse represents the set that feature extraction is carried out to large nuber of images
L。
As a kind of implementation, image search method in the present embodiment, as shown in figure 4, step S2, sets up target image
Feature database, including：
S21, extraction sift describe the second feature subset of image；Sift features are also a kind of more traditional extraction image
The conventional describing mode of expression, sift features are the local features of image, its scaling, brightness change, rotation to image
Maintain the invariance, to radiation conversion, visible change, noise also keeps a certain degree of stability, extracts sift and describes image
Second feature subset is Fs.
S22, extraction lbp describe the third feature subset of image；Lbp features are also a kind of more traditional extraction image table
The conventional describing mode for reaching, in Digital Image Processing and area of pattern recognition, lbp refers to local binary patterns, initial work(to lbp features
Can be auxiliary Image Warping, good representation can be carried out to the local detail of image, extract the 3rd spy that lbp describes image
Subset is levied for Fl.
S23, according to second feature subset, third feature subset, obtain the fourth feature subset of traditional description image；Because
Second feature subset, third feature subset are all the mode of traditional characteristic description, so unified combination obtains traditional characteristic description
The third feature subset of image is Dt={ Fs, Fl }；
S24, the fifth feature subset for obtaining individual target image, with reference to the fisrt feature of above-mentioned dramatic characteristics of image
Subset D d={ Fd }, obtains describing fifth feature subset D f={ Fd, Fs, Fl } of individual target image.
As a kind of implementation, the image search method in the present embodiment, above-mentioned steps S15, step S23 and step
In S24 fourth feature subset include second feature subset and third feature subset, fifth feature subset include fisrt feature subset,
Second feature subset and third feature subset.As described above：Because fisrt feature subset is Dd={ Fd }, second feature subset is
Fs, third feature subset is Fl, and fourth feature subset is Dt={ Fs, Fl }, and fifth feature subset is Df={ Fd, Fs, Fl }.
As a kind of implementation, image search method in the present embodiment, target feature library is comprising multiple in step S2
Target image with fifth feature subset to be retrieved.Large nuber of images is extracted according to depth network structure model, is built
Vertical magnanimity target image feature database L, L=Df1, Df2, Df3 ... Dfn }.
S3, the character subset for being input into image to be retrieved；As it is input into picture to be checked.
Used as a kind of implementation, image search method in the present embodiment, step S3 is input into feature of image to be retrieved
Collection, image is extracted using S1 the step of building multitask depth network structure model, extracts the feature description of image to be checked
Character subset Ds.
S4, the character subset for calculating image to be retrieved and every similarity distance of image in characteristics of image storehouse.
Used as a kind of implementation, image search method in the present embodiment, step S4 calculates feature of image to be retrieved
Collection and every similarity distance of image in characteristics of image storehouse, including：
Formula Dis=0.75*d (d)+0.125*d (s)+0.125*d (l)；Wherein, d (d) represents the spy of image to be retrieved
Levy the distance of subset and fisrt feature subset；D (s) represents the character subset of image to be retrieved and the distance of second feature subset；d
L () represents the character subset of image to be retrieved and the distance of third feature subset；Dis represents final similarity distance.Calculate
The every group of feature gone out in the Feature Descriptor collection Ds and retrieval character storehouse L of image to be retrieved carries out Euclidean distance calculating respectively, obtains
Distance between three groups of features, is finally weighted fusion and obtains final apart from Dis.
S5, according to the image minimum with image distance to be retrieved is obtained after distance ordered arrangement from small to large, finally away from
From compared with being closer to a distance from novel improving eyesight logo image and image to be retrieved, according to the order that Dis is ascending, output is relative
The retrieval image answered is returned as a result, and ranking is more forward to represent that similarity is higher, realizes scheming to search the image retrieval function of figure.
Embodiment 2
The present embodiment provides a kind of image retrieving apparatus, including such as lower unit：
Construction unit 51, for building multitask depth network structure model；
Unit 52 is set up, for setting up target image characteristics storehouse；
Input block 53, the character subset for being input into image to be retrieved；
Computing unit 54, character subset and every phase of characteristics of image in characteristics of image storehouse for calculating image to be retrieved
Like degree distance；
Acquiring unit 55, according to the figure obtained after distance ordered arrangement from small to large with image distance to be retrieved minimum
Picture.
As a kind of implementation, image retrieving apparatus in the present embodiment, as shown in fig. 6, construction unit 51, including：
First determining module 511, attribute classification and training figure corresponding with attribute classification for determining target image
As subset；
Input module 512, for being input into multiple attribute task datas；
Sharing module 513, for sharing relevant task data resource；
Second determining module 514, for according to multiple attribute tasks, determining corresponding loss function；
First acquisition module 515, for according to 256 dimension depth characteristic descriptions, obtaining the fisrt feature subset of target image.
As a kind of implementation, image retrieving apparatus in the present embodiment, the second determining module 514；Including：
Formula Lt=- (1-t) * ㏒ (1-Pt)-t* ㏒ (pt)；
Wherein t represents whether the label that a certain learning tasks learning is arrived is equal with true tag, equal then t=1, not phase
Deng t=0；Pt represents that study is output as the probability of true tag.
As a kind of implementation, image retrieving apparatus in the present embodiment, as shown in fig. 7, unit 52 is set up, including：
First extraction module 521, the second feature subset of image is described for extracting sift；
Second extraction module 522, the third feature subset of image is described for extracting lbp；
Second acquisition module 523, according to second feature subset, third feature subset, obtains the 4th of traditional description image
Character subset；
3rd acquisition module 524, the fifth feature subset for obtaining individual target image.
Used as a kind of implementation, image retrieving apparatus in the present embodiment, fourth feature subset includes second feature subset
With third feature subset, fifth feature subset include fisrt feature subset, second feature subset and third feature subset.
Used as a kind of implementation, image retrieving apparatus in the present embodiment, the target feature library set up in unit 52 is bag
Containing multiple target images with fifth feature subset to be retrieved.
Used as a kind of implementation, image retrieving apparatus in the present embodiment are input into image to be retrieved in input block 53
Character subset extracts image using construction unit.
Used as a kind of implementation, image retrieving apparatus in the present embodiment calculate image to be retrieved in computing unit 54
Character subset and every similarity distance of image in characteristics of image storehouse, including：
Formula Dis=0.75*d (d)+0.125*d (s)+0.125*d (l)；
Wherein, d (d) represents the character subset of image to be retrieved and the distance of fisrt feature subset；D (s) represents to be retrieved
The character subset of image and the distance of second feature subset；
D (l) represents the character subset of image to be retrieved and the distance of third feature subset；
Dis represents final similarity distance.
Claims (16)
1. a kind of image search method, it is characterised in that comprise the following steps：
Build multitask depth network structure model；
Set up target image characteristics storehouse；
It is input into the character subset of image to be retrieved；
Calculate the character subset and every similarity distance of characteristics of image in described image feature database of the image to be retrieved；
According to the image obtained after distance ordered arrangement from small to large with the image distance minimum to be retrieved.
2. method according to claim 1, it is characterised in that the structure multitask depth network structure model, including：
Determine the attribute classification and training image subset corresponding with the attribute classification of target image；
The multiple attribute task datas of input；
Share the related task data resource；
According to the multiple attribute task, corresponding loss function is determined；
According to 256 dimension depth characteristic descriptions, the fisrt feature subset of target image is obtained.
3. method according to claim 2, it is characterised in that described according to the multiple attribute task, determines corresponding
The step of loss function；Including：
Formula Lt=- (1-t) * ㏒ (1-Pt)-t* ㏒ (pt)；
Wherein t represents whether the label that a certain learning tasks learning is arrived is equal with true tag, equal then t=1, unequal t
=0；Pt represents that study is output as the probability of the true tag.
4. method according to claim 1, it is characterised in that described to set up target image characteristics storehouse, including：
Extract the second feature subset that sift describes image；
Extract the third feature subset that lbp describes image；
According to the second feature subset, third feature subset, the fourth feature subset of traditional description image is obtained；
Obtain the fifth feature subset of individual target image.
5. the method according to claim 2 or 4, it is characterised in that the fourth feature subset includes second feature subset
With third feature subset, fifth feature subset include fisrt feature subset, second feature subset and third feature subset.
6. the method according to claim 1 or 4, it is characterised in that the target feature library is to be retrieved comprising multiple
Target image with the fifth feature subset.
7. method according to claim 1, it is characterised in that the character subset of the input image to be retrieved is using described
The step of building multitask depth network structure model extracts image.
8. method according to claim 1, it is characterised in that calculate the character subset and the figure of the image to be retrieved
As every similarity distance of image in feature database, including：
Formula Dis=0.75*d (d)+0.125*d (s)+0.125*d (l)；
Wherein, d (d) represents the character subset of the image to be retrieved and the distance of the fisrt feature subset；D (s) represents institute
State the character subset of image to be retrieved and the distance of the second feature subset；
D (l) represents the character subset of the image to be retrieved and the distance of the third feature subset；
Dis represents final similarity distance.
9. a kind of image retrieving apparatus, it is characterised in that including such as lower unit：
Construction unit, for building multitask depth network structure model；
Unit is set up, for setting up target image characteristics storehouse；
Input block, the character subset for being input into image to be retrieved；
Computing unit, for calculating in the character subset of the image to be retrieved and described image feature database every characteristics of image
Similarity distance；
Acquiring unit, according to the figure obtained after distance ordered arrangement from small to large with the image distance minimum to be retrieved
Picture.
10. device according to claim 9, it is characterised in that the construction unit, including：
First determining module, attribute classification and training image corresponding with the attribute classification for determining target image
Collection；
Input module, for being input into multiple attribute task datas；
Sharing module, for sharing the related task data resource；
Second determining module, for according to the multiple attribute task, determining corresponding loss function；
First acquisition module, for according to 256 dimension depth characteristic descriptions, obtaining the fisrt feature subset of target image.
11. devices according to claim 10, it is characterised in that second determining module；Including：
Formula Lt=- (1-t) * ㏒ (1-Pt)-t* ㏒ (pt)；
Wherein t represents whether the label that a certain learning tasks learning is arrived is equal with true tag, equal then t=1, unequal t
=0；Pt represents that study is output as the probability of the true tag.
12. devices according to claim 9, it is characterised in that described to set up unit, including：
First extraction module, the second feature subset of image is described for extracting sift；
Second extraction module, the third feature subset of image is described for extracting lbp；
Second acquisition module, according to the second feature subset, third feature subset, obtains the fourth feature of traditional description image
Subset；
3rd acquisition module, the fifth feature subset for obtaining individual target image.
13. device according to claim 10 or 12, it is characterised in that the fourth feature subset includes second feature
Collection and third feature subset, fifth feature subset include fisrt feature subset, second feature subset and third feature subset.
14. device according to claim 9 or 12, it is characterised in that the target feature library set up in unit is bag
Containing multiple target images with the fifth feature subset to be retrieved.
15. devices according to claim 9, it is characterised in that extract image using construction unit in the input block.
16. devices according to claim 1, it is characterised in that calculate the character subset of the image to be retrieved with it is described
Every similarity distance of image in characteristics of image storehouse, including：
Formula Dis=0.75*d (d)+0.125*d (s)+0.125*d (l)；
Wherein, d (d) represents the character subset of the image to be retrieved and the distance of the fisrt feature subset；D (s) represents institute
State the character subset of image to be retrieved and the distance of the second feature subset；
D (l) represents the character subset of the image to be retrieved and the distance of the third feature subset；
Dis represents final similarity distance.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710038542.1A CN106886573A (en) | 2017-01-19 | 2017-01-19 | A kind of image search method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710038542.1A CN106886573A (en) | 2017-01-19 | 2017-01-19 | A kind of image search method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106886573A true CN106886573A (en) | 2017-06-23 |
Family
ID=59176022
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710038542.1A Pending CN106886573A (en) | 2017-01-19 | 2017-01-19 | A kind of image search method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106886573A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107315837A (en) * | 2017-07-17 | 2017-11-03 | 梧州井儿铺贸易有限公司 | The accurate image indexing system of one kind retrieval |
| CN107844784A (en) * | 2017-12-08 | 2018-03-27 | 广东美的智能机器人有限公司 | Face identification method, device, computer equipment and readable storage medium storing program for executing |
| CN108108657A (en) * | 2017-11-16 | 2018-06-01 | 浙江工业大学 | A kind of amendment local sensitivity Hash vehicle retrieval method based on multitask deep learning |
| CN108960412A (en) * | 2018-06-29 | 2018-12-07 | 北京京东尚科信息技术有限公司 | Image-recognizing method, device and computer readable storage medium |
| CN109063112A (en) * | 2018-07-30 | 2018-12-21 | 成都快眼科技有限公司 | A kind of fast image retrieval method based on multi-task learning deep semantic Hash, model and model building method |
| WO2019020049A1 (en) * | 2017-07-28 | 2019-01-31 | 杭州海康威视数字技术股份有限公司 | Image retrieval method and apparatus, and electronic device |
| WO2020014903A1 (en) * | 2018-07-18 | 2020-01-23 | Shenzhen Malong Technologies Co., Ltd. | Complexity-based progressive training for machine vision models |
| CN113722509A (en) * | 2021-09-07 | 2021-11-30 | 中国人民解放军32801部队 | Knowledge graph data fusion method based on entity attribute similarity |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103810505A (en) * | 2014-02-19 | 2014-05-21 | 北京大学 | Vehicle identification method and system based on multilayer descriptors |
| CN104317946A (en) * | 2014-10-31 | 2015-01-28 | 上海交通大学 | Multi-key image-based image content retrieval method |
| CN105469076A (en) * | 2015-12-29 | 2016-04-06 | 东莞市微模式软件有限公司 | Face comparing verification method based on multi-instance learning |
| CN105574215A (en) * | 2016-03-04 | 2016-05-11 | 哈尔滨工业大学深圳研究生院 | Instance-level image search method based on multiple layers of feature representations |
| CN105631037A (en) * | 2015-12-31 | 2016-06-01 | 北京恒冠网络数据处理有限公司 | Image retrieval method |
| CN105740906A (en) * | 2016-01-29 | 2016-07-06 | 中国科学院重庆绿色智能技术研究院 | Depth learning based vehicle multi-attribute federation analysis method |
| CN105808691A (en) * | 2016-03-03 | 2016-07-27 | 博康智能网络科技股份有限公司 | Gate vehicle retrieval method and system |
-
2017
- 2017-01-19 CN CN201710038542.1A patent/CN106886573A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103810505A (en) * | 2014-02-19 | 2014-05-21 | 北京大学 | Vehicle identification method and system based on multilayer descriptors |
| CN104317946A (en) * | 2014-10-31 | 2015-01-28 | 上海交通大学 | Multi-key image-based image content retrieval method |
| CN105469076A (en) * | 2015-12-29 | 2016-04-06 | 东莞市微模式软件有限公司 | Face comparing verification method based on multi-instance learning |
| CN105631037A (en) * | 2015-12-31 | 2016-06-01 | 北京恒冠网络数据处理有限公司 | Image retrieval method |
| CN105740906A (en) * | 2016-01-29 | 2016-07-06 | 中国科学院重庆绿色智能技术研究院 | Depth learning based vehicle multi-attribute federation analysis method |
| CN105808691A (en) * | 2016-03-03 | 2016-07-27 | 博康智能网络科技股份有限公司 | Gate vehicle retrieval method and system |
| CN105574215A (en) * | 2016-03-04 | 2016-05-11 | 哈尔滨工业大学深圳研究生院 | Instance-level image search method based on multiple layers of feature representations |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107315837A (en) * | 2017-07-17 | 2017-11-03 | 梧州井儿铺贸易有限公司 | The accurate image indexing system of one kind retrieval |
| US11586664B2 (en) | 2017-07-28 | 2023-02-21 | Hangzhou Hikvision Digital Technology Co., Ltd. | Image retrieval method and apparatus, and electronic device |
| WO2019020049A1 (en) * | 2017-07-28 | 2019-01-31 | 杭州海康威视数字技术股份有限公司 | Image retrieval method and apparatus, and electronic device |
| CN108108657B (en) * | 2017-11-16 | 2020-10-30 | 浙江工业大学 | Method for correcting locality sensitive Hash vehicle retrieval based on multitask deep learning |
| CN108108657A (en) * | 2017-11-16 | 2018-06-01 | 浙江工业大学 | A kind of amendment local sensitivity Hash vehicle retrieval method based on multitask deep learning |
| CN107844784A (en) * | 2017-12-08 | 2018-03-27 | 广东美的智能机器人有限公司 | Face identification method, device, computer equipment and readable storage medium storing program for executing |
| CN108960412A (en) * | 2018-06-29 | 2018-12-07 | 北京京东尚科信息技术有限公司 | Image-recognizing method, device and computer readable storage medium |
| WO2020014903A1 (en) * | 2018-07-18 | 2020-01-23 | Shenzhen Malong Technologies Co., Ltd. | Complexity-based progressive training for machine vision models |
| US11062180B2 (en) | 2018-07-18 | 2021-07-13 | Shenzhen Malong Technologies Co., Ltd. | Complexity-based progressive training for machine vision models |
| CN109063112B (en) * | 2018-07-30 | 2022-04-01 | 成都快眼科技有限公司 | Rapid image retrieval method, model and model construction method based on multitask learning deep semantic hash |
| CN109063112A (en) * | 2018-07-30 | 2018-12-21 | 成都快眼科技有限公司 | A kind of fast image retrieval method based on multi-task learning deep semantic Hash, model and model building method |
| CN113722509A (en) * | 2021-09-07 | 2021-11-30 | 中国人民解放军32801部队 | Knowledge graph data fusion method based on entity attribute similarity |
| CN113722509B (en) * | 2021-09-07 | 2022-03-01 | 中国人民解放军32801部队 | Knowledge graph data fusion method based on entity attribute similarity |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106886573A (en) | A kind of image search method and device | |
| Pulido | Geographies of race and ethnicity III: Settler colonialism and nonnative people of color | |
| CN105183869B (en) | Building knowledge mapping database and its construction method | |
| US10963700B2 (en) | Character recognition | |
| US10789317B2 (en) | Apparatus and user device for providing feedback effect about posts on social network services | |
| CN115481969A (en) | Resume screening method and device, electronic equipment and readable storage medium | |
| CN104778224A (en) | Target object social relation identification method based on video semantics | |
| CN107633022A (en) | Personnel's portrait analysis method, device and storage medium | |
| Neyland et al. | Visual accountability | |
| CN111143613A (en) | Method, system, electronic device and storage medium for selecting video cover | |
| CN110706136A (en) | Study on AI + digitized campus construction scheme | |
| US11373057B2 (en) | Artificial intelligence driven image retrieval | |
| EP3850538A1 (en) | Image management with region-based metadata indexing | |
| CN115063589A (en) | Knowledge distillation-based vehicle component segmentation method and related equipment | |
| Stahlbock et al. | Advances in Data Science and Information Engineering | |
| CN112052332A (en) | Retrieval method, retrieval device, electronic equipment and readable storage medium | |
| US20150035836A1 (en) | Data display and data display method | |
| CN112396060B (en) | Identification card recognition method based on identification card segmentation model and related equipment thereof | |
| Vanky et al. | Urban-semantic computer vision: a framework for contextual understanding of people in urban spaces | |
| Wang | The digital presentation of human-oriented urban design | |
| Ahmed et al. | An overview: Big data analysis by deep learning and image processing | |
| Patel et al. | Single-object detection from video streaming | |
| Maurya et al. | Application of google lens to promote information services beyond the traditional techniques | |
| Verma et al. | Crosswalk detection from satellite imagery for pedestrian network completion | |
| Zhang et al. | Automated data collection and safety analysis at intersections based on a novel video processing system |
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: 20170623 |
|
| RJ01 | Rejection of invention patent application after publication |