CN108229289A - Target retrieval method, apparatus and electronic equipment - Google Patents
Target retrieval method, apparatus and electronic equipment Download PDFInfo
- Publication number
- CN108229289A CN108229289A CN201710500550.3A CN201710500550A CN108229289A CN 108229289 A CN108229289 A CN 108229289A CN 201710500550 A CN201710500550 A CN 201710500550A CN 108229289 A CN108229289 A CN 108229289A
- Authority
- CN
- China
- Prior art keywords
- similarity
- target
- checked
- image
- feature vector
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Abstract
This application discloses target retrieval method and apparatus.One specific embodiment of the above method includes：Multiple images set is obtained, wherein, each image collection includes at least one image for containing at least one target to be checked；The clarification of objective to be checked vector that at least part image is included in multiple images set is extracted respectively；In at least part feature vector of extraction, the similarity ranking information in the image collection belonging to the target to be checked of the similarity and the similarity of each two feature vector in two feature vector instructions is determined respectively；According to each similarity and each similarity ranking information, the probability that the target to be checked indicated by each two feature vector is same target is determined.The embodiment realizes the preliminary search to target, reduces the workload of mark personnel, improves the efficiency of target retrieval.
Description
Technical field
This application involves computer vision fields, and in particular to image processing field more particularly to target retrieval method, dress
It puts and electronic equipment.
Background technology
Computer vision is using a kind of simulation of computer and relevant device to biological vision, can use camera
The data of subject and information are obtained with computer.Target retrieval is an important research in computer vision research
Direction, can be according to the image of a target of input, and it is same to be searched out in large-scale data set with the target of input
All images of one target；This similarity value to image can also be obtained to a pair of of image of its input.
Target retrieval generally requires to rely on the target retrieval data largely marked.If all marks are all by marking people
The artificial mark of member, then mark person's workload is larger, less efficient.
Invention content
The purpose of the application is to propose a kind of target retrieval method and apparatus and electronic equipment, to solve background above skill
The technical issues of art part is mentioned.
In a first aspect, this application provides a kind of target retrieval method, including：Multiple images set is obtained, wherein, each
Image collection includes at least one image for containing at least one target to be checked；At least portion is extracted in above-mentioned multiple images set respectively
The clarification of objective to be checked vector that partial image is included；In at least part feature vector of extraction, determine that each two is special respectively
Levy the phase in the image collection belonging to the target to be checked of the similarity and the similarity of vector in two feature vector instructions
Like degree ranking information；According to each similarity and each similarity ranking information, the mesh to be checked indicated by each two feature vector is determined
It is designated as the probability of same target.
In some embodiments, the above method further includes：In response to the target to be checked indicated by there are two feature vectors
Probability for same target meets the first preset condition, generates to inquire whether two feature vectors indicate same target
Prompt message.
In some embodiments, above-mentioned first preset condition includes at least one of：Above-mentioned probability is more than predetermined probability
Threshold value；Above-mentioned probability is located at the preceding predetermined ratio range that all probability sort from big to small.
In some embodiments, above-mentioned image collection derives from video source；And above-mentioned above-mentioned multiple images are extracted respectively
The clarification of objective to be checked vector that at least part image is included in set, including：The identification information of each image collection is obtained,
Above-mentioned identification information includes the first flag information of the video source belonging to the image collection and the image collection is regarded in affiliated
Second identifier information in frequency source；The identification information of each two image collection is compared, generation compares list；It extracts respectively
The clarification of objective to be checked vector that at least part image is included in each image collection, and determine each in above-mentioned comparison list
Compare the similarity between indicated feature vector.
In some embodiments, the image in each image collection includes same target to be checked；And the above method also wraps
It includes：For each comparison in above-mentioned comparison list, the first flag information of two image collections indicated by the comparison is determined
It is whether identical；It is identical in response to the first flag information of two image collections indicated by the comparison, it determines indicated by the comparison
Two image collections in the presence or absence of generation moment identical image；And/or in response to two images indicated by the comparison
In the presence of the image that the generation moment is identical in set, then the comparison is deleted, to optimize above-mentioned comparison list.
In some embodiments, the above method further includes：According to the generation moment of every image in each image collection, really
The average generation moment of fixed each image collection；For each comparison in above-mentioned comparison list, the two of comparison instruction is determined
Whether the difference at the average generation moment of a image collection is more than preset duration；In response to two image collections of comparison instruction
Averagely the difference at generation moment is more than preset duration, the comparison is deleted, to optimize above-mentioned comparison list.
In some embodiments, it is above-mentioned extract respectively at least part image in above-mentioned multiple images set included it is to be checked
Clarification of objective vector, including：Using preset first nerves network extract respectively above-mentioned multiple images set it is above-mentioned at least
The clarification of objective to be checked vector that every image is included in parts of images.
In some embodiments, the above-mentioned similarity for determining each two feature vector respectively and the similarity are at this two
The similarity ranking information in image collection belonging to the target to be checked of feature vector instruction, including：Determine above-mentioned comparison list
In each compare similarity between two indicated feature vectors；Determine the similarity set of each target to be checked, wherein,
Above-mentioned similarity set includes the corresponding similarity of comparison that the target to be checked is contained in above-mentioned comparison list；It will be above-mentioned similar
Each similarity in degree set is arranged according to descending or ascending sequence, determines each target to be checked each
From the similarity ranking information in affiliated image collection.
In some embodiments, it is above-mentioned according to each similarity and each similarity ranking information, determine each two feature vector
Indicated target to be checked is the probability of same target, including：By the similarity between each two feature vector and this two
Ranking information of the target to be checked of feature vector instruction in respective sequencing of similarity inputs preset grader, based on above-mentioned
Grader determines probability of the target to be checked indicated by each two feature vector for same target.
In some embodiments, above-mentioned grader is pre-established using following methods：Obtain the multiple images marked in advance
Set, wherein, the information of mark is used to represent that multiple images set includes whether two targets are same target；Extraction is each
The clarification of objective vector that image collection is included, and determine the similarity between each two feature vector；According to each two spy
Similarity between sign vector, determines the sequencing of similarity between each feature vector and other feature vectors；According to above-mentioned phase
It sorts like degree, determines ranking of each two feature vector in respective sequencing of similarity；According to above-mentioned multiple images set
Similarity and above-mentioned ranking training grader between markup information, each two feature vector.
In some embodiments, above-mentioned acquisition multiple images set, including：Obtain multiple video sources, each video source packet
Include at least one image collection；At least part image in above-mentioned multiple video sources is detected using preset nervus opticus network,
Determine the target to be checked that every image in above-mentioned at least part image is included；To each target to be checked for detecting into rower
Note obtains mark image collection corresponding with each target to be checked.
In some embodiments, it is above-mentioned that each target to be checked detected is labeled, it obtains and each target to be checked
Corresponding mark image collection, including：The each target to be checked detected is labeled using minimum enclosed rectangle frame；It cuts
Each tab area obtains cutting image collection corresponding with each target to be checked.
In some embodiments, above-mentioned acquisition multiple images set, including：It determines to wrap in each above-mentioned cutting image collection
The quantity of the cutting image contained and the following parameter of every cutting image：The first pixel quantity along the first direction, along second
The ratio of second pixel quantity in direction, above-mentioned first pixel quantity and above-mentioned second pixel quantity, wherein, above-mentioned first direction
It is respectively the extending direction of two adjacent length of sides of above-mentioned minimum enclosed rectangle with above-mentioned second direction；Choose each cutting image
The preset quantity for meeting the second preset condition in set cuts image, forms new image collection.
Second aspect, this application provides a kind of target retrieval device, above device includes：Image collection acquiring unit,
For obtaining multiple images set, wherein, each image collection includes at least one image for containing at least one target to be checked；It is special
The vectorial extraction unit of sign, for extracting the spy of target to be checked that at least part image is included in above-mentioned multiple images set respectively
Sign vector；Similarity determining unit, at least part feature vector of extraction, determining each two feature vector respectively
The similarity ranking in image collection belonging to the target to be checked of similarity and the similarity in two feature vector instructions
Information；Probability determining unit, for according to each similarity and each similarity ranking information, determining indicated by each two feature vector
Target to be checked be same target probability.
In some embodiments, above device further includes：Prompt message generation unit, in response to there are two features
Target to be checked indicated by vector meets the first preset condition for the probability of same target, generate for inquire two features to
Whether amount indicates the prompt message of same target.
In some embodiments, above-mentioned first preset condition includes at least one of：Above-mentioned probability is more than predetermined probability
Threshold value；Above-mentioned probability is located at the preceding predetermined ratio range that all probability sort from big to small.
In some embodiments, above-mentioned image collection derives from video source；And features described above vector extraction unit includes：
Identification information acquisition module, for obtaining the identification information of each image collection, above-mentioned identification information includes the image collection institute
Second identifier information of the first flag information and the image collection of the video source of category in affiliated video source；Compare list
Generation module, for the identification information of each two image collection to be compared, generation compares list；Characteristic vector pickup mould
Block for extracting the clarification of objective to be checked vector that at least part image in each image collection is included respectively, and determines
State the similarity for comparing and each being compared in list between indicated feature vector.
In some embodiments, the image in each image collection includes same target to be checked；And features described above vector
Extraction unit, which further includes, compares list optimization module, and above-mentioned comparison list optimization module is used for：For in above-mentioned comparison list
It is each to compare, determine whether the first flag information of two image collections indicated by the comparison is identical；In response to the comparison institute
The first flag information of two image collections indicated is identical, determines to whether there is in two indicated by the comparison image collections
Generate moment identical image；It is and/or identical in response to there is the generation moment in two image collections indicated by the comparison
Image then deletes the comparison, to optimize above-mentioned comparison list.
In some embodiments, features described above vector extraction unit, which further includes, compares list optimization module, above-mentioned comparison row
Table optimization module is used for：According to the generation moment of every image in each image collection, the average life of each image collection is determined
Into the moment；For each comparison in above-mentioned comparison list, when determining the average generation of two image collections of comparison instruction
Whether the difference carved is more than preset duration；It is more than in advance in response to the difference at the average generation moment of two image collections of comparison instruction
If duration, the comparison is deleted, to optimize above-mentioned comparison list.
In some embodiments, features described above vector extraction unit is further used for：Utilize preset first nerves network
Extract respectively every image is included in above-mentioned at least part image of above-mentioned multiple images set clarification of objective to be checked to
Amount.
In some embodiments, above-mentioned similarity determining unit includes：Similarity determining module, for determining above-mentioned comparison
The similarity between two indicated feature vectors is each compared in list；Similarity set determining module, it is every for determining
The similarity set of a target to be checked, wherein, above-mentioned similarity set includes containing the target to be checked in above-mentioned comparison list
The corresponding similarity of comparison；Similarity set ranking module, for by each similarity in above-mentioned similarity set according to
Descending or ascending sequence is arranged, and determines that each target to be checked is similar in respectively affiliated image collection
Spend ranking information.
In some embodiments, above-mentioned probability determining unit is further used for：It will be similar between each two feature vector
The ranking information of degree and the target to be checked of two feature vector instructions in respective sequencing of similarity inputs preset point
Class device determines the probability that the target to be checked indicated by each two feature vector is same target based on above-mentioned grader.
In some embodiments, above-mentioned grader is established unit by grader and is pre-established, and above-mentioned grader is established single
Member includes：Image collection acquisition module is marked, for obtaining the multiple images set marked in advance, wherein, the information of mark is used
Include whether two targets are same target in expression multiple images set；Similarity determining module, for extracting each figure
Image set closes included clarification of objective vector, and determines the similarity between each two feature vector；Sequencing of similarity determines
Module, for according to the similarity between each two feature vector, determining between each feature vector and other feature vectors
Sequencing of similarity；Ranking determining module, for according to above-mentioned sequencing of similarity, determining each two feature vector respective similar
Ranking in degree sequence；Training module, for according to above-mentioned multiple images set markup information, between each two feature vector
Similarity and above-mentioned ranking training grader.
In some embodiments, above-mentioned image collection acquiring unit includes：Video source acquisition module, for obtaining multiple regard
Frequency source, each video source include at least one image collection；Module of target detection, for being examined using preset nervus opticus network
At least part image in above-mentioned multiple video sources is surveyed, it is to be checked to determine that every image in above-mentioned at least part image is included
Target；Target labeling module for being labeled to each target to be checked detected, obtains corresponding with each target to be checked
Mark image collection.
In some embodiments, above-mentioned target labeling module is further used for：Using minimum enclosed rectangle frame to detecting
Each target to be checked be labeled；Each tab area is cut, obtains cutting image collection corresponding with each target to be checked.
In some embodiments, above-mentioned target labeling module is further used for：It determines in each above-mentioned cutting image collection
Comprising cutting image quantity and every cutting image following parameter：The first pixel quantity along the first direction, along
The ratio of second pixel quantity in two directions, above-mentioned first pixel quantity and above-mentioned second pixel quantity, wherein, above-mentioned first party
To the extending direction for two adjacent length of sides for above-mentioned second direction being respectively above-mentioned minimum enclosed rectangle；Each cut is chosen to scheme
Image set meets the second preset condition preset quantity in closing cuts image, forms new image collection.
The third aspect, the embodiment of the present application provide a kind of server, including：One or more processors；Storage device,
For storing one or more programs, when said one or multiple programs are performed by said one or multiple processors so that on
It states one or more processors and realizes any of the above-described described method of embodiment.
Fourth aspect, the embodiment of the present application provide a kind of computer readable storage medium, are stored thereon with computer journey
Sequence, the program realize any of the above-described embodiment described method when being executed by processor.
The target retrieval method and apparatus that the application provides, it is first determined then multiple images set extracts multiple images
The clarification of objective to be checked vector that at least part image is included in set, then at least part feature vector of extraction,
It determines belonging to the target to be checked of the similarity and each similarity of each two feature vector in two feature vector instructions
Similarity ranking information in image collection finally according to each similarity and each similarity ranking information, determines each two spy
Levy the probability that the target to be checked indicated by vector is same target.In this way, it can be obtained by different images using electronic equipment
Target to be checked be same target probability, can realize the preliminary search to target, reduce the workload of mark personnel, carry
The high efficiency of target retrieval.
Description of the drawings
By reading the detailed description made to non-limiting example made with reference to the following drawings, the application's is other
Feature, objects and advantages will become more apparent upon：
Fig. 1 is the flow chart according to one embodiment of an object of the application search method；
Fig. 2 is the flow chart according to the extraction feature vector of an object of the application search method；
Fig. 3 is the flow chart according to the acquisition multiple images set of an object of the application search method；
Fig. 4 is the structure diagram of one embodiment that device is retrieved according to an object of the application；
Fig. 5 is adapted for the structure diagram of the electronic equipment for realizing the embodiment of the present application.
Specific embodiment
The application is described in further detail with reference to the accompanying drawings and examples.It is understood that this place is retouched
The specific embodiment stated is used only for explaining related invention rather than the restriction to the invention.It also should be noted that in order to
Convenient for description, illustrated only in attached drawing and invent relevant part with related.
It should be noted that in the absence of conflict, the feature in embodiment and embodiment in the application can phase
Mutually combination.The application is described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.
Fig. 1 shows the flow 100 according to one embodiment of an object of the application search method.As shown in Figure 1, this reality
The target retrieval method of example is applied, is included the following steps：
Step 101, multiple images set is obtained.
In the present embodiment, the electronic equipment (such as terminal or server) of target retrieval method operation thereon can pass through
Wired connection mode multiple images set or acquisition user that either radio connection acquisition is locally stored is defeated by terminal
The multiple images set entered.Each image collection in above-mentioned image collection can include multiple images, and multiple above-mentioned images
In can include an at least image for including target to be checked, the quantity of above-mentioned target to be checked can be one or more
It is a.Above-mentioned target to be checked can be the target of any required retrieval such as pedestrian, vehicle, non-motor vehicle.
Step 102, the clarification of objective to be checked that at least part image is included in above-mentioned multiple images set is extracted respectively
Vector.
After above-mentioned multiple images set is obtained, at least part image institute in above-mentioned multiple images set can be extracted respectively
Comprising clarification of objective to be checked vector.It is understood that there may be multiple to include mesh to be checked in above-mentioned multiple images set
Target image, can be with selected part or all images, with what is included in parts of images selected by extraction in multiple above-mentioned images
Clarification of objective vector to be checked.Features described above vector is used to characterize the spy different from other targets to be checked of each target to be checked
Sign.
Step 103, at least part feature vector of extraction, determine respectively each two feature vector similarity and
The similarity ranking information in image collection belonging to target to be checked of the similarity in two feature vector instructions.
After part clarification of objective vector to be checked is extracted, can in the feature vector of extraction selected part or whole
Feature vector carries out following handle：Determine the similarity of each two feature vector in the feature vector chosen.On being calculated
After stating similarity, for each similarity, it may be determined that belonging to target to be checked of the similarity in two feature vector instructions
Image collection in similarity ranking information.
In the present embodiment, above-mentioned similarity can be related to the distance between two feature vectors, above-mentioned distance, Ke Yishi
Euclidean distance, Ming Shi distances, manhatton distance, Chebyshev's distance, mahalanobis distance etc..Above-mentioned similarity can also be vector
Space cosine similarity, Pearson correlation coefficients, Jaccard similarity factors, adjustment cosine similarity etc..
It is understood that for two feature vectors in different images set, it may be determined that each feature vector with
The similarity between other feature vector in image collection belonging to another feature vector in addition to this feature vector, then may be used
To obtain two similarity set.According to the similarity between two obtained similarity set and two feature vectors,
It can determine similarity between the two feature vectors ranking information in two similarity set respectively.For example,
For the feature vector 2 in the feature vector 1 in image collection A and image collection B, feature vector 1 can be calculated first
With the similarity between feature vector 2.Then it can calculate between the other feature vector in feature vector 1 and image collection B
Similarity, obtain the similarity set of each feature vector and feature vector 1 in image collection B.Then feature can be calculated
The similarity between other feature vector in 2 and image collection A of vector obtains each feature vector in image collection A and spy
The similarity set of sign vector 2.Finally, it can determine that this is similar according to the similarity between feature vector 1 and feature vector 2
Spend the ranking information in two similarity set respectively.It is understood that after two similarity set are obtained, it can be right
Each similarity in similarity set, which is had, arrives greatly small sequence or ascending sequence, then determines the ranking letter of the similarity
Breath.
Step 104, it according to each similarity and each similarity ranking information, determines to be checked indicated by each two feature vector
Target is the probability of same target.
After above-mentioned similarity and similarity ranking information is obtained, it may be determined that the mesh to be checked of two feature vector instructions
It is designated as the probability of same target.For example, the similarity value of two feature vectors is larger, the similarity in two similarity set
Ranking in two similarity set is located at upstream, then it can be assumed that the target to be checked of two feature vector instructions is same
The probability of target is larger.
In some optional realization methods of the present embodiment, when the probability being calculated in step 104 meets first in advance
If it during condition, generates to inquire whether two feature vectors indicate the prompt message of same target.
In this realization method, above-mentioned prompt message can be used for mark personnel is reminded to treat two feature vector instructions
Whether inspection target is that same target this problem is reaffirmed, so as to efficiently reduce the workload of mark personnel.
In some optional realization methods of the present embodiment, above-mentioned first preset condition can include following at least one
：Above-mentioned probability is more than predetermined probability threshold value, above-mentioned probability is located at the preceding predetermined ratio range that all probability sort from big to small.
When it is very likely same target that above-mentioned probability, which indicates two targets to be checked, prompt message can be exported.This reality
In existing mode, it can determine whether two targets to be checked are very likely same mesh by setting probability threshold value or probability proportion
Mark.
The target retrieval method that above-described embodiment of the application provides, it is first determined then multiple images set is extracted more
At least part image is included in a image collection clarification of objective to be checked vector, then at least part feature of extraction to
In amount, the target institute to be checked that the similarity of each two feature vector and each similarity are indicated in two feature vectors is determined
Similarity ranking information in the image collection of category finally according to each similarity and each similarity ranking information, determines every two
Target to be checked indicated by a feature vector is the probability of same target.In this way, it can be obtained by different figures using electronic equipment
Probability and probability ranking information in its image collection of the target to be checked for same target as in, by probability with
And the ranking information of probability can realize the preliminary search to target, reduce the workload of mark personnel, improve target inspection
The efficiency of rope.
With continued reference to Fig. 2, it illustrates the flows 200 of the extraction feature vector according to an object of the application search method.
As shown in Fig. 2, in the present embodiment, each image collection source and video source, above-mentioned video source can be monitor video, above-mentioned prison
Preset value can be less than by controlling the distance between center of monitoring range of video, this preset value can according to practical application scene into
Row setting, for example, can be 300 meters the distance between (on street two the adjacent center of monitoring range of monitoring camera),
Can also be 10-50 meters (the distance between centers of monitoring range of different monitoring camera in shop).In the present embodiment, prison
The center of control range can be the center of the drop shadow spread of range that monitoring camera is monitored on the ground, for example, when monitoring
When the drop shadow spread of the range that camera is monitored on the ground is oval, center is elliptical center.It is understood that
It can include multiple images set in each video source.
Extraction feature vector can be realized in the present embodiment by following steps：
Step 201, the identification information of each image collection is obtained.
In the present embodiment, above-mentioned identification information can include the image collection belonging to video source first flag information with
And second identifier information of the image collection in affiliated video source.Above-mentioned first flag information can be the image collection institute
Number of the video source of category in multiple video sources, above-mentioned second identifier information can be the image collections in affiliated video source
In number, such as the identification information of image collection can be 2-5, represent the 5th image collection in the 2nd video source.
Step 202, the identification information of each two image collection is compared, generation compares list.
The identification information of each two image collection in above-mentioned multiple monitor videos is compared, generation compares list.
Above-mentioned comparison list includes multiple comparisons, and each compare can be represented using the identification information of image collection, such as can be with
For 2-5：3-1.
Step 203, the clarification of objective to be checked vector that at least part image is included in each image collection is extracted respectively,
And it determines each to compare the similarity between indicated feature vector in above-mentioned comparison list.
It, can be as unit of comparing the comparison in list in the similarity between determining each two clarification of objective vector
It carries out, display and subsequent manual examination and verification conducive to target retrieval result.It is understood that above-mentioned audit can by manually Lai
It carries out, can also be realized by other algorithms, the present embodiment does not limit this.
In some optional realization methods of the present embodiment, when extracting feature vector, preset first can be utilized
Neural network extract respectively every image is included at least part image of multiple images set clarification of objective to be checked to
Amount.
Every image at least part image in each image collection can be extracted using preset first nerves network
Included in clarification of objective to be checked vector, specifically, multiple convolutional layers in above-mentioned first nerves network can be utilized to carry
Take the feature vector in every image.After at least part feature vector of each image collection is obtained, can to each feature to
Measure average or calculate weighted mean, using obtained average vector as the image collection indicated by clarification of objective to be checked to
Amount.
In some optional realization methods of the present embodiment, the image in each image collection includes same mesh to be checked
Mark, that is to say, that there are one the targets to be checked in each image collection.Then after generation compares list, the above method also wraps
Include the step of optimization compares list.The step of above-mentioned optimization comparison list, includes：For each comparison in above-mentioned comparison list,
Determine whether the first flag information of two image collections indicated by the comparison is identical；In response to two indicated by the comparison
The first flag information of image collection is identical, determines in two image collections indicated by the comparison with the presence or absence of generation moment phase
Same image；And/or in response in the presence of the image that the generation moment is identical, then being deleted in two image collections indicated by the comparison
Except the comparison, to optimize comparison list.
Whether each first flag information for comparing indicated image collection that detection is compared in list is identical, if phase
Together, then illustrate that two image collections indicated by the comparison belong to same video source.Determining that two image collections belong to same
After video source, then detect with the presence or absence of the image that the generation moment is identical in two image collections, if it is present explanation above-mentioned two
Exist in an at least frame image there are two targets to be checked indicated by above-mentioned two image collection in a image collection, then this two
A target to be checked can not possibly be same target, this comparison be deleted, to optimize comparison list.
In some optional realization methods of the present embodiment, the step of above-mentioned optimization compares list, can also include following
Step：According to the generation moment of every image in each image collection, the average generation moment of each image collection is determined；For
Each comparison in list is compared, determines whether the difference at the average generation moment of two image collections of comparison instruction is more than in advance
If duration；It is more than preset duration in response to the difference at the average generation moment of two image collections of comparison instruction, deletes the ratio
It is right, to optimize comparison list.
According to the generation moment of frame image every in each image collection, it may be determined that during the average generation of each image collection
It carves.Detection compares whether the difference at the average generation moment of two image collections each compared in list is more than preset duration.This
Place, when video source is monitor video, preset duration can be according to the location of practical monitoring camera and to be checked
The normal walking speed of target and it is determining.Come using target to be checked as pedestrian for example, if two image collections are averaged
The moment is generated more than preset duration, then illustrates the probability that the pedestrian indicated by two image collections of the comparison is same pedestrian
It is smaller, then the comparison is deleted.For example, positioned at two monitoring cameras in same street, monitoring range is adjacent, each
The length that monitoring camera can monitor this street is 150 meters, then two monitoring cameras can monitor the street of 300 meters of length
Road.The speed of travel of normal pedestrian is about 1 meter per second, it is assumed that the pedestrian passes by above-mentioned 300 meters of monitoring range, takes around
300 seconds.The difference at the average generation moment between two image collections then obtained in two monitor videos is about 150 seconds,
It is 180 seconds that preset duration can be set herein, when the difference at above-mentioned averagely generation moment is more than 180 seconds, it is believed that above-mentioned two row
People is not same pedestrian.
In some optional realization methods of the present embodiment, in step 103 when determining similarity ranking information, may be used also
To be realized by following steps：It determines to compare the similarity each compared in list between two indicated feature vectors；
Determine the similarity set of each target to be checked；By each similarity in similarity set according to descending or ascending
Sequence arranged, determine the similarity ranking information in image collection of each target to be checked belonging to respectively.
After comparison list is determined, what it is due to each comparison instruction is two image collections, can be determined first each
Then the feature vector of image collection determines the similarity between the feature vector of each two image collection, then each comparison pair
Answer a similarity.Then the similarity set of each target to be checked is determined, above-mentioned similarity set includes comparing wraps in list
The corresponding similarity of comparison of the target to be checked is contained.After the similarity set of each target to be checked is obtained, it can incite somebody to action each
Similarity is arranged according to descending or ascending sequence, determines each target to be checked in respectively affiliated image set
Similarity ranking information in conjunction.For example, one compared in list is compared as 2-5：3-1, and share 10 image sets
It closes, the ascending distance-taxis for now obtaining image collection 2-5 is：2-1、1-3、1-2、3-2、2-4、3-1、2-2、2-3、1-
1, then rankings of the image collection 3-1 in above-mentioned distance-taxis is the 6th.Similarly, the distance-taxis of image collection 3-1 is being obtained
Afterwards, rankings of the image collection 2-5 in this distance-taxis can also be determined.
In this realization method, the feature vector of each image collection can be to be checked by least partly being included in the image collection
The feature vector of the image of target weights to obtain.
In some optional realization methods of the present embodiment, determined indicated by each two feature vector in step 104
Target to be checked be same target probability when, can specifically be realized by following steps unshowned in Fig. 2：By each two
Row of the target to be checked of similarity and two feature vector instructions between feature vector in respective sequencing of similarity
Name information inputs preset grader, and it is same that the target to be checked indicated by each two feature vector is determined based on above-mentioned grader
The probability of target.
In this realization method, grader may be used to determine whether two targets to be checked are same target.It is above-mentioned default
Grader can be it is various can the grader based on numerical value output probability, such as can be support vector machine classifier.
In some optional realization methods of the present embodiment, above-mentioned grader can be by unshowned following in Fig. 2
Step is established：Obtain the multiple images set marked in advance；The clarification of objective vector that each image collection is included is extracted,
And determine the similarity between each two feature vector；According to the similarity between each two feature vector, each feature is determined
Sequencing of similarity between vector and other feature vectors；According to above-mentioned sequencing of similarity, determine each two feature vector each
From sequencing of similarity in ranking；According to the similarity between the markup information of multiple images set, each two feature vector
And above-mentioned ranking training grader.
Wherein, the information marked in above-mentioned multiple images set is used to represent that multiple images set includes two targets and is
No is same target, that is to say, that includes the multiple images collection that two targets are same target in above-mentioned multiple images set
It closes, further includes the multiple images set that two targets are not same target.Extract the spy for the target that each image collection is included
Then sign vector calculates the similarity between each two feature vector, the similarity for belonging to same target to be checked is arranged
Sequence obtains sequencing of similarity.It determines in above-mentioned sequencing of similarity, row of each two feature vector in respective sequencing of similarity
Name.For example, two feature vectors are respectively labeled as 1 and 5,10 feature vectors are shared, then the similar of feature vector 1 are being determined
After degree sequence, it may be determined that ranking of the feature vector 5 in above-mentioned sequencing of similarity；Similarly, in the phase that feature vector 5 is determined
After sorting like degree, it may be determined that ranking of the feature vector 1 in above-mentioned sequencing of similarity.After above-mentioned ranking is obtained, in utilization
The similarity and above-mentioned ranking training grader between the markup information of multiple images set, each two feature vector are stated, just
It can obtain preset grader.That is, utilize similarity, feature vector 1 between feature vector 1 and feature vector 5
The ranking of ranking and feature vector 5 in the sequencing of similarity of feature vector 1 in the sequencing of similarity of feature vector 5 with
And annotation results (whether the target indicated by target and feature vector 5 indicated by feature vector 1 is same target) training point
Class device, by such multiple training, it is possible to obtain preset grader.
In the present embodiment, for multiple images set, a large amount of useless images may be included, that is to say, that more
It is same target that may there was only the target that small part image is included in a image collection.If every in each image collection
Image is required for if mark personnel are labeled, and workload is very big.After this realization method is combined with embodiment illustrated in fig. 1,
Mark personnel are given by the way that two targets to be checked are very likely exported for the image of same target, mark people can be greatly reduced
The workload of member.Meanwhile after mark personnel confirm above-mentioned output information, the image after can will confirm that is used to classify
The training of device can also improve the accuracy of classifier training.
The target retrieval method that above-described embodiment of the application provides after clarification of objective vector to be checked is extracted, passes through
Structure compares list, and retrieval result can be made clear；List is compared by optimization simultaneously, follow-up calculating can be effectively reduced
Workload, improve computational efficiency.
With continued reference to Fig. 3, Fig. 3 is the flow according to the acquisition multiple images set of the target retrieval method of the present embodiment
300.As shown in figure 3, in the present embodiment, multiple images set can be obtained by following steps：
Step 301, multiple video sources are obtained.
In the present embodiment, above-mentioned video source can be the various video datas for containing at least one target to be checked, each
Video data can include multiple images.Each video source can include at least one image collection.
Step 302, at least part image in multiple video sources is detected using preset nervus opticus network, is determined above-mentioned
The target to be checked that every image at least part image is included.
In the present embodiment, preset nervus opticus network can be utilized at least part figure in multiple video sources of acquisition
As being detected, to determine target to be checked that every image in above-mentioned at least part image is included.Above-mentioned preset second
Neural network can be trained convolutional neural networks, can detect the target to be checked in the image of input.
Step 303, each target to be checked detected is labeled using minimum enclosed rectangle frame.
After the target in detecting every frame image, each target detected can be labeled.It, can in mark
To be labeled using various callout box, such as circle, rectangle, ellipse etc..In the present embodiment, it may be used what is detected
The minimum enclosed rectangle of target is labeled each target detected.It is understood that in the present embodiment, to detection
When the target gone out is labeled, it can also determine whether the clarity of the target detected meets the requirements first, if be unsatisfactory for
It is required that the image can be rejected, to ensure that the clarity for marking out the image come is all preferable.
Step 304, each tab area is cut, obtains cutting image collection corresponding with each target to be checked.
After being labeled to each target, each tab area can be cut, obtains cutting image corresponding with each target
Set.It is understood that being cut per frame only comprising a target in image, and each cut in image collection per frame image packet
Containing same target.
Step 305, determine that each cutting the quantity for cutting image included in image collection and every cuts image
Following parameter：The first pixel quantity along the first direction, the second pixel quantity in a second direction, above-mentioned first pixel quantity with
The ratio of above-mentioned second pixel quantity.
Since every cutting image is rectangle, it is possible to determine that every cuts the first pixel of image along its length
Quantity and the second pixel quantity in the width direction can also determine the first pixel quantity and width direction of length direction
The ratio of second pixel quantity.Wherein, above-mentioned first direction, second direction are respectively one kind of length direction and width direction.
In the present embodiment, the quantity of cutting image included in each cutting image collection can also be determined, it in this way can be by quantity mistake
Few cutting image collection is deleted.In this way, when establishing comparison list, the number for comparing and being compared in list can be efficiently reduced
Amount, reduces calculation amount.
Step 306, each preset quantity cutting image for cutting and meeting the second preset condition in image collection, structure are chosen
The image collection of Cheng Xin.
In order to enable the every image cut in image collection can clearly reflect clarification of objective, in the present embodiment,
Each image that cuts cut in image collection can be screened.Above-mentioned screening can choose the first pixel of length direction
Quantity is more than a certain numerical value, the second pixel quantity of width direction is more than a certain numerical value and cuts the big Mr. Yu of length-width ratio of image
The cutting image of one numerical value, such as length direction can be chosen and be more than 30 pixels, length-width ratios more than 60 pixels, width direction
Cutting image more than 2.It is understood that screening when can also according to clarity, integrality of image of image etc. because
Usually screen.
Simultaneously in order to improve subsequent arithmetic speed, preset quantity can be chosen from the cutting image collection after screening
Image is cut, forms new image collection.Above-mentioned preset quantity can be set according to practical application, the present embodiment pair
This is not limited.Such as can be 10 or 20 representative images.
The target retrieval method that above-described embodiment of the application provides reduces the size of the image of each target, simultaneously
Reduce the quantity of image in each image collection, subsequent operation efficiency can be effectively improved.
With further reference to Fig. 4, as the realization to method shown in above-mentioned each figure, this application provides a kind of target retrieval dresses
The one embodiment put, the device embodiment is corresponding with embodiment of the method shown in FIG. 1, which specifically can be applied to respectively
In kind electronic equipment.
As shown in figure 4, the target retrieval device 400 of the present embodiment includes：Image collection acquiring unit 401, feature vector
Extraction unit 402, similarity determining unit 403 and probability determining unit 404.
Wherein, image collection acquiring unit 401, for obtaining multiple images set.
Wherein, each image collection includes at least one image for containing at least one target to be checked.
Characteristic vector pickup unit 402 is treated for extract that at least part image in multiple images set included respectively
Examine clarification of objective vector.
Similarity determining unit 403, at least part feature vector of extraction, determine respectively each two feature to
The similarity in image collection belonging to the target to be checked of the similarity of amount and the similarity in two feature vector instructions
Ranking information.
Probability determining unit 404, for according to each similarity and each similarity ranking information, determining each two feature vector
Indicated target to be checked is the probability of same target.
In some optional realization methods of the present embodiment, above device 400 can also include unshowned in Fig. 4 carry
Show information generating unit, in response to there are probability satisfaction of the target to be checked indicated by two feature vectors for same target
First preset condition generates to inquire whether two feature vectors indicate the prompt message of same target.
In some optional realization methods of the present embodiment, above-mentioned first preset condition includes at least one of：On
Probability is stated more than predetermined probability threshold value；Above-mentioned probability is located at the preceding predetermined ratio range that all probability sort from big to small.
In some optional realization methods of the present embodiment, above-mentioned image collection derives from video source.Features described above to
Amount extraction unit 402 may further include unshowned identification information acquisition module in Fig. 4, compare List Generating Module and
Characteristic vector pickup module.
Identification information acquisition module, for obtaining the identification information of each image collection.
Above-mentioned identification information includes the first flag information of the video source belonging to the image collection and the image collection exists
Second identifier information in affiliated video source.
List Generating Module is compared, for the identification information of each two image collection to be compared, generation compares list.
Characteristic vector pickup module, for extracting the mesh to be checked that at least part image in each image collection is included respectively
Target feature vector, and determine to compare the similarity each compared in list between indicated feature vector.
In some optional realization methods of the present embodiment, the image in each image collection includes same mesh to be checked
Mark.Features described above vector extraction unit 402 can also include unshowned comparison list optimization module in Fig. 4.Above-mentioned comparison row
Table optimization module is used for：For each comparison in above-mentioned comparison list, two image collections indicated by the comparison are determined
Whether first flag information is identical；It is identical in response to the first flag information of two image collections indicated by the comparison, it determines
With the presence or absence of the image that the generation moment is identical in two image collections indicated by the comparison；And/or in response to comparison meaning
In the presence of the image that the generation moment is identical in two image collections shown, then the comparison is deleted, to optimize above-mentioned comparison list.
In some optional realization methods of the present embodiment, above-mentioned comparison list optimization module is used for：According to each figure
The generation moment of every image during image set closes determines the average generation moment of each image collection；For in above-mentioned comparison list
Each comparison, determine the comparison instruction two image collections the average generation moment difference whether be more than preset duration；It rings
The difference at the average generation moment of two image collections that should be indicated in the comparison is more than preset duration, the comparison is deleted, with optimization
Above-mentioned comparison list.
In some optional realization methods of the present embodiment, features described above vector extraction unit 402 can also be further
For：Extract every figure in above-mentioned at least part image of above-mentioned multiple images set respectively using preset first nerves network
As the clarification of objective to be checked vector included.
In some optional realization methods of the present embodiment, above-mentioned similarity determining unit 403 may further include
Unshowned similarity determining module, similarity set determining module and similarity set ranking module in Fig. 4.
Wherein, similarity determining module, for determine to compare in list each compare two indicated feature vectors it
Between similarity.
Similarity set determining module, for determining the similarity set of each target to be checked, wherein, above-mentioned similarity collection
It closes and includes comparing the corresponding similarity of comparison for containing the target to be checked in list.
Similarity set ranking module, for by each similarity in similarity set according to it is descending or by it is small to
Big sequence is arranged, and determines similarity ranking information of each target to be checked in respectively affiliated image collection.
In some optional realization methods of the present embodiment, above-mentioned probability determining unit 404 can be further used for：It will
The target to be checked of similarity and two feature vector instructions between each two feature vector is in respective sequencing of similarity
In ranking information input preset grader, the target to be checked indicated by each two feature vector is determined based on above-mentioned grader
Probability for same target.
In some optional realization methods of the present embodiment, above device 400 can also include in Fig. 4 unshowned point
Class device establishes unit.The grader establishes unit for pre-establishing grader, including mark image collection acquisition module, phase
Like degree determining module, sequencing of similarity determining module, ranking determining module and training module.
Wherein, image collection acquisition module is marked, for obtaining the multiple images set marked in advance.Wherein, mark
Information is used to represent that multiple images set includes whether two targets are same target.
Similarity determining module for extracting the clarification of objective vector that each image collection is included, and determines every two
Similarity between a feature vector.
Sequencing of similarity determining module, for according to the similarity between each two feature vector, determine each feature to
Sequencing of similarity between amount and other feature vectors.
Ranking determining module, for according to above-mentioned sequencing of similarity, determining each two feature vector in respective similarity
Ranking in sequence.
Training module, for the markup information according to multiple images set, the similarity between each two feature vector with
And above-mentioned ranking training grader.
In some optional realization methods of the present embodiment, above-mentioned image collection acquiring unit 401 can be wrapped further
Include unshowned video source acquisition module, module of target detection and target labeling module in Fig. 4.
Wherein, video source acquisition module, for obtaining multiple video sources, each video source includes at least one image set
It closes.
Module of target detection, for detecting at least part in above-mentioned multiple video sources using preset nervus opticus network
Image determines the target to be checked that every image in above-mentioned at least part image is included.
Target labeling module for being labeled to each target to be checked detected, obtains and each target pair to be checked
The mark image collection answered.
In some optional realization methods of the present embodiment, above-mentioned target labeling module can be further used for：It utilizes
Minimum enclosed rectangle frame is labeled each target to be checked detected；Each tab area is cut, is obtained and each mesh to be checked
Mark corresponding cutting image collection.
In some optional realization methods of the present embodiment, above-mentioned target labeling module can be further used for：It determines
The quantity of the cutting image included in each above-mentioned cutting image collection and the following parameter of every cutting image：Along first party
To the first pixel quantity, the second pixel quantity in a second direction, above-mentioned first pixel quantity and above-mentioned second pixel quantity
Ratio, wherein, above-mentioned first direction and above-mentioned second direction are respectively two adjacent length of sides of above-mentioned minimum enclosed rectangle
Extending direction；Each preset quantity cutting image for cutting and meeting the second preset condition in image collection is chosen, is formed new
Image collection.
The target retrieval device that above-described embodiment of the application provides, it is first determined then multiple images set is extracted more
At least part image is included in a image collection clarification of objective to be checked vector, then at least part feature of extraction to
In amount, the target institute to be checked that the similarity of each two feature vector and each similarity are indicated in two feature vectors is determined
Similarity ranking information in the image collection of category finally according to each similarity and each similarity ranking information, determines every two
Target to be checked indicated by a feature vector is the probability of same target.In this way, it can be obtained by different figures using electronic equipment
Target to be checked as in is the probability of same target, can realize the preliminary search to target, reduce the work of mark personnel
Amount improves the efficiency of target retrieval.
It should be appreciated that target retrieval device 400 described in unit 401 to unit 404 respectively with reference to described in figure 1
Each step in method is corresponding.It is equally applicable to fill above with respect to the operation and feature of the description of target retrieval method as a result,
400 and unit wherein included are put, details are not described herein.The corresponding units of device 400 can be mutual with the unit in server
Coordinate the scheme to realize the embodiment of the present application.
The embodiment of the present invention additionally provides a kind of electronic equipment, such as can be mobile terminal, personal computer (PC), put down
Plate computer, server etc..Below with reference to Fig. 5, it illustrates suitable for being used for realizing the terminal device of the embodiment of the present application or service
The structure diagram of the electronic equipment 500 of device：As shown in figure 5, computer system 500 includes one or more processors, communication
Portion etc., said one or multiple processors are for example：One or more central processing unit (CPU) 501 and/or one or more
Image processor (GPU) 513 etc., processor can according to the executable instruction being stored in read-only memory (ROM) 502 or
From the executable instruction that storage section 508 is loaded into random access storage device (RAM) 503 perform various appropriate actions and
Processing.Communication unit 512 may include but be not limited to network interface card, and above-mentioned network interface card may include but be not limited to IB (Infiniband) network interface card.
Processor can communicate to perform executable instruction with read-only memory 502 and/or random access storage device 503, lead to
It crosses bus 504 with communication unit 512 to be connected and communicate with other target devices through communication unit 512, so as to complete the embodiment of the present application
The corresponding operation of any one method of offer, for example, multiple images set is obtained, wherein, each image collection includes at least one
The image of at least one targets to be checked of Zhang Hanyou；Extract respectively at least part image in above-mentioned multiple images set included it is to be checked
Clarification of objective vector；In at least part feature vector of extraction, determine respectively each two feature vector similarity and
The similarity ranking information in image collection belonging to target to be checked of the similarity in two feature vector instructions；According to each
Similarity and each similarity ranking information determine the probability that the target to be checked indicated by each two feature vector is same target.
In addition, in RAM 503, it can also be stored with various programs and data needed for device operation.CPU 501、ROM
502 and RAM 503 is connected with each other by bus 504.In the case where there is RAM 503, ROM 502 is optional module.RAM
Executable instruction is written in 503 storage executable instructions into ROM 502 at runtime, and executable instruction holds processor 501
The corresponding operation of the above-mentioned communication means of row.Input/output (I/O) interface 505 is also connected to bus 504.Communication unit 512 can collect
Into setting, may be set to be with multiple submodule (such as multiple IB network interface cards), and in bus link.
I/O interfaces 505 are connected to lower component：Importation 506 including keyboard, mouse etc.；It is penetrated including such as cathode
The output par, c 507 of spool (CRT), liquid crystal display (LCD) etc. and loud speaker etc.；Storage section 508 including hard disk etc.；
And the communications portion 509 of the network interface card including LAN card, modem etc..Communications portion 509 via such as because
The network of spy's net performs communication process.Driver 510 is also according to needing to be connected to I/O interfaces 505.Detachable media 511, such as
Disk, CD, magneto-optic disk, semiconductor memory etc. are mounted on driver 510, as needed in order to be read from thereon
Computer program be mounted into storage section 508 as needed.
Need what is illustrated, framework as shown in Figure 5 is only a kind of optional realization method, can root during concrete practice
The component count amount and type of above-mentioned Fig. 5 are selected, are deleted, increased or replaced according to actual needs；It is set in different function component
Put, can also be used it is separately positioned or integrally disposed and other implementations, such as GPU and CPU separate setting or can be by GPU collection
Into on CPU, communication unit separates setting, can also be integrally disposed on CPU or GPU, etc..These interchangeable embodiments
Each fall within protection domain disclosed by the invention.
Particularly, in accordance with an embodiment of the present disclosure, it may be implemented as computer above with reference to the process of flow chart description
Software program.For example, embodiment of the disclosure includes a kind of computer program product, it is machine readable including being tangibly embodied in
Computer program on medium, computer program are included for the program code of the method shown in execution flow chart, program code
It may include the corresponding instruction of corresponding execution method and step provided by the embodiments of the present application, for example, multiple images set is obtained,
In, each image collection includes at least one image for containing at least one target to be checked；Above-mentioned multiple images set is extracted respectively
The clarification of objective to be checked vector that middle at least part image is included；In at least part feature vector of extraction, determine respectively
Image set belonging to the target to be checked of the similarity of each two feature vector and the similarity in two feature vector instructions
Similarity ranking information in conjunction；According to each similarity and each similarity ranking information, determine indicated by each two feature vector
Target to be checked be same target probability.In such embodiments, the computer program can by communications portion 509 from
It is downloaded and installed on network and/or is mounted from detachable media 511.In the computer program by central processing unit
(CPU) during 501 execution, the above-mentioned function of being limited in the present processes is performed.
Methods and apparatus of the present invention, equipment may be achieved in many ways.For example, software, hardware, firmware can be passed through
Or any combinations of software, hardware, firmware realize methods and apparatus of the present invention, equipment.The step of for method
Sequence is stated merely to illustrate, the step of method of the invention is not limited to sequence described in detail above, unless with other
Mode illustrates.In addition, in some embodiments, the present invention can be also embodied as recording program in the recording medium, this
A little programs include being used to implement machine readable instructions according to the method for the present invention.Thus, the present invention also covering stores to hold
The recording medium of the program of row according to the method for the present invention.
Description of the invention provides for the sake of example and description, and is not exhaustively or will be of the invention
It is limited to disclosed form.Many modifications and variations are obvious for the ordinary skill in the art.It selects and retouches
It states embodiment and is to more preferably illustrate the principle of the present invention and practical application, and those of ordinary skill in the art is enable to manage
The solution present invention is so as to design the various embodiments with various modifications suitable for special-purpose.
Claims (10)
- A kind of 1. target retrieval method, which is characterized in that including：Multiple images set is obtained, wherein, each image collection includes at least one image for containing at least one target to be checked；The clarification of objective to be checked vector that at least part image is included in described multiple images set is extracted respectively；In at least part feature vector of extraction, the similarity of determining each two feature vector and the similarity are at this respectively The similarity ranking information in image collection belonging to the target to be checked of two feature vector instructions；According to each similarity and each similarity ranking information, it is same mesh to determine the target to be checked indicated by each two feature vector Target probability.
- 2. according to the method described in claim 1, it is characterized in that, the method further includes：It is raw in response to meeting the first preset condition for the probability of same target there are the target to be checked indicated by two feature vectors Into for inquiring whether two feature vectors indicate the prompt message of same target.
- 3. according to the method described in claim 2, it is characterized in that, first preset condition includes at least one of：Institute Probability is stated more than predetermined probability threshold value；The probability is located at the preceding predetermined ratio range that all probability sort from big to small.
- 4. according to claim 1-3 any one of them methods, which is characterized in that described image set derives from video source；With AndIt is described to extract the clarification of objective to be checked vector that at least part image is included in described multiple images set, packet respectively It includes：The identification information of each image collection is obtained, the identification information includes the first mark of the video source belonging to the image collection Know the second identifier information of information and the image collection in affiliated video source；The identification information of each two image collection is compared, generation compares list；The clarification of objective to be checked vector that at least part image in each image collection is included is extracted respectively, and determines the ratio To each comparing the similarity between indicated feature vector in list.
- 5. according to claim 1-4 any one of them methods, which is characterized in that described to extract described multiple images set respectively The clarification of objective to be checked vector that middle at least part image is included, including：Extract every figure at least part image of described multiple images set respectively using preset first nerves network As the clarification of objective to be checked vector included.
- 6. according to the method described in claim 4, it is characterized in that, the similarity for determining each two feature vector respectively with And the similarity ranking information in the image collection belonging to target to be checked of the similarity in two feature vector instructions, packet It includes：It determines each to compare the similarity between two indicated feature vectors in the comparison list；Determine the similarity set of each target to be checked, wherein, the similarity set includes containing in the comparison list The corresponding similarity of comparison of the target to be checked；Each similarity in the similarity set according to descending or ascending sequence is arranged, is determined every Similarity ranking information of a target to be checked in respectively affiliated image collection.
- 7. according to the method described in claim 6, it is characterized in that, described according to each similarity and each similarity ranking information, Determine the probability that the target to be checked indicated by each two feature vector is same target, including：By the similarity between each two feature vector and the target to be checked of two feature vector instructions respective similar Ranking information in degree sequence inputs preset grader, determines to treat indicated by each two feature vector based on the grader Examine the probability that target is same target.
- 8. a kind of target retrieval device, which is characterized in that described device includes：Image collection acquiring unit, for obtaining multiple images set, wherein, each image collection includes at least one containing extremely The image of a few target to be checked；Characteristic vector pickup unit, for extracting the mesh to be checked that at least part image in described multiple images set is included respectively Target feature vector；Similarity determining unit, at least part feature vector of extraction, determining the phase of each two feature vector respectively Similarity ranking letter in the image collection belonging to target to be checked indicated like degree and the similarity in two feature vectors Breath；Probability determining unit, for according to each similarity and each similarity ranking information, determining indicated by each two feature vector Target to be checked be same target probability.
- 9. a kind of electronic equipment, which is characterized in that including：One or more processors；Storage device, for storing one or more programs,When one or more of programs are performed by one or more of processors so that one or more of processors are real The now method as described in any in claim 1-7.
- 10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is by processor The method as described in any in claim 1-7 is realized during execution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710500550.3A CN108229289B (en) | 2017-06-27 | 2017-06-27 | Target retrieval method and device and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710500550.3A CN108229289B (en) | 2017-06-27 | 2017-06-27 | Target retrieval method and device and electronic equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108229289A true CN108229289A (en) | 2018-06-29 |
| CN108229289B CN108229289B (en) | 2021-02-05 |
Family
ID=62657338
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710500550.3A Active CN108229289B (en) | 2017-06-27 | 2017-06-27 | Target retrieval method and device and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108229289B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110209895A (en) * | 2019-06-06 | 2019-09-06 | 阿里巴巴集团控股有限公司 | Vector index method, apparatus and equipment |
| CN110929545A (en) * | 2018-09-19 | 2020-03-27 | 传线网络科技（上海）有限公司 | Human face image sorting method and device |
| CN110929546A (en) * | 2018-09-19 | 2020-03-27 | 传线网络科技（上海）有限公司 | Face comparison method and device |
| CN110955792A (en) * | 2019-12-13 | 2020-04-03 | 云粒智慧科技有限公司 | Searching method and device based on picture, electronic equipment and storage medium |
| WO2021203823A1 (en) * | 2020-04-10 | 2021-10-14 | Oppo广东移动通信有限公司 | Image classification method and apparatus, storage medium, and electronic device |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7130849B2 (en) * | 2002-02-05 | 2006-10-31 | Hitachi, Ltd. | Similarity-based search method by relevance feedback |
| US20140032580A1 (en) * | 2011-03-11 | 2014-01-30 | Amco Llc | Relativistic concept measuring system for data clustering |
| US8831358B1 (en) * | 2011-11-21 | 2014-09-09 | Google Inc. | Evaluating image similarity |
| CN104063719A (en) * | 2014-06-27 | 2014-09-24 | 深圳市赛为智能股份有限公司 | Method and device for pedestrian detection based on depth convolutional network |
| CN105095884A (en) * | 2015-08-31 | 2015-11-25 | 桂林电子科技大学 | Pedestrian recognition system and pedestrian recognition processing method based on random forest support vector machine |
| CN106354735A (en) * | 2015-07-22 | 2017-01-25 | 杭州海康威视数字技术股份有限公司 | Image target searching method and device |
| CN106682095A (en) * | 2016-12-01 | 2017-05-17 | 浙江大学 | Subjectterm and descriptor prediction and ordering method based on diagram |
| CN106778464A (en) * | 2016-11-09 | 2017-05-31 | 深圳市深网视界科技有限公司 | A kind of pedestrian based on deep learning recognition methods and device again |
-
2017
- 2017-06-27 CN CN201710500550.3A patent/CN108229289B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7130849B2 (en) * | 2002-02-05 | 2006-10-31 | Hitachi, Ltd. | Similarity-based search method by relevance feedback |
| US20140032580A1 (en) * | 2011-03-11 | 2014-01-30 | Amco Llc | Relativistic concept measuring system for data clustering |
| US8831358B1 (en) * | 2011-11-21 | 2014-09-09 | Google Inc. | Evaluating image similarity |
| CN104063719A (en) * | 2014-06-27 | 2014-09-24 | 深圳市赛为智能股份有限公司 | Method and device for pedestrian detection based on depth convolutional network |
| CN106354735A (en) * | 2015-07-22 | 2017-01-25 | 杭州海康威视数字技术股份有限公司 | Image target searching method and device |
| CN105095884A (en) * | 2015-08-31 | 2015-11-25 | 桂林电子科技大学 | Pedestrian recognition system and pedestrian recognition processing method based on random forest support vector machine |
| CN106778464A (en) * | 2016-11-09 | 2017-05-31 | 深圳市深网视界科技有限公司 | A kind of pedestrian based on deep learning recognition methods and device again |
| CN106682095A (en) * | 2016-12-01 | 2017-05-17 | 浙江大学 | Subjectterm and descriptor prediction and ordering method based on diagram |
Non-Patent Citations (2)
| Title |
|---|
| SUH-YIN LEE等: "Spatial reasoning and similarity retrieval of images using 2D C-string knowledge representation", 《PATTERN RECOGNITION》 * |
| 石宽: "基于多相似度融合图像检索技术研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110929545A (en) * | 2018-09-19 | 2020-03-27 | 传线网络科技（上海）有限公司 | Human face image sorting method and device |
| CN110929546A (en) * | 2018-09-19 | 2020-03-27 | 传线网络科技（上海）有限公司 | Face comparison method and device |
| CN110209895A (en) * | 2019-06-06 | 2019-09-06 | 阿里巴巴集团控股有限公司 | Vector index method, apparatus and equipment |
| CN110209895B (en) * | 2019-06-06 | 2023-09-05 | 创新先进技术有限公司 | Vector retrieval method, device and equipment |
| CN110955792A (en) * | 2019-12-13 | 2020-04-03 | 云粒智慧科技有限公司 | Searching method and device based on picture, electronic equipment and storage medium |
| WO2021203823A1 (en) * | 2020-04-10 | 2021-10-14 | Oppo广东移动通信有限公司 | Image classification method and apparatus, storage medium, and electronic device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108229289B (en) | 2021-02-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107341805B (en) | Background segment and network model training, image processing method and device before image | |
| CN108229289A (en) | Target retrieval method, apparatus and electronic equipment | |
| CN108304835A (en) | character detecting method and device | |
| CN109815868A (en) | A kind of image object detection method, device and storage medium | |
| JP6779641B2 (en) | Image classification device, image classification system and image classification method | |
| CN110458154A (en) | Face identification method, device and computer readable storage medium | |
| CN109978918A (en) | A kind of trajectory track method, apparatus and storage medium | |
| CN109359666A (en) | A kind of model recognizing method and processing terminal based on multiple features fusion neural network | |
| CN108280477A (en) | Method and apparatus for clustering image | |
| CN108875676A (en) | Biopsy method, apparatus and system | |
| CN108229341A (en) | Sorting technique and device, electronic equipment, computer storage media, program | |
| CN108229509A (en) | For identifying object type method for distinguishing and device, electronic equipment | |
| CN108090508A (en) | A kind of classification based training method, apparatus and storage medium | |
| CN106909870A (en) | The search method and device of facial image | |
| CN109284729A (en) | Method, apparatus and medium based on video acquisition human face recognition model training data | |
| CN108229344A (en) | Image processing method and device, electronic equipment, computer program and storage medium | |
| CN108171191B (en) | Method and apparatus for detecting face | |
| CN109644255A (en) | Mark includes the method and apparatus of the video flowing of a framing | |
| CN103164687B (en) | A kind of method and system of pornographic image detecting | |
| CN109214280A (en) | Shop recognition methods, device, electronic equipment and storage medium based on streetscape | |
| CN109389589A (en) | Method and apparatus for statistical number of person | |
| CN107730515A (en) | Panoramic picture conspicuousness detection method with eye movement model is increased based on region | |
| CN109472193A (en) | Method for detecting human face and device | |
| CN110032916A (en) | A kind of method and apparatus detecting target object | |
| CN109145828A (en) | Method and apparatus for generating video classification detection model |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |