CN108491528A - A kind of image search method, system and device - Google Patents
A kind of image search method, system and device Download PDFInfo
- Publication number
- CN108491528A CN108491528A CN201810264981.9A CN201810264981A CN108491528A CN 108491528 A CN108491528 A CN 108491528A CN 201810264981 A CN201810264981 A CN 201810264981A CN 108491528 A CN108491528 A CN 108491528A
- Authority
- CN
- China
- Prior art keywords
- image
- hash
- classification
- retrieved
- model
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
-
- 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
Abstract
This application discloses a kind of image search methods, system and device, including image to be retrieved classification judgement is carried out via integral retrieval model, it is that image to be retrieved chooses corresponding local search model according to the classification judging result of block mold, image to be retrieved carries out signature analysis by corresponding local search model and carries out Hash coding mapping, obtain the Hash coding of image to be retrieved, simultaneously, the Hash of retrieval and image to be retrieved encodes similar Hash and encodes in target Hash code database corresponding with local search model, finally therefrom choose the similar Hash coded set for meeting condition of similarity, and similar diagram image set corresponding with similar Hash coded set is found in image data base, complete retrieval；The application pre-establishes the integral retrieval model for carrying out classification judgement, integral retrieval model is recycled to generate local search model, the retrieval accuracy to image is improved, and establishes target Hash code database corresponding with local search model, further improves retrieval accuracy.
Description
Technical field
The present invention relates to field of image search, more particularly to a kind of image search method, system and device.
Background technology
With the arrival in big data epoch, the Internet images resource rapidly increases, and the collection of image data resource neutralizes scale
Increase to bring to image retrieval and also brings challenge while opportunity.Compared to traditional text based image retrieval management and
Inquiry, it is difficult, it is difficult to meet actual requirement.Therefore, content-based image retrieval (Content-Based Image
Retrieval, CBIR) by the extensive research of many scholars.The characteristic information of image how is effectively described, and what is used
A series of problems, such as kind data structure carries out efficient index and Fast Similarity Retrieval becomes the research hotspot in this direction.
In the prior art, it is proposed that a kind of supervision hash method CNNH and CNNH+, this method is first training image data
Pairs of semantic similarity matrix factorisation is encoded at approximate Hash, then utilizes these approximate Hash codings and image tag
One depth convolutional network of training, but the matrix decomposition in CNNH and CNNH+ can bring additional mistake so that training objective
Deviate；And another hash method DLBHC for being directly based upon CNN, one full articulamentum two-value is turned to binary system using threshold value
As a result, to obtain Hash coding；But above method retrieval effectiveness on the image being difficult to differentiate between is still very poor.
A kind of therefore, it is necessary to retrieval effectiveness more preferable, the higher image search method of precision.
Invention content
In view of this, the purpose of the present invention is to provide a kind of image search method, system and device, image retrieval is improved
Effect and precision.Its concrete scheme is as follows：
A kind of image search method, including：
Image category retrieval is carried out to image to be retrieved using integral retrieval model, obtains the first of the image to be retrieved
Tag along sort；
Image category is carried out to the image to be retrieved using local search model corresponding with first tag along sort
Retrieval, obtains the second tag along sort；
Judge whether first tag along sort and second tag along sort are consistent；
If it is, carrying out Hash coding mapping to the image to be retrieved using the local search model, institute is obtained
State the Hash coding of image to be retrieved；
It is searched in target Hash code database corresponding with the local search model full with Hash coding similarity
The similar Hash coded set of the preset condition of similarity of foot；
Corresponding similar diagram image set is found in image data base using the similar Hash coded set；
Wherein, the integral retrieval model be using training image collection, it is initial whole to being established based on convolutional neural networks
Body retrieval model is trained, and the training image collection includes the image category of all images to be retrieved；
The local search model is using training image subset corresponding with classification subset, to being based on convolutional neural networks
What the initial local retrieval model of foundation was trained；Wherein, using the integral retrieval model to each of verification collection
Authentication image carries out image category retrieval, obtains the tag along sort for recording each authentication image, and generation records each authentication image
Tag along sort classification results collection, from the classification results concentrate filter out meet preset class condition the classification son
Collection；Training image corresponding with image category in classification subset is filtered out from training image concentration, obtains the training figure
As subset；
The target Hash code database preserves the local search model to each instruction in the training image subset
Practice the Hash coding that image carries out each training image of Hash coding mapping generation.
Optionally, described concentrated from the classification results filters out the classification subset for meeting preset condition of similarity
Process, including：
Using each of described in standard picture classification and the classification results collection of each authentication image counted in advance
The tag along sort of authentication image compares one by one, obtains the classification accuracy collection for recording every class image classification accuracy rate；
The classification accuracy per class image is concentrated to be arranged in order according to descending order the classification accuracy, successively
It brings the adjacent accuracy rate of image classification two-by-two into classification threshold formula, calculates classification thresholds；
Adjacent image classification accuracy rate using the classification thresholds more than predetermined threshold value is as cut-point, by the classification
Accuracy collection is divided into multiple classification subsets；Wherein,
The classification thresholds formula is：
In formula, i*Indicate the classification thresholds,Indicate siThe classification accuracy of class,Indicate si-1The classification of class
Accuracy.
Optionally, described that Hash coding mapping is carried out to the image to be retrieved using the local search model, it obtains
The process of the Hash coding of the image to be retrieved, including：
Output and the Hash mapping for mapping full articulamentum in the local search model using the image to be retrieved are public
Formula carries out Hash coding mapping to the image to be retrieved, obtains the Hash coding of the image to be retrieved；Wherein,
The Hash mapping formula is：
In formula,Indicate j-th of coding in the Hash coding of i-th image,Indicate that i-th image connects entirely in mapping
Meet j-th of output valve, θ in the output of layer1For the mapping threshold value of model.
Optionally, described searched in target Hash code database corresponding with the local search model is compiled with the Hash
Code similarity meets the process of the similar Hash coding of preset condition of similarity, including：
It is searched in the target Hash code database similar to the predetermined number of Hash coding Hamming distance minimum
Hash encodes, and obtains similar Hash coded set.
Optionally, it is described judge whether first tag along sort and second tag along sort consistent after, further include：
When first tag along sort and second tag along sort are inconsistent, then utilize the integral retrieval model to institute
It states image to be retrieved and carries out Hash coding mapping, obtain the Hash coding of the image to be retrieved；
It is searched in whole Hash code database corresponding with the integral retrieval model full with Hash coding similarity
The similar Hash coded set of the foot condition of similarity；
Wherein, the whole Hash code database is to carry out Hash to the training image collection using the integral retrieval model
What coding mapping obtained.
The invention also discloses a kind of image indexing systems, including：
Whole sort module obtains institute for carrying out image category retrieval to image to be retrieved using integral retrieval model
State the first tag along sort of image to be retrieved；
Local sort module, for utilizing local search model corresponding with first tag along sort to described to be retrieved
Image carries out image category retrieval, obtains the second tag along sort；
Classification judgment module, for judging whether first tag along sort and second tag along sort are consistent；
Hash mapping module, for judging that first tag along sort is classified with described second when the classification judgment module
Label is consistent, then carries out Hash coding mapping to the image to be retrieved using the local search model, obtain described to be checked
The Hash of rope image encodes；
Hash retrieve module, in target Hash code database corresponding with the local search model search with it is described
Hash coding similarity meets the similar Hash coded set of preset condition of similarity；
Image retrieval module, it is corresponding similar for being found in image data base using the similar Hash coded set
Image set；
Block mold generation module, for utilizing training image collection, to the initial entirety established based on convolutional neural networks
Retrieval model is trained to obtain integral retrieval model；Wherein, the training image collection includes the image of all images to be retrieved
Classification；
Partial model generation module, for utilizing training image subset corresponding with classification subset, to being based on convolutional Neural
The initial local retrieval model that network is established is trained to obtain local search model；
Wherein, the partial model generation module, including：
Classification generation unit, for carrying out image class to each authentication image of verification collection using the integral retrieval model
It does not retrieve, obtains the tag along sort for recording each authentication image, generate the classification knot for the tag along sort for recording each authentication image
Fruit collects；
Subset screening unit filters out the classification for meeting preset class condition for being concentrated from the classification results
Subset；
Optical sieving unit, for filtering out instruction corresponding with image category in classification subset from training image concentration
Practice image, obtains the training image subset；
Hash library generation module preserves the local search model to the training for the target Hash code database
Each training image in image subset carries out the Hash coding of each training image of Hash coding mapping generation.
Optionally, the subset screening unit, including：
Contrast subunit, the standard picture classification for utilizing each authentication image counted in advance and the classification results
The tag along sort of each authentication image described in collection compares one by one, and the classification for obtaining recording every class image classification accuracy rate is accurate
Degree collection；
Threshold value subelement, for concentrating the classification accuracy per class image according to suitable from big to small the classification accuracy
Sequence is arranged in order, and is brought the adjacent accuracy rate of image classification two-by-two into classification threshold formula successively, is calculated classification thresholds；
Divide subelement, for using the classification thresholds be more than predetermined threshold value adjacent image classification accuracy rate as divide
The classification accuracy collection is divided into multiple classification subsets by cutpoint；Wherein,
The classification thresholds formula is：
In formula, i*Indicate the classification thresholds,Indicate siThe classification accuracy of class,Indicate si-1The classification of class
Accuracy.
Optionally, the Hash mapping module is specifically used for using the image to be retrieved in the local search model
The output of the middle full articulamentum of mapping and Hash mapping formula carry out Hash coding mapping to the image to be retrieved, obtain described
The Hash of image to be retrieved encodes；Wherein,
The Hash mapping formula is：
In formula,Indicate j-th of coding in the Hash coding of i-th image,Indicate that i-th image connects entirely in mapping
Meet j-th of output valve, θ in the output of layer1For the mapping threshold value of model.
Optionally, further include：
Whole Hash mapping module, it is inconsistent for working as first tag along sort and second tag along sort, then it is sharp
Hash coding mapping is carried out to the image to be retrieved with the integral retrieval model, obtains the Kazakhstan of the image to be retrieved
Uncommon coding；
Whole Hash retrieves module, for searched in whole Hash code database corresponding with the integral retrieval model with
The Hash coding similarity meets the similar Hash coded set of the condition of similarity；
Wherein, the whole Hash code database is to carry out Hash to the training image collection using the integral retrieval model
What coding mapping obtained.
The invention also discloses a kind of image retrieving apparatus, including：
Memory, for storing instruction；Wherein, described instruction carries out image to be retrieved including the use of integral retrieval model
Image category is retrieved, and the first tag along sort of the image to be retrieved is obtained；Utilize office corresponding with first tag along sort
Portion's retrieval model carries out image category retrieval to the image to be retrieved, obtains the second tag along sort；Judge first classification
Whether label and second tag along sort are consistent；If it is, using the local search model to the image to be retrieved
Hash coding mapping is carried out, the Hash coding of the image to be retrieved is obtained；In target corresponding with the local search model
The similar Hash coded set for meeting preset condition of similarity to Hash coding similarity is searched in Hash code database；Using institute
It states similar Hash coded set and finds corresponding similar diagram image set in image data base；Wherein, the integral retrieval model is
Utilize training image collection, to what is be trained based on the initial integral retrieval model that convolutional neural networks are established, the instruction
Practice the image category that image set includes all images to be retrieved；The local search model is to utilize instruction corresponding with classification subset
Practice image subset, the initial local retrieval model established based on convolutional neural networks is trained；Wherein, institute is utilized
It states integral retrieval model and image category retrieval is carried out to each authentication image of verification collection, obtain point for recording each authentication image
Class label generates the classification results collection for the tag along sort for recording each authentication image, is filtered out from classification results concentration full
The classification subset of the preset class condition of foot；It is filtered out and image category pair in classification subset from training image concentration
The training image answered obtains the training image subset；The target Hash code database preserves the local search model pair
Each training image in the training image subset carries out the Hash coding of each training image of Hash coding mapping generation；
Processor, for executing the instruction in the memory.
In the present invention, image search method, including image to be retrieved carry out big classification via integral retrieval model first and sentence
It is disconnected, it is that image to be retrieved chooses corresponding local search model according to the first tag along sort of classification judging result of block mold,
Range of search is reduced, retrieval precision is improved, image to be retrieved carries out signature analysis by corresponding local search model and obtains accurately
Higher second tag along sort of rate is further ensured that point by judging whether the first tag along sort and the second tag along sort are consistent
The accuracy of class, and carry out Hash coding mapping obtains the Hash coding of image to be retrieved, meanwhile, with local search model
The Hash of retrieval and image to be retrieved encodes similar Hash and encodes in corresponding target Hash code database, further to reduce inspection
Rope range improves retrieval precision, finally therefrom chooses the similar Hash coded set for meeting condition of similarity, and in image data base
Similar diagram image set corresponding with similar Hash coded set is found, the retrieval to image to be retrieved is completed；The present invention pre-establishes
The integral retrieval model of big classification judgement is carried out to image to be retrieved, and integral retrieval model is recycled to generate corresponding part image class
Other local search model improves the retrieval accuracy to image, and establishes target Hash corresponding with local search model and compile
Code library, improves the retrieval accuracy of similar image.
Description of the drawings
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is a kind of image search method flow diagram provided in an embodiment of the present invention；
Fig. 2 is a kind of image indexing system structural schematic diagram provided in an embodiment of the present invention.
Specific implementation mode
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation describes, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
Shown in Figure 1 the embodiment of the invention discloses a kind of image search method, this method includes：
S11：Image category retrieval is carried out to image to be retrieved using integral retrieval model, obtains the first of image to be retrieved
Tag along sort.
Specifically, when needing to find the image generic with figure to be retrieved from image data base, it will be to be retrieved
Image is input to the integral retrieval model being generated in advance, and integral retrieval model carries out image category retrieval to image to be retrieved, point
The image category of image to be retrieved is precipitated, and obtains the first tag along sort of image to be retrieved.
Wherein, integral retrieval model is the initial whole inspection to being established based on convolutional neural networks using training image collection
Rope model is trained；Training image collection is the pre-prepd set for recording and having each training image classification information,
Initial integral retrieval model may include convolutional layer, pond layer and full articulamentum, and training image collection includes all images to be retrieved
Image category, initial integral retrieval model is substituted by the training image for concentrating training image and obtains the initial of training image
Training result is compared the classification information for the training image recorded in advance using initial training result, corrects initial integral retrieval mould
The error of type presets error until meeting, obtains integral retrieval model.
It is understood that can the image of training image collection be divided into training set and verification collection, profit according to preset ratio
Model is trained with the image in training set, the image concentrated using verification verifies the model of generation, calculates accurate
True rate.
S12：Image category retrieval is carried out to image to be retrieved using local search model corresponding with the first tag along sort,
Obtain the second tag along sort.
Specifically, being retrieved again to image to be retrieved using local search model, it is higher precision can be obtained
The second tag along sort of retrieval result, meanwhile, if integral retrieval model leads to the first contingency table to image classification mistake to be retrieved
Image mistake to be retrieved is assigned to the local search model for not including image type to be retrieved by label, can also be according to follow-up first point
Whether class label is consistent with the second tag along sort, is judged and corrects mistake.
S13：Judge whether the first tag along sort is consistent with the second tag along sort.
Wherein, if the first tag along sort and the second tag along sort are inconsistent, the inspection to image to be retrieved can be terminated
Rope waits for user's intervention, can also re-execute S11, be judged again by integral retrieval model.
S14：If it is, carrying out Hash coding mapping to image to be retrieved using local search model, obtain to be retrieved
The Hash of image encodes.
Specifically, carrying out Hash volume using to the higher local search model of image retrieval feature extraction precision to be retrieved
Code mapping obtains the Hash coding of image to be retrieved；For example, the first tag along sort is aircraft, First partial retrieval model corresponds to
Type include cat, dog, frog, the corresponding type of the second local retrieval model includes bird, truck, aircraft, then utilizes second game
Portion's retrieval model carries out Hash coding mapping to image to be retrieved, obtains the Hash coding of image to be retrieved.
Wherein, parts of images type in local search model correspondence image database, therefore, to the parts of images type
Analysis degree higher, precision higher；Local search model is using training image subset corresponding with classification subset, to being based on
What the initial local retrieval model that convolutional neural networks are established was trained；Wherein, using integral retrieval model to verification
It concentrates each authentication image to carry out image category retrieval, obtains the tag along sort for recording each authentication image, it is each to generate record
The classification results collection of the tag along sort of authentication image is concentrated from classification results and filters out the classification for meeting preset class condition
Collection, preset class condition include filtering out classification accurately from classification results concentration according to the classification accuracy of integral retrieval model
Degree meets the classification subset of accuracy threshold value；Training corresponding with image category in classification subset is filtered out from training image concentration
Image obtains training image subset.
There is the classification of different images classification accurate it is understood that classification subset is the record divided according to classification accuracy
Therefore training image collection according to the image category information described in classification subset, is divided into multiple training images by the set of exactness
Subset, a training image subset may include a variety of image categories；Verification collection has each proof diagram for the record being generated in advance
The image set of the image category information of picture, verification, which integrates, to be the subset of training image collection.
S15：It searches in target Hash code database corresponding with local search model and meets in advance with Hash coding similarity
If condition of similarity similar Hash coded set.
Specifically, pre-set condition of similarity, when finding the Hash with image to be retrieved in target Hash code database
Coding similarity meets the Hash coding of condition of similarity, then is preserved as the element in similar Hash coded set, to
Meet the similar Hash coded set that the Hash of condition of similarity encodes to multiple similarities are preserved.
Wherein, target Hash code database preserve local search model to each training image in training image subset into
The Hash coding for each training image that row Hash coding mapping generates, therefore, the Hash coding of image to be retrieved can be in mesh
Mark Hash code database finds the Hash coding of similar image.
S16：Corresponding similar diagram image set is found in image data base using similar Hash coded set.
Specifically, after obtaining similar Hash coded set, in the image data base for preserving image, using Hash coding with
The mapping relations of image find similar diagram image set corresponding with similar Hash coded set, complete to image to be retrieved in image
The task of similar image is searched in database.
As it can be seen that the integral retrieval model that big classification judgement is carried out to image to be retrieved is pre-established in the embodiment of the present invention,
It recycles integral retrieval model to generate the local search model of corresponding part image category, improves the retrieval accuracy to image,
And target Hash code database corresponding with local search model is established, improve the retrieval accuracy of similar image；Image to be retrieved
Big classification judgement is carried out via integral retrieval model first, is to wait for according to the first tag along sort of classification judging result of block mold
It retrieves image and chooses corresponding local search model, reduce range of search, improve retrieval precision, image to be retrieved is by corresponding office
Portion's retrieval model carries out signature analysis and obtains higher second tag along sort of accuracy rate, by judging the first tag along sort and second
Whether tag along sort is consistent, further ensures that the accuracy of classification, and carry out Hash coding mapping, obtains the Kazakhstan of image to be retrieved
Uncommon coding, meanwhile, retrieval and the Hash of image to be retrieved encode in target Hash code database corresponding with local search model
Similar Hash coding, further reduces range of search, improves retrieval precision, finally therefrom choose the phase for meeting condition of similarity
Like Hash coded set, and similar diagram image set corresponding with similar Hash coded set is found in image data base, complete to treat
Retrieve the retrieval of image.
The embodiment of the invention discloses a kind of specific image search methods, relative to a upper embodiment, the present embodiment pair
Technical solution has made further instruction and optimization.Specifically：
In the embodiment of the present invention, above-mentioned concentrated from classification results filters out the classification subset for meeting preset condition of similarity
Detailed process, including：
S21：Using each of described in standard picture classification and the classification results collection of each authentication image counted in advance
The tag along sort of authentication image compares one by one, obtains the classification accuracy collection for recording every class image classification accuracy rate.
Specifically, classifying using each authentication image that whole verification model concentrates verification, each verification is obtained
The tag along sort of image, it is whether consistent with standard picture classification by comparing the tag along sort that whole verification model analysis goes out, it obtains
To record per the classification accuracy collection of class image classification accuracy rate.
S22：The accuracy that will classify concentrates the classification accuracy per class image to be arranged in order according to descending order, successively
It brings the adjacent accuracy rate of image classification two-by-two into classification threshold formula, calculates classification thresholds.
Specifically, be respectively arranged in order the classification accuracy of the every class image arranged according to descending order, and
It brings the adjacent accuracy rate of image classification two-by-two into classification threshold formula successively from the beginning to calculate, calculates classification threshold
Value.
Wherein, classification thresholds formula is：
In formula, i*Presentation class threshold value,Indicate siThe classification accuracy of class,Indicate si-1The classification of class is accurate
Degree.
S23：The adjacent image classification accuracy rate that classification thresholds are more than predetermined threshold value will classify accurately as cut-point
Degree collection is divided into multiple classification subsets.
Specifically, predetermined threshold value can be arranged according to historical experience, when point for calculating adjacent image classification accuracy rate
Class threshold value is more than predetermined threshold value then from when starting point is divided at the first two image classification accuracy rate seat, and the accuracy collection that will classify is divided into
Two parts, for example, predetermined threshold value is 7, classification accuracy is ordered as the first accuracy rate, the 4th accuracy rate, the second accuracy rate, third
Accuracy rate, the 6th accuracy rate and the 5th accuracy rate, classification thresholds are followed successively by 2,5,7,6,3, then the first accuracy rate, the 4th accuracy rate
It is a classification subset with the second accuracy rate, third accuracy rate, the 6th accuracy rate and the 5th accuracy rate are another classification subset.
In the embodiment of the present invention, above-mentioned S14 carries out Hash coding mapping using local search model to image to be retrieved, obtains
To the detailed process of the Hash coding of image to be retrieved, may include：
Output and the Hash mapping formula for mapping full articulamentum in local search model using image to be retrieved, to be checked
Rope image carries out Hash coding mapping, obtains the Hash coding of image to be retrieved；Wherein, it can be part inspection to map full articulamentum
The full articulamentum of rope model penultimate；
Wherein, Hash mapping formula is：
In formula,Indicate j-th of coding in the Hash coding of i-th image,Indicate that i-th image connects entirely in mapping
Meet j-th of output valve, θ in the output of layer1For the mapping threshold value of model；Wherein, full articulamentum is mapped in local search model
Export oiIt can be expressed asThe Hash coding h generated after mappingiIt can be expressed asK is code length.
It should be noted that being searched in target Hash code database corresponding with local search model similar to Hash coding
The detailed process that degree meets the similar Hash coding of preset condition of similarity is included in lookup and Hash in target Hash code database
The similar Hash of the predetermined number of the Hamming distance minimum of coding encodes, and obtains similar Hash coded set；By smallest hamming distance
As a preset condition, and the similar Hash of selection predetermined number Hamming distance minimum encodes from target Hash code database,
To ensure to find and the most similar generic image of image to be retrieved.
It is understood that since integral retrieval model can not ensure that 100% is correct in assorting process, it is thus possible to go out
Now by the first tag along sort classification error of image to be retrieved the case where, so as to cause image to be retrieved in wrong local search
Model carries out image retrieval, keeps the first tag along sort and the second tag along sort inconsistent, and is re-started by integral retrieval model
Image retrieval may not obtain correct first tag along sort, cause retrieval fail, but in image data base there may be with
Therefore the similar picture of image to be retrieved can be completed the Hash coding mapping to image to be retrieved by integral retrieval model, and
Entirety Hash code database corresponding with integral retrieval model is established, when the first tag along sort and the second tag along sort are inconsistent, then
Hash coding mapping to image to be retrieved is completed by integral retrieval model, and is searched in whole Hash code database, ensures inspection
The success rate of rope.
Specifically, when the first tag along sort and the second tag along sort are inconsistent, then utilize integral retrieval model to be retrieved
Image carries out Hash coding mapping, obtains the Hash coding of image to be retrieved；In whole Hash corresponding with integral retrieval model
The similar Hash coded set for meeting condition of similarity to Hash coding similarity is searched in code database；
Wherein, whole Hash code database is to carry out Hash coding mapping to training image collection using integral retrieval model to obtain
's.
In addition, the embodiment of the invention also discloses the concrete application scenes of image search method, specifically：
In the embodiment of the present invention, image data base includes the cromogram of 60000 32*32 using CIFAR-10 data sets
Picture shares 10 classes, can set the image collection in image data base and be expressed as X={ X1,X2,…,Xc, corresponding class label
It is expressed as Y={ 1,2 ..., c }, wherein c is the class number of image data base, and c=10 extracts to form instruction by the 90% of X
Practice collection Xtrain, remaining 10% as verification collection Xvalidation。
In embodiments of the present invention, the structure of the convolutional neural networks of integral retrieval model and local retrieval model can be equal
3 convolutional layers, 3 pond layers and 3 full articulamentums are set as, the size of convolution kernel is 3 × 3 wherein in convolutional layer, and step-length is
1, the pond window size of pond layer is 2 × 2, and step-length 1, the 1st layer of full articulamentum has 500 nodes, the second layer to have 48 sections
Point, last layer have 10 nodes.
Specifically, by using training set XtrainAs training image collection, first convolutional neural networks of training generate whole
Physical examination rope MODEL C NN1, verification is collected into XvalidationSubstitute into block mold CNN1, calculate XvalidationIn point per class authentication image
Class accuracy rate is denoted as CP={ p1,p2,…,pc, by CP={ p1,p2,…,pcElement in set arranged from big to small by its value
Sequence generates the set after a sequenceThat is classification accuracy collection, whereinRepresent siThe classification of class
Accuracy.
Specifically, CP={ 0.92,0.96,0.81,0.76,0.91,0.81,0.96,0.89,0.93,0.94 }；Using point
Class threshold formulaBy CPSIt is divided into two subsets, the respectively first classification subsetWith
Second classification subsetIt, can be using the first classification subset as X according to this divisiongood, indicate the first classification
Collect by the image construction of the classification of good classification effect, the second classification subset is then that remaining image category constitutes Xbad=X-
Xgood；In embodiments of the present invention, preset classification thresholds can be i*=7, thus Ygood={ 1,2,5,7,9,10,3 }, Ybad
={ 4,6,8 }.
Further, the first classification subset X can be utilizedgoodSecond convolutional neural networks of training, generate First partial
Retrieval model CNN2；Utilize the second classification subset XbadTraining third convolutional neural networks generate the second local retrieval model
CNN3。
Wherein it is possible to which the integral retrieval model of all images in image data base X is mapped integral Hash code database
Hglobal, concrete methods of realizing can be：I-th image x in XiIt is input to MODEL C NN1, record its layer output second from the bottomThe Hash coded representation generated after being mapped isMapping side
Formula is：
Wherein θ1For MODEL C NN1Mapping threshold value can take, θ1=7.
Similarly, respectively MODEL C NN2And CNN3Determine suitable mapping threshold θ2And θ3, by image subset XgoodAnd XbadPoint
It is not mapped to target Hash code database HgoodAnd Hbad, mapping threshold value can be respectively θ2=5, θ3=3.
Further, in the complete integral retrieval model of training and two local search models, and whole Hash is generated respectively and is compiled
After Ma Ku and two target Hash code database, for an image x to be retrieved, retrieved from image data base X most like
The detailed process of preceding T images is as follows.
MODEL C NN is used first1Obtain the first tag along sort y of xglobalIf yglobalIt is YgoodIn classification, then x is inputted
To MODEL C NN2, the tag along sort remembered is ygoodIf yglobalIt is YbadIn classification, then use MODEL C NN3To predict this
The classification of image, is denoted as ybad, i.e., image to be retrieved is substituted by corresponding local search model according to the first tag along sort.
It selects suitable Hash code database to be retrieved for x, enables the target Hash code database be：
With generation HgoalModel by x be mapped to Hash coding, be denoted as Hretrieval, for example, HgoalBy CNN2It generates, then
Utilize CNN2X is mapped to Hash coding.
Finally, lookup and H in corresponding target Hash code database or whole Hash code databaseretrievalWith the minimum Chinese
The preceding T of prescribed distance similar Hash codings find out similar diagram corresponding with preceding T similar Hash codings in image data base
Picture completes retrieval.
In addition, the embodiment of the invention also discloses a kind of image indexing system, shown in Figure 2, which includes：
Whole sort module 11 is obtained for carrying out image category retrieval to image to be retrieved using integral retrieval model
First tag along sort of image to be retrieved；
Local sort module 12, for using local search model corresponding with the first tag along sort to image to be retrieved into
Row image category is retrieved, and the second tag along sort is obtained；
Classification judgment module 13, for judging whether the first tag along sort is consistent with the second tag along sort；
Hash mapping module 14, for judging that the first tag along sort is consistent with the second tag along sort when classification judgment module,
Hash coding mapping then is carried out to image to be retrieved using local search model, obtains the Hash coding of image to be retrieved；
Hash retrieves module 15, is compiled with Hash for being searched in target Hash code database corresponding with local search model
Code similarity meets the similar Hash coded set of preset condition of similarity；
Image retrieval module 16, for finding corresponding similar diagram in image data base using similar Hash coded set
Image set；
Block mold generation module 17, it is initial whole to being established based on convolutional neural networks for utilizing training image collection
Body retrieval model is trained to obtain integral retrieval model；Wherein, training image collection includes the image class of all images to be retrieved
Not；
Partial model generation module 18, for utilizing training image subset corresponding with classification subset, to being based on convolution god
The initial local retrieval model established through network is trained to obtain local search model；
Wherein, partial model generation module 18, including：
Classification generation unit, for carrying out image category inspection to each authentication image of verification collection using integral retrieval model
Rope obtains the tag along sort for recording each authentication image, generates the classification results collection for the tag along sort for recording each authentication image；
Subset screening unit filters out the classification subset for meeting preset class condition for being concentrated from classification results；
Optical sieving unit is schemed for filtering out training corresponding with image category in classification subset from training image concentration
Picture obtains training image subset；
Hash library generation module 19 preserves local search model in training image subset for target Hash code database
Each training image carry out Hash coding mapping generation each training image Hash coding.
As it can be seen that the integral retrieval model that big classification judgement is carried out to image to be retrieved is pre-established in the embodiment of the present invention,
It recycles integral retrieval model to generate the local search model of corresponding part image category, improves the retrieval accuracy to image,
And target Hash code database corresponding with local search model is established, improve the retrieval accuracy of similar image；Image to be retrieved
Big classification judgement is carried out via integral retrieval model first, is to wait for according to the first tag along sort of classification judging result of block mold
It retrieves image and chooses corresponding local search model, reduce range of search, improve retrieval precision, image to be retrieved is by corresponding office
Portion's retrieval model carries out signature analysis and obtains higher second tag along sort of accuracy rate, by judging the first tag along sort and second
Whether tag along sort is consistent, further ensures that the accuracy of classification, and carry out Hash coding mapping, obtains the Kazakhstan of image to be retrieved
Uncommon coding, meanwhile, retrieval and the Hash of image to be retrieved encode in target Hash code database corresponding with local search model
Similar Hash coding, further reduces range of search, improves retrieval precision, finally therefrom choose the phase for meeting condition of similarity
Like Hash coded set, and similar diagram image set corresponding with similar Hash coded set is found in image data base, complete to treat
Retrieve the retrieval of image.
In the embodiment of the present invention, above-mentioned subset screening unit, including contrast subunit, threshold value subelement and segmentation are single
Member；Wherein,
Contrast subunit is concentrated for the standard picture classification using each authentication image counted in advance with classification results
The tag along sort for each authentication image recorded compares one by one, obtains the classification accuracy for recording every class image classification accuracy rate
Collection；
Threshold value subelement, for the accuracy that will classify concentrate the classification accuracy of often class image according to descending order according to
Secondary arrangement brings the adjacent accuracy rate of image classification two-by-two into classification threshold formula successively, calculates classification thresholds；
Divide subelement, for classification thresholds to be more than to the adjacent image classification accuracy rate of predetermined threshold value as segmentation
Classification accuracy collection is divided into multiple classification subsets by point；Wherein,
Classification thresholds formula is：
In formula, i*Presentation class threshold value,Indicate siThe classification accuracy of class,Indicate si-1The classification of class is accurate
Degree.
Specifically, above-mentioned Hash mapping module 14, specifically for being mapped in local search model using image to be retrieved
The output of full articulamentum and Hash mapping formula carry out Hash coding mapping to image to be retrieved, obtain the Kazakhstan of image to be retrieved
Uncommon coding；Wherein,
Hash mapping formula is：
In formula,Indicate j-th of coding in the Hash coding of i-th image,Indicate that i-th image connects entirely in mapping
Meet j-th of output valve, θ in the output of layer1For the mapping threshold value of model.
Specifically, above-mentioned Hash retrieves module 15, it is specifically used for searching in target Hash code database and the Hash coding Chinese
The similar Hash of the predetermined number of prescribed distance minimum encodes, and obtains similar Hash coded set.
In the embodiment of the present invention, further include：
Whole Hash mapping module, it is inconsistent for working as the first tag along sort and the second tag along sort, then utilize whole inspection
Rope model carries out Hash coding mapping to image to be retrieved, obtains the Hash coding of image to be retrieved；
Whole Hash retrieves module, for lookup and Hash in whole Hash code database corresponding with integral retrieval model
Coding similarity meets the similar Hash coded set of condition of similarity；
Wherein, whole Hash code database is to carry out Hash coding mapping to training image collection using integral retrieval model to obtain
's.
Further, the embodiment of the invention also discloses a kind of image retrieving apparatus, including：
Memory, for storing instruction；Wherein, instruction carries out image including the use of integral retrieval model to image to be retrieved
Classification is retrieved, and the first tag along sort of image to be retrieved is obtained；Utilize local search model pair corresponding with the first tag along sort
Image to be retrieved carries out image category retrieval, obtains the second tag along sort；Judge the first tag along sort is with the second tag along sort
It is no consistent；If it is, carrying out Hash coding mapping to image to be retrieved using local search model, image to be retrieved is obtained
Hash encodes；Lookup meets preset with Hash coding similarity in target Hash code database corresponding with local search model
The similar Hash coded set of condition of similarity；Corresponding similar image is found in image data base using similar Hash coded set
Collection；Wherein, integral retrieval model is using training image collection, to the initial integral retrieval model established based on convolutional neural networks
It is trained, training image collection includes the image category of all images to be retrieved；Local search model is to utilize and divide
The corresponding training image subset of class subset, is trained to obtain to the initial local retrieval model established based on convolutional neural networks
's；Wherein, image category retrieval is carried out to each authentication image of verification collection using integral retrieval model, obtains record and each tests
The tag along sort of image is demonstrate,proved, the classification results collection for the tag along sort for recording each authentication image is generated, concentrates and sieves from classification results
Select the classification subset for meeting preset class condition；It is filtered out from training image concentration corresponding with image category in classification subset
Training image, obtain training image subset；Target Hash code database preserves local search model in training image subset
Each training image carry out Hash coding mapping generation each training image Hash coding；
Processor, for executing the instruction in memory.
Detail about the store instruction in memory in the embodiment of the present invention can refer to aforementioned image retrieval side
Content in method embodiment is not being repeated herein.
Finally, it is to be noted that, herein, relational terms such as first and second and the like be used merely to by
One entity or operation are distinguished with another entity or operation, without necessarily requiring or implying these entities or operation
Between there are any actual relationship or orders.Moreover, the terms "include", "comprise" or its any other variant meaning
Covering non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only that
A little elements, but also include other elements that are not explicitly listed, or further include for this process, method, article or
The intrinsic element of equipment.In the absence of more restrictions, the element limited by sentence "including a ...", is not arranged
Except there is also other identical elements in the process, method, article or apparatus that includes the element.
Professional further appreciates that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure
And algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software generally describes each exemplary composition and step according to function in the above description.These
Function is implemented in hardware or software actually, depends on the specific application and design constraint of technical solution.Profession
Technical staff can use different methods to achieve the described function each specific application, but this realization is not answered
Think beyond the scope of this invention.
A kind of image search method provided by the present invention, system and device are described in detail above, herein
Applying specific case, principle and implementation of the present invention are described, and the explanation of above example is only intended to help
Understand the method and its core concept of the present invention；Meanwhile for those of ordinary skill in the art, according to the thought of the present invention,
There will be changes in the specific implementation manner and application range, in conclusion the content of the present specification should not be construed as to this
The limitation of invention.
Claims (10)
1. a kind of image search method, which is characterized in that including：
Image category retrieval is carried out to image to be retrieved using integral retrieval model, obtains the first classification of the image to be retrieved
Label；
Image category retrieval is carried out to the image to be retrieved using local search model corresponding with first tag along sort,
Obtain the second tag along sort；
Judge whether first tag along sort and second tag along sort are consistent；
If it is, carrying out Hash coding mapping to the image to be retrieved using the local search model, described wait for is obtained
Retrieve the Hash coding of image；
It searches in target Hash code database corresponding with the local search model and meets in advance with Hash coding similarity
If condition of similarity similar Hash coded set；
Corresponding similar diagram image set is found in image data base using the similar Hash coded set；
Wherein, the integral retrieval model is the initial whole inspection to being established based on convolutional neural networks using training image collection
Rope model is trained, and the training image collection includes the image category of all images to be retrieved；
The local search model is to be established using training image subset corresponding with classification subset to being based on convolutional neural networks
Initial local retrieval model be trained；Wherein, each verification using the integral retrieval model to verification collection
Image carries out image category retrieval, obtains the tag along sort for recording each authentication image, generates point for recording each authentication image
The classification results collection of class label is concentrated from the classification results and filters out the classification subset for meeting preset class condition；
Training image corresponding with image category in classification subset is filtered out from training image concentration, obtains training image
Collection；
The target Hash code database is preserved the local search model and is schemed to the training of each of described training image subset
As the Hash coding for each training image for carrying out Hash coding mapping generation.
2. image search method according to claim 1, which is characterized in that described to be filtered out from classification results concentration
Meet the process of the classification subset of preset condition of similarity, including：
Standard picture classification using each authentication image counted in advance and each verification described in the classification results collection
The tag along sort of image compares one by one, obtains the classification accuracy collection for recording every class image classification accuracy rate；
The classification accuracy per class image is concentrated to be arranged in order according to descending order the classification accuracy, successively by phase
The adjacent accuracy rate of image classification two-by-two brings classification threshold formula into, calculates classification thresholds；
The adjacent image classification accuracy rate that the classification thresholds are more than predetermined threshold value is classified accurately as cut-point by described
Degree collection is divided into multiple classification subsets；Wherein,
The classification thresholds formula is：
In formula, i*Indicate the classification thresholds,Indicate siThe classification accuracy of class,Indicate si-1The classification of class is accurate
Degree.
3. image search method according to claim 1, which is characterized in that described to utilize the local search model to institute
It states image to be retrieved and carries out Hash coding mapping, obtain the process of the Hash coding of the image to be retrieved, including：
Output and the Hash mapping formula of full articulamentum are mapped in the local search model using the image to be retrieved, it is right
The image to be retrieved carries out Hash coding mapping, obtains the Hash coding of the image to be retrieved；Wherein,
The Hash mapping formula is：
In formula,Indicate j-th of coding in the Hash coding of i-th image,Indicate that i-th image is mapping full articulamentum
Output in j-th of output valve, θ1For the mapping threshold value of model.
4. image search method according to claim 1, which is characterized in that described corresponding with the local search model
Target Hash code database in search with the Hash coding similarity meet preset condition of similarity similar Hash encode
Process, including：
The similar Hash for the predetermined number that Hamming distance minimum is encoded to the Hash is searched in the target Hash code database
Coding, obtains similar Hash coded set.
5. image search method according to any one of claims 1 to 4, which is characterized in that described to judge described first point
After whether class label and second tag along sort are consistent, further include：
When first tag along sort and second tag along sort are inconsistent, then waited for described using the integral retrieval model
It retrieves image and carries out Hash coding mapping, obtain the Hash coding of the image to be retrieved；
It is searched in whole Hash code database corresponding with the integral retrieval model and meets institute with Hash coding similarity
State the similar Hash coded set of condition of similarity；
Wherein, the whole Hash code database is to carry out Hash coding to the training image collection using the integral retrieval model
What mapping obtained.
6. a kind of image indexing system, which is characterized in that including：
Whole sort module obtains described wait for for carrying out image category retrieval to image to be retrieved using integral retrieval model
Retrieve the first tag along sort of image；
Local sort module, for utilizing local search model corresponding with first tag along sort to the image to be retrieved
Image category retrieval is carried out, the second tag along sort is obtained；
Classification judgment module, for judging whether first tag along sort and second tag along sort are consistent；
Hash mapping module, for judging first tag along sort and second tag along sort when the classification judgment module
Unanimously, then Hash coding mapping is carried out to the image to be retrieved using the local search model, obtains the figure to be retrieved
The Hash of picture encodes；
Hash retrieves module, for being searched and the Hash in target Hash code database corresponding with the local search model
Coding similarity meets the similar Hash coded set of preset condition of similarity；
Image retrieval module, for finding corresponding similar image in image data base using the similar Hash coded set
Collection；
Block mold generation module, for utilizing training image collection, to the initial integral retrieval established based on convolutional neural networks
Model is trained to obtain integral retrieval model；Wherein, the training image collection includes the image category of all images to be retrieved；
Partial model generation module, for utilizing training image subset corresponding with classification subset, to being based on convolutional neural networks
The initial local retrieval model of foundation is trained to obtain local search model；
Wherein, the partial model generation module, including：
Classification generation unit, for carrying out image category inspection to each authentication image of verification collection using the integral retrieval model
Rope obtains the tag along sort for recording each authentication image, generates the classification results collection for the tag along sort for recording each authentication image；
Subset screening unit filters out the classification for meeting preset class condition for being concentrated from the classification results
Collection；
Optical sieving unit is schemed for filtering out training corresponding with image category in classification subset from training image concentration
Picture obtains the training image subset；
Hash library generation module preserves the local search model to the training image for the target Hash code database
Each training image in subset carries out the Hash coding of each training image of Hash coding mapping generation.
7. image indexing system according to claim 6, which is characterized in that the subset screening unit, including：
Contrast subunit is concentrated for the standard picture classification using each authentication image counted in advance with the classification results
The tag along sort for each authentication image recorded compares one by one, obtains the classification accuracy for recording every class image classification accuracy rate
Collection；
Threshold value subelement, for by the classification accuracy concentrate the classification accuracy per class image according to descending order according to
Secondary arrangement brings the adjacent accuracy rate of image classification two-by-two into classification threshold formula successively, calculates classification thresholds；
Divide subelement, for the classification thresholds to be more than to the adjacent image classification accuracy rate of predetermined threshold value as segmentation
The classification accuracy collection is divided into multiple classification subsets by point；Wherein,
The classification thresholds formula is：
In formula, i*Indicate the classification thresholds,Indicate siThe classification accuracy of class,Indicate si-1The classification of class is accurate
Degree.
8. image indexing system according to claim 6, which is characterized in that the Hash mapping module is specifically used for profit
Output and the Hash mapping formula for mapping full articulamentum in the local search model with the image to be retrieved, are waited for described
It retrieves image and carries out Hash coding mapping, obtain the Hash coding of the image to be retrieved；Wherein,
The Hash mapping formula is：
In formula,Indicate j-th of coding in the Hash coding of i-th image,Indicate that i-th image is mapping full articulamentum
J-th of output valve in output, θ1For the mapping threshold value of model.
9. according to claim 6 to 8 any one of them image indexing system, which is characterized in that further include：
Whole Hash mapping module, it is inconsistent for working as first tag along sort and second tag along sort, then utilize institute
It states integral retrieval model and Hash coding mapping is carried out to the image to be retrieved, the Hash for obtaining the image to be retrieved is compiled
Code；
Whole Hash retrieves module, for searched in whole Hash code database corresponding with the integral retrieval model with it is described
Hash coding similarity meets the similar Hash coded set of the condition of similarity；
Wherein, the whole Hash code database is to carry out Hash coding to the training image collection using the integral retrieval model
What mapping obtained.
10. a kind of image retrieving apparatus, which is characterized in that including：
Memory, for storing instruction；Wherein, described instruction carries out image including the use of integral retrieval model to image to be retrieved
Classification is retrieved, and the first tag along sort of the image to be retrieved is obtained；It is examined using part corresponding with first tag along sort
Rope model carries out image category retrieval to the image to be retrieved, obtains the second tag along sort；Judge first tag along sort
It is whether consistent with second tag along sort；If it is, being carried out to the image to be retrieved using the local search model
Hash coding mapping obtains the Hash coding of the image to be retrieved；In target Hash corresponding with the local search model
The similar Hash coded set for meeting preset condition of similarity to Hash coding similarity is searched in code database；Utilize the phase
Corresponding similar diagram image set is found in image data base like Hash coded set；Wherein, the integral retrieval model is to utilize
Training image collection, to what is be trained based on the initial integral retrieval model that convolutional neural networks are established, the training figure
Image set includes the image category of all images to be retrieved；The local search model is to be schemed using training corresponding with classification subset
As subset, the initial local retrieval model established based on convolutional neural networks is trained；Wherein, using described whole
Body retrieval model carries out image category retrieval to each authentication image of verification collection, obtains the contingency table for recording each authentication image
Label generate the classification results collection for the tag along sort for recording each authentication image, and it is pre- to filter out satisfaction from classification results concentration
If class condition the classification subset；It is filtered out from training image concentration corresponding with image category in classification subset
Training image obtains the training image subset；The target Hash code database preserves the local search model to described
Each training image in training image subset carries out the Hash coding of each training image of Hash coding mapping generation；
Processor, for executing the instruction in the memory.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810264981.9A CN108491528B (en) | 2018-03-28 | 2018-03-28 | Image retrieval method, system and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810264981.9A CN108491528B (en) | 2018-03-28 | 2018-03-28 | Image retrieval method, system and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108491528A true CN108491528A (en) | 2018-09-04 |
| CN108491528B CN108491528B (en) | 2021-09-21 |
Family
ID=63316571
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810264981.9A Active CN108491528B (en) | 2018-03-28 | 2018-03-28 | Image retrieval method, system and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108491528B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109582813A (en) * | 2018-12-04 | 2019-04-05 | 广州欧科信息技术股份有限公司 | A kind of search method, device, equipment and the storage medium of historical relic showpiece |
| CN109766470A (en) * | 2019-01-15 | 2019-05-17 | 北京旷视科技有限公司 | Image search method, device and processing equipment |
| CN109918532A (en) * | 2019-03-08 | 2019-06-21 | 苏州大学 | Image search method, device, equipment and computer readable storage medium |
| CN111628866A (en) * | 2020-05-22 | 2020-09-04 | 深圳前海微众银行股份有限公司 | Neural network verification method, device and equipment and readable storage medium |
| CN111813975A (en) * | 2020-07-09 | 2020-10-23 | 国网电子商务有限公司 | Image retrieval method and device and electronic equipment |
| CN112053399A (en) * | 2020-09-04 | 2020-12-08 | 厦门大学 | Method for positioning digestive tract organs in capsule endoscope video |
| WO2020244437A1 (en) * | 2019-06-06 | 2020-12-10 | 腾讯科技（深圳）有限公司 | Image processing method and apparatus, and computer device |
| CN114241484A (en) * | 2022-02-23 | 2022-03-25 | 北京嘉沐安科技有限公司 | Social network-oriented image big data accurate retrieval method and system |
| CN116610805A (en) * | 2023-07-20 | 2023-08-18 | 恒辉信达技术有限公司 | Unstructured data application method, system, equipment and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104008395A (en) * | 2014-05-20 | 2014-08-27 | 中国科学技术大学 | Intelligent bad video detection method based on face retrieval |
| 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 |
| CN106503106A (en) * | 2016-10-17 | 2017-03-15 | 北京工业大学 | A kind of image hash index construction method based on deep learning |
-
2018
- 2018-03-28 CN CN201810264981.9A patent/CN108491528B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104008395A (en) * | 2014-05-20 | 2014-08-27 | 中国科学技术大学 | Intelligent bad video detection method based on face retrieval |
| 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 |
| CN106503106A (en) * | 2016-10-17 | 2017-03-15 | 北京工业大学 | A kind of image hash index construction method based on deep learning |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109582813A (en) * | 2018-12-04 | 2019-04-05 | 广州欧科信息技术股份有限公司 | A kind of search method, device, equipment and the storage medium of historical relic showpiece |
| CN109582813B (en) * | 2018-12-04 | 2021-10-01 | 广州欧科信息技术股份有限公司 | Retrieval method, device, equipment and storage medium for cultural relic exhibit |
| CN109766470A (en) * | 2019-01-15 | 2019-05-17 | 北京旷视科技有限公司 | Image search method, device and processing equipment |
| CN109918532A (en) * | 2019-03-08 | 2019-06-21 | 苏州大学 | Image search method, device, equipment and computer readable storage medium |
| CN109918532B (en) * | 2019-03-08 | 2023-08-18 | 苏州大学 | Image retrieval method, device, equipment and computer readable storage medium |
| WO2020182019A1 (en) * | 2019-03-08 | 2020-09-17 | 苏州大学 | Image search method, apparatus, device, and computer-readable storage medium |
| WO2020244437A1 (en) * | 2019-06-06 | 2020-12-10 | 腾讯科技（深圳）有限公司 | Image processing method and apparatus, and computer device |
| WO2021233183A1 (en) * | 2020-05-22 | 2021-11-25 | 深圳前海微众银行股份有限公司 | Neural network verification method, apparatus and device, and readable storage medium |
| CN111628866A (en) * | 2020-05-22 | 2020-09-04 | 深圳前海微众银行股份有限公司 | Neural network verification method, device and equipment and readable storage medium |
| CN111813975A (en) * | 2020-07-09 | 2020-10-23 | 国网电子商务有限公司 | Image retrieval method and device and electronic equipment |
| CN112053399A (en) * | 2020-09-04 | 2020-12-08 | 厦门大学 | Method for positioning digestive tract organs in capsule endoscope video |
| CN112053399B (en) * | 2020-09-04 | 2024-02-09 | 厦门大学 | Method for positioning digestive tract organs in capsule endoscope video |
| CN114241484A (en) * | 2022-02-23 | 2022-03-25 | 北京嘉沐安科技有限公司 | Social network-oriented image big data accurate retrieval method and system |
| CN116610805A (en) * | 2023-07-20 | 2023-08-18 | 恒辉信达技术有限公司 | Unstructured data application method, system, equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108491528B (en) | 2021-09-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108491528A (en) | A kind of image search method, system and device | |
| CN109918532B (en) | Image retrieval method, device, equipment and computer readable storage medium | |
| CN106407352B (en) | Traffic image search method based on deep learning | |
| KR102458463B1 (en) | A method for training a convolutional neural network for image recognition using image-conditioned masked language modeling | |
| CN111159407B (en) | Method, apparatus, device and medium for training entity recognition and relation classification model | |
| CN110362723B (en) | Topic feature representation method, device and storage medium | |
| US8422832B2 (en) | Annotating images | |
| CN103823824B (en) | A kind of method and system that text classification corpus is built automatically by the Internet | |
| CN111460249A (en) | Personalized learning resource recommendation method based on learner preference modeling | |
| CN111026841B (en) | Automatic coding method and device based on retrieval and deep learning | |
| CN110689081A (en) | Weak supervision target classification and positioning method based on bifurcation learning | |
| Chow et al. | Multilayer SOM with tree-structured data for efficient document retrieval and plagiarism detection | |
| CN109935337B (en) | Medical record searching method and system based on similarity measurement | |
| Weyand et al. | Visual landmark recognition from internet photo collections: A large-scale evaluation | |
| CN109829065B (en) | Image retrieval method, device, equipment and computer readable storage medium | |
| CN104346438A (en) | Data management service system based on large data | |
| CN109063113A (en) | A kind of fast image retrieval method based on the discrete Hash of asymmetric depth, retrieval model and model building method | |
| CN110196918B (en) | Unsupervised deep hashing method based on target detection | |
| WO2016057000A1 (en) | Customs tariff code classification | |
| CN111914062A (en) | Long text question-answer pair generation system based on keywords | |
| CN109145161A (en) | Chinese Place Names querying method, device and equipment | |
| CN110580339A (en) | Method and device for perfecting medical term knowledge base | |
| CN113722478A (en) | Multi-dimensional feature fusion similar event calculation method and system and electronic equipment | |
| CN111180060B (en) | Disease diagnosis automatic coding method and device | |
| CN113836269B (en) | Chapter-level core event extraction method based on question-answering 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |