CN107885764A - Based on the quick Hash vehicle retrieval method of multitask deep learning - Google Patents
Based on the quick Hash vehicle retrieval method of multitask deep learning Download PDFInfo
- Publication number
- CN107885764A CN107885764A CN201710857318.5A CN201710857318A CN107885764A CN 107885764 A CN107885764 A CN 107885764A CN 201710857318 A CN201710857318 A CN 201710857318A CN 107885764 A CN107885764 A CN 107885764A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msub
- msup
- msubsup
- vehicle
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
-
- 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
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
One kind be based on the quick Hash vehicle retrieval method of multitask deep learning, including for deep learning and the multitask depth convolutional neural networks of training identification, the Feature fusion of the compact Hash codes of the segmentation for improving retrieval precision and search method practicality and example aspects, retrieve for being lifted the local sensitivity Hash of performance sort algorithm and cross-module state search method for lifting search engine robustness and accuracy again.First, propose a kind of multitask depth convolutional network speced learning Hash codes method, image, semantic and graphical representation are combined, retrieval precision and refined image feature are improved using the contact between inter-related task, while the vehicle characteristics for arriving study using Image Coding is minimized have more robustness；Secondly, the example aspects of vehicle image are extracted again from feature pyramid network；Then, using local sensitivity Hash, sort method is retrieved to the feature extracted again；Finally, to the special circumstances of enquiring vehicle target image can not be obtained using cross-module state auxiliary vehicle retrieval method.
Description
Technical field
Led the present invention relates to artificial intelligence, Digital Image Processing, convolutional neural networks and computer vision in public safety
Application in terms of domain, belongs to intelligent transportation field.
Background technology
In today that smart city, intelligent transportation develop rapidly, large-scale image monitoring and video in public safety system
The demand sharp increase of the vehicle identification, vehicle retrieval of database.
In the prior art, for the license board information of the search method of vehicle, mainly extraction target vehicle.Then root then,
Motor vehicles to be retrieved are retrieved according to license board information.General specific practice is that the license plate number of vehicle is identified from monitoring image
Code, then identify whether the motor vehicles with the number-plate number in other monitoring images again.It is this to rely solely on license plate number
Although the method retrieved easily is realized, the motor vehicles for license plate information can not be got, such as false-trademark fake-licensed car
Then can not effectively it be retrieved.
Vehicle retrieval technology based on external appearance characteristic not only can overcome the disadvantages that the limitation and deficiency of traditional vehicle board recognition methods, in intelligence
Can vehicle retrieval, especially in inspection violating the regulations, accident chase, suspicion of crime Vehicle locking, fake license plate vehicle differentiates, and accelerates criminal investigation
Efficiency of solving a case and speed all have very important realistic meaning and very wide application prospect.
Existing vehicle retrieval method is substantially to utilize the whole of sift, surf, dog scheduling algorithm extraction target vehicle image
Width characteristics of image, using this feature as target signature, utilize the view picture of each vehicle image in identical algorithm extraction database
Characteristics of image, using this feature as feature to be matched, the Euclidean distance between calculating target signature and each feature to be matched will
Vehicle is as target vehicle corresponding to the nearest feature to be matched of Euclidean distance.
The vehicle of specific objective is found out in vehicle retrieval requirement in a series of similar vehicle of profiles, and this allows for work change
Challenge must be had more；Moreover, it is contemplated that actual conditions, monitors environment, weather condition and illumination condition etc.
The influence of factor.
Recent years, deep learning in the technology of computer vision field obtained rapid development, and deep learning can utilize
Substantial amounts of training sample and hidden layer successively in depth learn the abstracted information of image, more comprehensively directly obtain characteristics of image.
Digital picture is described with matrix, and convolutional neural networks describe the whole of image preferably from local message block
Body structure, therefore solve problem using convolutional neural networks mostly in computer vision field, deep learning method.Around
Improve accuracy of detection and detection time, depth convolutional neural networks technology is from R-CNN, FasterR-CNN to Fasterer R-
CNN.Be embodied in further precision improvement, acceleration, end-to-end and more practical, almost cover from be categorized into detection, point
Cut, position every field.It will be that the research for having very much actual application value is led that depth learning technology, which applies to vehicle retrieval,
Domain.
It is the technology for the lifting retrieval performance commonly used in image retrieval technologies to sort again, for example, can be by between image pair
Visual signature matching relationship initial retrieval result is sorted again.However, the effect that sorts again is greatly dependent on used
Whether visual signature can sufficiently effective express image.
In similar vehicle search, due to many vehicles, often profile is much like, and the visual signature extracted can also compare
It is similar, different automobile types are cannot be distinguished by, so as to cause this sort method again for directly using the matching relationship between image pair not
Similar vehicle can preferably be retrieved.
Query expansion is to be used for improving the common method of recall rate and accuracy rate in retrieval technique.Query expansion technology be by
Originally inquire about the method that sentence increases new keyword to inquire about again, such as the inquiry sentence elder generation that search engine can input user
Primary retrieval is done, according to the file being retrieved, select suitable keyword, be added to inquiry sentence and retrieve again, look for whereby
Go out more associated documents.It can be seen that query expansion can effectively improve the recall rate of information retrieval, but do not have also in the prior art
There is this special object of the vehicle proposition being directed in image targetedly enquiry expanding method.
" a kind of vehicle inspection based on similarity study disclosed in the Chinese patent application of Application No. 201510744990.4
Suo Fangfa ", relative to vehicle retrieval method of the tradition based on the number-plate number, the method for proposition not only effectively prevent to be known to car plate
The dependence of the other degree of accuracy, title can be retrieved to false-trademark car and fake-licensed car.But the technology still fall within before deep learning when
The computer vision technique in generation.
A kind of vehicle retrieval side based on big data disclosed in the Chinese patent application of Application No. 201610671729.0
Method and device, this method include：Extract the brand identity of target vehicle in target vehicle image；Determine every in target vehicle image
Individual pixel corresponds to the probability of each mark, wherein, mark includes the one or more in annual test mark, goods of furniture for display rather than for use, ornament；Root
According to each pixel in target vehicle image correspond to each mark probability and each mark corresponding to probability threshold value, really
Fixed position of each mark in target vehicle image；It is each according to position extraction of each mark in target vehicle image
The characteristics of image of individual mark；It is special according to the brand of the characteristics of image of each mark and target vehicle in target vehicle image
Sign searched targets vehicle in multiple vehicle images to be retrieved.Although the technology employs depth learning technology, but still belong to
In the depth learning technology of single task；But vehicle retrieval belongs to a kind of depth learning technology of typical multitask.
The inquiry that the Chinese patent application of Application No. 201410381577.1 is disclosed in a kind of similar vehicle retrieval is expanded
Method and device is opened up, wherein methods described includes：According to the image to be checked including vehicle, the car of the image to be checked is determined
Type information；The multiple samples for meeting preparatory condition are chosen in vehicle ATL corresponding to vehicle information from the image to be checked
Image；The sample image is formed into query expansion image collection, so that the sample image in the query expansion image collection
Inquired about instead of the image to be checked in target database；Wherein, the vehicle ATL includes：Multiple different illumination
Under the conditions of vehicle sample image, the vehicle sample image of the vehicle sample image of different shooting angles and different scenes, by upper
The method of stating can improve the recall rate and accuracy rate of vehicle image retrieval.The figure in deep learning epoch before this method is still fallen within fact
As retrieval technique.
Application No. 201410652730.X Chinese patent application discloses a kind of motor vehicles retrieval based on image
Method and device.Methods described includes：Obtain the first image for including motor vehicle information to be retrieved；From described first image
Determine the first appearance profile of the motor vehicles to be retrieved；Image in first appearance profile is divided into multiple areas
Domain, the characteristics of image of regional is extracted using different step-lengths；Merge the characteristics of image of regional, obtain to be retrieved motor-driven
The general image feature of vehicle；By the general image feature of the motor vehicles to be retrieved and the subject vehicle that in advance extracts
General image feature be compared, obtain comparison result.The image inspection in deep learning epoch before this method is still fallen within fact
Rope technology.
The content of the invention
For how efficiently to utilize caused massive video data in public safety field, the vehicle in big data epoch is lifted
The problems such as recall precision, the present invention propose one kind and are based on the quick Hash search method of multitask deep learning, effectively utilize inspection
Relevance, the diversity of bayonet vehicle essential information surveyed between identification mission realize the purpose of real-time retrieval；It is final to provide
A kind of quick Hash vehicle retrieval method of multitask deep learning that retrieval precision is high, robustness is good.
The technical solution adopted for the present invention to solve the technical problems is：
One kind is based on the quick Hash vehicle retrieval method of multitask deep learning, and the vehicle retrieval method includes following step
Suddenly：
The first step, build the multitask depth convolutional neural networks for deep learning and training identification；
Second step, using the Feature fusion for being segmented compact Hash codes and example aspects；
3rd step, using local sensitivity Hash sort algorithm again；
4th step, using cross-module state search method, realize vehicle retrieval.
Further, in the first step, using facilities networks of the Faster R-CNN as multitask convolutional neural networks；Net
Network foremost is 3 × 3 convolutional networks, referred to as conv1, is named as conv2_x followed by 4 stacking convolution modules and arrives
Conv5_x, each module contain { 2,3,3,3 } individual unit respectively, and conv1 to conv4_3 is as shared network；It is followed by
RPN, i.e. region suggest network, and RPN networks are using any scalogram picture as input, the collection of output rectangular target Suggestion box
Close, each frame includes 4 position coordinates variables and a score；For formation zone Suggestion box, in last shared volume
Small network is slided in the convolution Feature Mapping of lamination output, this network is connected to n × n of input convolution Feature Mapping sky entirely
Between on window；Each sliding window is mapped on a low-dimensional vector, and a sliding window of each Feature Mapping is corresponding one
Numerical value；This vector exports the layer of the full connection at the same level to two；
Described RPN；RPN networks export the set of rectangular target Suggestion box using any scalogram picture as inputting,
Each frame includes 4 position coordinates variables and a score；The target of described rectangular target Suggestion box refers to Vehicle Object；
It is the estimated probability of target/non-targeted to each Suggestion box, is the classification realized with the softmax layers of two classification
Layer；K Suggestion box is parameterized by the corresponding k Suggestion box for being referred to as anchor；
Each anchor is centered on current sliding window mouth center, and a kind of corresponding yardstick and length-width ratio, uses 3 kinds of yardsticks
With 3 kinds of length-width ratios, so just there is k=9 anchor in each sliding position；
In order to train RPN networks, a binary label is distributed to each anchor, is to mark the anchor with this
It is not target；Then distribute positive label and give this two class anchor：(I) with some real target bounding box, i.e. Ground Truth,
GT has the ratio between highest IoU, i.e. Interse-ction-over-Union, common factor union, overlapping anchor；(II) it is and any
GT bounding boxs have the overlapping anchor of the IoU more than 0.7；Notice that a GT bounding box may give multiple anchor distribution positive mark
Label；The IoU ratios that the negative label of distribution gives all GT bounding boxs are below 0.3 anchor；Anon-normal non-negative anchor is to instruction
Practicing target does not have any effect, then abandons；
The multitask loss in Faster R-CNN is followed, minimizes object function；The loss function of one image is determined
Justice is：
Here, i is anchor index, piIt is the prediction probability that anchor is the i-th target, if anchor is
Just, GT labelsIt is exactly 1, if anchor is negative,It is exactly 0；tiIt is a vector, represents 4 parameters of the bounding box of prediction
Change coordinate,It is the coordinate vector of GT bounding boxs corresponding with positive anchor；λ is a balance weight, NclsIt is the normalizing of cls items
Change value be mini-batch size, NregBe reg items normalized value be anchor positions quantity, Classification Loss function Lcls
Two classifications, i.e. the logarithm loss of motor vehicles target vs. road backgrounds：
For returning loss function Lreg, defined to minor function：
In formula, LregTo return loss function, R is the loss function of robust, and smooth L are calculated with formula (4)1；
In formula, smoothL1For smooth L1Loss function, x are variable.
Further, in the multitask depth convolutional neural networks, it is heavy to closing to design a kind of multitask object function
Want；Multitask object function is represented with formula (5)；
In formula,It is input feature valueWith weight parameter wtMapping function, L () is loss function, Φ
(wt) be weight parameter regularization value, T is general assignment number, is designated as the training data of t-th of taskWherein t ∈
(1, T), i ∈ (1, N), N are total number of training,The respectively characteristic vector of the i-th sample and mark label；
For loss function, coordinate log-likelihood cost function to train the feature of last layer using softmax, realize more
Task image is classified, and softmax loss functions are defined with formula (6),
In formula, xiFor the i-th depth characteristic, WjFor the jth row of weight in last full articulamentum, b is bias term, and m, n divide
Sample size and classification number Wei not handled.
Further, in the second step, the Feature fusion process for being segmented compact Hash codes and example aspects is as follows：
In vehicle image feature extraction phases, first, by softmax activation primitives by threshold restriction between [0,1]；
Then, promote the output of binary system Hash codes by fragmentation threshold function, Hash is reduced using speced learning and coding strategy
The redundancy of intersymbol improves feature robustness；Finally, by way of Fusion Features by speced learning come Hash codes carry out
Fusion finally gives vehicle characteristics and is segmented compact Hash codes；
For the example aspects of vehicle, implementation method is：It is each that stacking convolution is shared from conv2_x to conv5_x respectively
The output of last unit combination RPN network of module, pyramid pond layer and vectorial flat layer are added to adapt to different chis
Very little convolution characteristic pattern input, while one-dimensional characteristic vector, the reality of the vector referred to as vehicle are turned to by convolution three-dimensional feature is flat
Example feature；
Finally, the compact feature and example aspects of the compact Hash codes of vehicle subsection are merged to obtain the spy for retrieval again
Sign vector.
Described vehicle characteristics are segmented compact Hash codes and are realized by the following method；If share T task, Mei Geren
C be present under businesstIndividual classification, uses mtThe full connection output vector of each task is represented, makes to connect entirely using softmax activation primitives
Layer output is calculated between [0,1] with formula (7)；
Wherein, θ represents random hyperplane, mtRepresent the full connection output vector of each task, ctRepresent to deposit under each task
Classification, qtRepresent full articulamentum output；
To encourage the output of the binary system of segment encoding module, threshold segmentation function binaryzation is reused：
Wherein, qtRepresent full articulamentum output, HtRepresent the binary system output of excitation segment encoding module；
Finally by HtFusion turns into the compact Hash codes vector f of vehicle subsectionA：
fA=[α1H1；α2H2；...；αtHt] (9)
Wherein, fARepresent the compact Hash code vector of vehicle subsection, αtCoefficient is represented, is calculated with formula (10), HtRepresent
Encourage the binary system output of segment encoding module, t ∈ (1, T)：
Wherein, αtCoefficient is represented, factor alpha is multiplied by before H vectorstTo make up between different task caused by inequality of classifying
Error.
In 3rd step, the characteristic vector for retrieval is by the compact feature and example of the compact Hash codes of vehicle subsection
Feature is merged what is obtained, and process is as follows：
{ 4 are selected respectively for conv2_x to conv5_x bottommost layer2,82,162,162Output chi as characteristic pattern
It is very little；Be h × w for the size for giving input picture I, convolution convx_x is activated as three-dimensional vector T, size for h ' × w ' ×
D, contain a series of two dimensional character figure S=S { Sn, n ∈ (1, d), SnSize correspond to n-th of channel characteristics for h ' × w '
The size of figure；T is sent into pyramid pond layer again and obtains three-dimensional vector T ', size is l × l × d, still comprising series of features
Scheme S '=S ' { S 'n, n ∈ (1, d), S 'nSize be l × l, to each S 'nIt is that k × k sliding windows are traveled through using size
Choose maximum, S 'nIt is changed into l/k × l/k, then the S ' to each passagenMerged to obtain one-dimensional vector, successively to d passage
Carry out same operation, the personal characteristics vector f finally obtainedBSize is (1, l/k × d)；Final searching characteristic vector f, meter
Shown in calculation method such as formula (11)；
F=[fA；fB] (11)
In formula, f is used for the characteristic vector of vehicle retrieval, fBExample aspects vector, i.e. personal characteristics vector, fARepresent vehicle
It is segmented compact Hash code vector.
In 3rd step, similar sample is mapped in the same same bucket with high probability；Local sensitivity
The hash function h () of Hash meets following condition：
s{h(fAq)=h (fA)=sim (fAq,fA) (12)
In formula, sim (fAq,fA) represent fAqWith fASimilarity, h (fA) represent fAHash function, h (fAq) represent fAq's
Hash function, similarity measurement therein and a distance function σ direct correlation, are calculated with formula (13),
The typical classification of local sensitivity hash function is provided by accidental projection and threshold value, is calculated with formula (14),
h(fA)=sign (WfA+b) (14)
In formula, W is a random hyperplane vector, and b is a random intercept.
In 3rd step, when query image recycles image by being segmented after compact Hash codes are mapped in similar bucket
Example aspects combination formula (15) image returned in bucket is sorted again；The computational methods that sort again such as formula (15) institute
Show：
In formula, k represents k-th of image in bucket,Represent penalty factor andCos represents COS distance formula, y tables
Show the f before mappingAqWithIt is whether equal；Y is 1 if equal, is otherwise 0,Represent that k-th of image vehicle subsection is compact
Hash code vector, fAqRepresent the compact Hash code vector of vehicle subsection after sorting againRetrieval result error image is allowed to be looked into input
It is farther to ask the distance between image；Dis is smaller to show that similarity is higher.
In 4th step, cross-module state search method is by building one group of deep neural network by image and text data
Common semantic space is mapped to by way of feature learning, to realize that the semanteme to different modalities data couples；Using depth
Degree convolutional neural networks extract the semantic feature of image modalities directly from input picture, and text is represented by the way of term vector
This, the semantic feature of middle extraction text modality is represented with one-dimensional convolutional neural networks from term vector；First, depth convolution god is passed through
The compact Hash codes f of segmentation of vehicle is generated through network dynamicA；Then, by text generation searching characteristic vector, so both are each
The characteristic vector of self-generating using same searching system with regard to can be retrieved.
The semantic feature of described text modality is that characteristic vector is extracted from text, first as the extraction algorithm first step
First need to split text；Entry of the characteristic vector of text from text, specific steps：
Input：Text O；Output：One group of rough similar image；
STEP1：Initialization：(1) text file analysis is vectorial into entry；(2) small word, repetitor are removed；(3) entry is checked
Ensure the correctness of parsing；
STEP2：The minimum vectorial R=(r of entry of random combine are taken out from O1,r2,...,rn)；
STEP3：To R and fAOrder and the compact Hash codes of segmentation are integrated, and obtain text attribute featureNow
fATxtDimension is less than R dimension；
STEP4：The hash algorithm that sorted again using local sensitivity is retrieved；
STEP5：Return to similar image group I；
Wherein, text attribute characteristic functionIt is expressed as with formula (16)：
In formula, ATBe expressed as the transposed matrix of the compact Hash codes of vehicle subsection, R be expressed as the entry minimum of random combine to
Amount,For text attribute characteristic function, it is sign function that sign, which is represented,；
In formula,Diagonal matrix is sought in diag expressions,Expression is that characteristic vector is extracted from text, car
The compact Hash codes A of segmentationTIt is initialized as complete 1 vector of (1 × c).
The present invention technical concept be：First, it is proposed that a kind of multitask depth convolutional network speced learning Hash codes side
Method, image, semantic and graphical representation are combined, improve retrieval precision using the contact between inter-related task and refined image is special
Sign, while the vehicle characteristics for arriving study using Image Coding is minimized have more robustness；Secondly, from feature pyramid network
Network extracts the example aspects of vehicle image again；Then, using local sensitivity Hash, sort method is carried out to the feature extracted again
Retrieval；Finally, to the special circumstances of enquiring vehicle target image can not be obtained using cross-module state auxiliary vehicle retrieval method.
Text generation searching characteristic vector is identical with the compact Hash code vector of the segmentation that convolutional network generates so that without warp
Extra training is crossed, the characteristic vector that both generate can be retrieved using same searching system.
Depth convolutional neural networks model constructed by this patent, as shown in figure 1, be an end-to-end learning system,
The model is jointly whole by Text Representation, characteristics of image study, text feature study, the retrieval of cross-module state and the task such as sort again
Close among same learning framework.
Beneficial effects of the present invention are mainly manifested in：
1) a kind of contour of the vehicle identification framework of multitask deep learning is devised.It is parallel using the correlation between task
Weights in processing procedure share the generalization ability of raising system, weaken influence of the over-fitting to neutral net, so solve because
Sample deficiency and caused by grader generalization ability it is not strong the problem of, and attempted heterogeneous networks structure, finally will be interrelated
Task merged, make network parameter is shared to reach maximization.
2) segmented method study Hash codes are used to reduce between binary system Hash codes with reference to multitask network structure
Redundancy.Each task is responsible for learning a part of Hash codes and connectionless between each other, then by set forth herein Vector Fusion
Method obtains accurately each vehicle image feature representation, and this feature is referred to as to the compact feature of segmentation of vehicle；Using shared heap
Folded convolutional layer, pyramid pond layer and vectorial flat layer, i.e. Vector flat layer, multiple layer combination construction feature it is golden
The example aspects of word tower network acquisition image, enter finally by by the graphical representation of get two kinds of different characteristic dimensional informations
The searching characteristic vector that row vector merges to the end again.
3) proposing local sensitivity Hash, ordering searching method carries out Rapid matching to meet to the retrieval character got again
The practical application request of intelligent transportation.This search method will inquire about image in storehouse first by the compact Hash codes of segmentation and be mapped to respectively
In individual " bucket ", then the minor sort again of image in bucket is filtered out most by vehicle different characteristic dimension using example aspects vector
TopK similar image, and using the mapping of coding vector so as to avoiding image from contrasting one by one to reach quick real-time retrieval
Effect.
4) for can not obtain enquiring vehicle image information, night camera view is fuzzy or daylight is too strong,
The special circumstances such as camera dead angle, the present invention propose cross-module state assisted retrieval mode to meet the actual requirement of varying environment；Root
Vehicle characteristics are summed up according to artificial judgement, and then changes into text data feeding retrieval network and realizes assisted retrieval.
Brief description of the drawings
Fig. 1 is that the quick Hash of multitask depth convolutional neural networks retrieves overall network framework；
Fig. 2 is the schematic diagram that sorts again；
Fig. 3 is Text eigenvector generating process explanation figure；
Fig. 4 is RPN network structures；
Fig. 5 is multitask Faster R-CNN depth convolutional network structure charts.
Embodiment
The invention will be further described below in conjunction with the accompanying drawings.
1~Fig. 5 of reference picture, one kind are based on the quick Hash vehicle retrieval method of multitask deep learning, including：
The first step, build the multitask depth convolutional neural networks for deep learning and training identification；
Second step, using the Feature fusion for being segmented compact Hash codes and example aspects；
3rd step, using local sensitivity Hash sort algorithm again；
4th step, using cross-module state search method, realize vehicle retrieval.
In the described first step, for deep learning and the multitask depth convolutional neural networks of training identification, such as Fig. 1 institutes
Show；Using facilities networks of the Faster R-CNN as multitask convolutional neural networks；Network foremost is 3 × 3 convolutional networks,
Referred to as conv1, stack convolution modules followed by 4 and be named as conv2_x to conv5_x, each module contain respectively 2,
3,3,3 } individual unit, conv1 to conv4_3 is as shared network；It is followed by RPN, i.e. network is suggested in region, and RPN networks are by one
Individual any scalogram picture exports the set of rectangular target Suggestion box, each frame includes 4 position coordinates variables and one as input
Individual score；For formation zone Suggestion box, small net is slided in the convolution Feature Mapping of last shared convolutional layer output
Network, this network are connected in n × n of input convolution Feature Mapping spatial window entirely；Each sliding window is mapped to one
On low-dimensional vector, the corresponding numerical value of a sliding window of each Feature Mapping；This vector exports at the same level to two complete
The layer of connection；
Described RPN, as shown in Figure 4；RPN networks are built any scalogram picture as input, output rectangular target
The set of frame is discussed, each frame includes 4 position coordinates variables and a score；What the target of described rectangular target Suggestion box referred to
It is Vehicle Object；
It is the estimated probability of target/non-targeted to each Suggestion box, is the classification realized with the softmax layers of two classification
Layer；K Suggestion box is parameterized by the corresponding k Suggestion box for being referred to as anchor；
Each anchor is centered on current sliding window mouth center, and a kind of corresponding yardstick and length-width ratio, uses 3 kinds of yardsticks
With 3 kinds of length-width ratios, so just there is k=9 anchor in each sliding position；
In order to train RPN networks, a binary label is distributed to each anchor, is to mark the anchor with this
It is not target；Then distribute positive label and give this two class anchor：(I) with some real target bounding box, i.e. GroundTruth,
GT has the ratio between highest IoU, i.e. Interse-ction-over-Union, common factor union, overlapping anchor；(II) it is and any
GT bounding boxs have the overlapping anchor of the IoU more than 0.7；Notice that a GT bounding box may give multiple anchor distribution positive mark
Label；The IoU ratios that the negative label of distribution gives all GT bounding boxs are below 0.3 anchor；Anon-normal non-negative anchor is to instruction
Practicing target does not have any effect, then abandons；
There are these to define, it then follows the multitask loss in Faster R-CNN, to minimize object function；To an image
Loss function be defined as：
Here, i is anchor index, piIt is the prediction probability that anchor is the i-th target, if anchor is
Just, GT labelsIt is exactly 1, if anchor is negative,It is exactly 0；tiIt is a vector, represents 4 parameters of the bounding box of prediction
Change coordinate,It is the coordinate vector of GT bounding boxs corresponding with positive anchor；λ is a balance weight, here λ=10, NclsIt is
The normalized value of cls items is mini-batch size, here Ncls=256, NregThe normalized value for being reg items is anchor positions
The quantity put, Nreg=2,400, Classification Loss function LclsIt is the logarithm of two classifications, i.e. motor vehicles target vs. road backgrounds
Loss：
For returning loss function Lreg, defined to minor function：
In formula, LregTo return loss function, R is the loss function of robust, and smooth L are calculated with formula (4)1；
In formula, smoothL1For smooth L1Loss function, x are variable.
The multitask depth convolutional neural networks, as shown in Figure 5；Learnt in order to which multiple-task is merged
And training, it is vital to design a kind of multitask object function；Multitask object function is represented with formula (5)；
In formula,It is input feature valueWith weight parameter wtMapping function, L () is loss function, Φ
(wt) be weight parameter regularization value, T is general assignment number, is designated as the training data of t-th of taskWherein t ∈
(1, T), i ∈ (1, N), N are total number of training,The respectively characteristic vector of the i-th sample and mark label；
For loss function, coordinate log-likelihood cost function to train the feature of last layer using softmax, realize more
Task image is classified, and softmax loss functions are defined with formula (6),
In formula, xiFor the i-th depth characteristic, WjFor the jth row of weight in last full articulamentum, b is bias term, and m, n divide
Sample size and classification number Wei not handled.
The described compact Hash codes of segmentation and the Feature fusion of example aspects, as shown in Figure 1；On the one hand, in vehicle
The image characteristics extraction stage, first, by softmax activation primitives by threshold restriction between [0,1]；Then, segmentation is passed through
Threshold function table promotes the output of binary system Hash codes, uses speced learning and coding strategy reduces the redundancy of Hash intersymbol and come
Improve feature robustness；Finally, by way of Fusion Features by speced learning come Hash codes carry out fusion and finally give car
Compact Hash codes of Feature Segmentation；
On the other hand, the example aspects on vehicle；Inspired by image pyramid technology, the vehicle of convolutional layer extraction is real
Example feature further merges with the compact feature extracted in multitask deep learning vehicle retrieval network so that retrieval result is more
It is precisely reliable；Implementation method is：Last unit for stacking each module of convolution is shared from conv2_x to conv5_x respectively
With reference to the output of RPN networks, add pyramid pond layer and vectorial flat layer and inputted with adapting to various sizes of convolution characteristic pattern,
Simultaneously by convolution three-dimensional feature it is flat turn to one-dimensional characteristic vector, the vector referred to as vehicle example aspects；
Finally, the compact feature and example aspects of the compact Hash codes of vehicle subsection are merged to obtain the spy for retrieval again
Sign vector.
Described vehicle characteristics are segmented compact Hash codes and are realized by the following method；If share T task, Mei Geren
C be present under businesstIndividual classification, uses mtThe full connection output vector of each task is represented, makes to connect entirely using softmax activation primitives
Layer output is calculated between [0,1] with formula (7)；
Wherein, θ represents random hyperplane, mtRepresent the full connection output vector of each task, ctRepresent to deposit under each task
Classification, qtRepresent full articulamentum output；
To encourage the output of the binary system of segment encoding module, threshold segmentation function binaryzation is reused：
Wherein, qtRepresent full articulamentum output, HtRepresent the binary system output of excitation segment encoding module；
Finally by HtFusion turns into the compact Hash codes vector f of vehicle subsectionA：
fA=[α1H1；α2H2；...；αtHt] (9)
Wherein, fARepresent the compact Hash code vector of vehicle subsection, αtCoefficient is represented, is calculated with formula (10), HtRepresent
Encourage the binary system output of segment encoding module, t ∈ (1, T)：
Wherein, αtCoefficient is represented, factor alpha is multiplied by before H vectorstTo make up between different task caused by inequality of classifying
Error.
The described characteristic vector for being used to retrieve is to enter the compact feature and example aspects of the compact Hash codes of vehicle subsection
Row fusion obtains, and concrete methods of realizing is：
{ 4 are selected respectively for conv2_x to conv5_x bottommost layer2,82,162,162Output chi as characteristic pattern
It is very little；Be h × w for the size for giving input picture I, convolution convx_x is activated as three-dimensional vector T, size for h ' × w ' ×
D, contain a series of two dimensional character figure S=S { Sn, n ∈ (1, d), SnSize correspond to n-th of channel characteristics for h ' × w '
The size of figure；T is sent into pyramid pond layer again and obtains three-dimensional vector T ', size is l × l × d, still comprising series of features
Scheme S '=S ' { S 'n, n ∈ (1, d), S 'nSize be l × l, to each S 'nIt is that k × k sliding windows are traveled through using size
Choose maximum, S 'nIt is changed into l/k × l/k, then the S ' to each passagenMerged to obtain one-dimensional vector, successively to d passage
Carry out same operation, the personal characteristics vector f finally obtainedBSize is (1, l/k × d)；Final searching characteristic vector f, meter
Shown in calculation method such as formula (11)；
F=[fA；fB] (11)
In formula, f is used for the characteristic vector of vehicle retrieval, fBExample aspects vector, i.e. personal characteristics vector, fARepresent vehicle
It is segmented compact Hash code vector.
The described local sensitivity Hash for being used to being lifted retrieval performance sort algorithm again, algorithm idea as shown in Fig. 2 be by
Similar sample is mapped in the same same bucket with high probability；The hash function h () of local sensitivity Hash meet with
Lower condition：
s{h(fAq)=h (fA)=sim (fAq,fA) (12)
In formula, sim (fAq,fA) represent fAqWith fASimilarity, h (fA) represent fAHash function, h (fAq) represent fAq's
Hash function, similarity measurement therein and a distance function σ direct correlation, are calculated with formula (13),
The typical classification of local sensitivity hash function is provided by accidental projection and threshold value, is calculated with formula (14),
h(fA)=sign (WfA+b) (14)
In formula, W is a random hyperplane vector, and b is a random intercept.
The Feature fusion of the compact Hash codes of described segmentation and example aspects, in order that similar image closer to,
The example aspects combination formula of image is recycled after query image is mapped in similar bucket by the compact Hash codes of segmentation
(15) image returned in bucket is sorted again；Shown in the computational methods that sort again such as formula (15)：
In formula, k represents k-th of image in bucket,Represent penalty factor andCos represents COS distance formula, y tables
Show the f before mappingAqWithIt is whether equal；Y is 1 if equal, is otherwise 0,Represent that k-th of image vehicle subsection is compact
Hash code vector, fAqRepresent the compact Hash code vector of vehicle subsection after sorting again；
Add coefficientPurpose is in order to ensure the correctness of LSH mappings, i.e., in the case of identical is segmented compact Hash codes
Ability calculated examples characteristic vector similarity, when being mapped into the compact Hash codes of different segmentations in same bucket, use penalty factorAllow
Retrieval result error image and the distance between input inquiry image are farther；Dis is smaller to show that similarity is higher.
Described cross-module state search method is that image and text data are passed through into spy by building one group of deep neural network
The mode of sign study maps to common semantic space, to realize that the semanteme to different modalities data couples；Using depth convolution
Neutral net extracts the semantic feature of image modalities directly from input picture, text is represented by the way of term vector, with one
Tie up the semantic feature that convolutional neural networks extract text modality from term vector expression；First, depth convolutional neural networks are passed through
The dynamic generation compact Hash codes f of segmentation of vehicleA；Then, by text generation searching characteristic vector, so both each self-generatings
Characteristic vector can just be retrieved using same searching system, specific implementation process it is as shown in Figure 3.
The semantic feature of described text modality is that characteristic vector is extracted from text, first as the extraction algorithm first step
First need to split text；Entry of the characteristic vector of text from text, specific steps：
Input：Text O；Output：One group of rough similar image；
STEP1：Initialization：(1) text file analysis is vectorial into entry；(2) small word, repetitor are removed；(3) entry is checked
Ensure the correctness of parsing；
STEP2：The minimum vectorial R=(r of entry of random combine are taken out from O1,r2,...,rn)；
STEP3：To R and fAOrder and the compact Hash codes of segmentation are integrated, and obtain text attribute featureNow
fATxtDimension is less than R dimension；
STEP4：The hash algorithm that sorted again using local sensitivity is retrieved；
STEP5：Return to similar image group I；
Wherein, text attribute characteristic functionIt is expressed as with formula (16)：
In formula, ATBe expressed as the transposed matrix of the compact Hash codes of vehicle subsection, R be expressed as the entry minimum of random combine to
Amount,For text attribute characteristic function, it is sign function that sign, which is represented,；
In formula,Diagonal matrix is sought in diag expressions,Expression is that characteristic vector is extracted from text, car
The compact Hash codes A of segmentationTIt is initialized as complete 1 vector of (1 × c).
The foregoing is only the preferable implementation example of the present invention, be not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent substitution and improvements made etc., it should be included in the scope of the protection.
Claims (10)
1. one kind is based on the quick Hash vehicle retrieval method of multitask deep learning, it is characterised in that：The vehicle retrieval method
Comprise the following steps：
The first step, build the multitask depth convolutional neural networks for deep learning and training identification；
Second step, using the Feature fusion for being segmented compact Hash codes and example aspects；
3rd step, using local sensitivity Hash sort algorithm again；
4th step, using cross-module state search method, realize vehicle retrieval.
2. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 1, it is characterised in that：It is described
In the first step, using facilities networks of the Faster R-CNN as multitask convolutional neural networks；Network foremost is volume 3 × 3
Product network, referred to as conv1, stack convolution module followed by 4 and be named as conv2_x to conv5_x, each module is wrapped respectively
{ 2,3,3,3 } individual unit is contained, conv1 to conv4_3 is as shared network；RPN is followed by, i.e. network, RPN nets are suggested in region
Network exports the set of rectangular target Suggestion box, each frame includes 4 position coordinateses and become using any scalogram picture as input
Amount and a score；It is sliding in the convolution Feature Mapping of last shared convolutional layer output for formation zone Suggestion box
Small network is moved, this network is connected in n × n of input convolution Feature Mapping spatial window entirely；Each sliding window mapping
Onto a low-dimensional vector, the corresponding numerical value of a sliding window of each Feature Mapping；This vector exports same to two
The layer of the full connection of level；
Described RPN；RPN networks export the set of rectangular target Suggestion box, each using any scalogram picture as input
Frame includes 4 position coordinates variables and a score；The target of described rectangular target Suggestion box refers to Vehicle Object；
It is the estimated probability of target/non-targeted to each Suggestion box, is the classification layer realized with the softmax layers of two classification；K
Suggestion box is parameterized by the corresponding k Suggestion box for being referred to as anchor；
Each anchor is centered on current sliding window mouth center, and a kind of corresponding yardstick and length-width ratio, uses 3 kinds of yardsticks and 3
Kind length-width ratio, so just has k=9 anchor in each sliding position；
In order to train RPN networks, a binary label is distributed to each anchor, is to mark the anchor with this
Target；Then distribute positive label and give this two class anchor：(I) have with some real target bounding box, i.e. Ground Truth, GT
The ratio between highest IoU, i.e. Interse-ction-over-Union, common factor union, overlapping anchor；(II) with any GT bags
Enclosing box has the overlapping anchor of the IoU more than 0.7；Notice that a GT bounding box may distribute positive label to multiple anchor；
The IoU ratios that the negative label of distribution gives all GT bounding boxs are below 0.3 anchor；Anon-normal non-negative anchor is to training mesh
No any effect is marked, then is abandoned；
The multitask loss in Faster R-CNN is followed, minimizes object function；The loss function of one image is defined as：
<mrow>
<mi>L</mi>
<mrow>
<mo>(</mo>
<mo>{</mo>
<msub>
<mi>p</mi>
<mi>i</mi>
</msub>
<mo>}</mo>
<mo>,</mo>
<mo>{</mo>
<msub>
<mi>t</mi>
<mi>i</mi>
</msub>
<mo>}</mo>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<msub>
<mi>N</mi>
<mrow>
<mi>c</mi>
<mi>l</mi>
<mi>s</mi>
</mrow>
</msub>
</mfrac>
<munder>
<mo>&Sigma;</mo>
<mi>i</mi>
</munder>
<msub>
<mi>L</mi>
<mrow>
<mi>c</mi>
<mi>l</mi>
<mi>s</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>p</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msubsup>
<mi>p</mi>
<mi>i</mi>
<mo>*</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>+</mo>
<mi>&lambda;</mi>
<mfrac>
<mn>1</mn>
<msub>
<mi>N</mi>
<mrow>
<mi>r</mi>
<mi>e</mi>
<mi>g</mi>
</mrow>
</msub>
</mfrac>
<munder>
<mo>&Sigma;</mo>
<mi>i</mi>
</munder>
<msubsup>
<mi>p</mi>
<mi>i</mi>
<mo>*</mo>
</msubsup>
<msub>
<mi>L</mi>
<mrow>
<mi>r</mi>
<mi>e</mi>
<mi>g</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>t</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msubsup>
<mi>t</mi>
<mi>i</mi>
<mo>*</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
Here, i is anchor index, piIt is the prediction probability that anchor is the i-th target, if anchor is just, GT marks
LabelIt is exactly 1, if anchor is negative,It is exactly 0；tiIt is a vector, represents 4 parametrization coordinates of the bounding box of prediction,It is the coordinate vector of GT bounding boxs corresponding with positive anchor；λ is a balance weight, NclsThe normalized value for being cls items is
Mini-batch size, NregBe reg items normalized value be anchor positions quantity, Classification Loss function LclsIt is two
The logarithm loss of classification, i.e. motor vehicles target vs. road backgrounds：
<mrow>
<msub>
<mi>L</mi>
<mrow>
<mi>c</mi>
<mi>l</mi>
<mi>s</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>p</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msubsup>
<mi>p</mi>
<mi>i</mi>
<mo>*</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>-</mo>
<mi>l</mi>
<mi>o</mi>
<mi>g</mi>
<mo>&lsqb;</mo>
<msubsup>
<mi>p</mi>
<mi>i</mi>
<mo>*</mo>
</msubsup>
<msub>
<mi>p</mi>
<mi>i</mi>
</msub>
<mo>+</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<msubsup>
<mi>p</mi>
<mi>i</mi>
<mo>*</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<msub>
<mi>p</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>&rsqb;</mo>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mo>)</mo>
</mrow>
</mrow>
For returning loss function Lreg, defined to minor function：
<mrow>
<msub>
<mi>L</mi>
<mrow>
<mi>r</mi>
<mi>e</mi>
<mi>g</mi>
</mrow>
</msub>
<mo>=</mo>
<mrow>
<mo>(</mo>
<msub>
<mi>t</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msubsup>
<mi>t</mi>
<mi>i</mi>
<mo>*</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mi>R</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>t</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msubsup>
<mi>t</mi>
<mi>i</mi>
<mo>*</mo>
</msubsup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>3</mn>
<mo>)</mo>
</mrow>
</mrow>
In formula, LregTo return loss function, R is the loss function of robust, and smooth L are calculated with formula (4)1；
<mrow>
<msub>
<mi>smooth</mi>
<mrow>
<mi>L</mi>
<mn>1</mn>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<mn>0.5</mn>
<msup>
<mi>x</mi>
<mn>2</mn>
</msup>
</mrow>
</mtd>
<mtd>
<mtable>
<mtr>
<mtd>
<mrow>
<mi>i</mi>
<mi>f</mi>
</mrow>
</mtd>
<mtd>
<mrow>
<mo>|</mo>
<mi>x</mi>
<mo>|</mo>
<mo><</mo>
<mn>1</mn>
</mrow>
</mtd>
</mtr>
</mtable>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mo>|</mo>
<mi>x</mi>
<mo>|</mo>
<mo>-</mo>
<mn>0.5</mn>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>o</mi>
<mi>t</mi>
<mi>h</mi>
<mi>e</mi>
<mi>r</mi>
<mi>w</mi>
<mi>i</mi>
<mi>s</mi>
<mi>e</mi>
<mo>,</mo>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>4</mn>
<mo>)</mo>
</mrow>
</mrow>
In formula, smoothL1For smooth L1Loss function, x are variable.
3. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 1 or 2, it is characterised in that：
In the multitask depth convolutional neural networks, it is vital to design a kind of multitask object function；Multitask target letter
Several formula (5) represent；
<mrow>
<munder>
<mrow>
<mi>arg</mi>
<mi> </mi>
<mi>min</mi>
</mrow>
<msubsup>
<mrow>
<mo>{</mo>
<msup>
<mi>w</mi>
<mi>t</mi>
</msup>
<mo>}</mo>
</mrow>
<mrow>
<mi>t</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>T</mi>
</msubsup>
</munder>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>t</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>T</mi>
</munderover>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>N</mi>
</munderover>
<mi>L</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mi>y</mi>
<mi>i</mi>
<mi>t</mi>
</msubsup>
<mo>,</mo>
<mi>f</mi>
<mo>(</mo>
<mrow>
<msubsup>
<mi>x</mi>
<mi>i</mi>
<mi>t</mi>
</msubsup>
<mo>;</mo>
<msup>
<mi>w</mi>
<mi>t</mi>
</msup>
</mrow>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mo>+</mo>
<mi>&Phi;</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>w</mi>
<mi>t</mi>
</msup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>5</mn>
<mo>)</mo>
</mrow>
</mrow>
In formula,It is input feature valueWith weight parameter wtMapping function, L () is loss function, Φ (wt) be
The regularization value of weight parameter, T are general assignment number, are designated as the training data of t-th of taskWherein t ∈ (1, T),
I ∈ (1, N), N are total number of training,The respectively characteristic vector of the i-th sample and mark label；
For loss function, coordinate log-likelihood cost function to train the feature of last layer using softmax, realize multitask
Image classification, softmax loss functions are defined with formula (6),
<mrow>
<msub>
<mi>L</mi>
<mi>S</mi>
</msub>
<mo>=</mo>
<mo>-</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>m</mi>
</munderover>
<mi>log</mi>
<mfrac>
<msup>
<mi>e</mi>
<mrow>
<msubsup>
<mi>W</mi>
<mrow>
<mi>j</mi>
<mi>i</mi>
</mrow>
<mi>T</mi>
</msubsup>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>+</mo>
<msub>
<mi>b</mi>
<mrow>
<mi>y</mi>
<mi>i</mi>
</mrow>
</msub>
</mrow>
</msup>
<mrow>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</munderover>
<msup>
<mi>e</mi>
<mrow>
<msubsup>
<mi>W</mi>
<mi>j</mi>
<mi>T</mi>
</msubsup>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>+</mo>
<msub>
<mi>b</mi>
<mi>j</mi>
</msub>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>6</mn>
<mo>)</mo>
</mrow>
</mrow>
In formula, xiFor the i-th depth characteristic, WjFor the jth row of weight in last full articulamentum, b is bias term, and m, n are respectively
Handle sample size and classification number.
4. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 1 or 2, it is characterised in that：
In the second step, the Feature fusion process for being segmented compact Hash codes and example aspects is as follows：
In vehicle image feature extraction phases, first, by softmax activation primitives by threshold restriction between [0,1]；So
Afterwards, promote the output of binary system Hash codes by fragmentation threshold function, Hash codes are reduced using speced learning and coding strategy
Between redundancy improve feature robustness；Finally, by way of Fusion Features by speced learning come Hash codes melted
Conjunction finally gives vehicle characteristics and is segmented compact Hash codes；
For the example aspects of vehicle, implementation method is：Shared respectively from conv2_x to conv5_x and stack each module of convolution
Last unit combination RPN network output, add pyramid pond layer and vectorial flat layer it is various sizes of to adapt to
Convolution characteristic pattern input, while by convolution three-dimensional feature it is flat turn to one-dimensional characteristic vector, the vector be referred to as vehicle example spy
Sign；
Finally, by the compact feature and example aspects of the compact Hash codes of vehicle subsection merge to obtain again feature for retrieval to
Amount.
5. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 4, it is characterised in that：It is described
Vehicle characteristics be segmented compact Hash codes and be realized by the following method；If sharing T task, c be present under each tasktIt is individual
Classification, use mtRepresent the full connection output vector of each task, using softmax activation primitives make the output of full articulamentum [0,
1] between, calculated with formula (7)；
<mrow>
<msup>
<mi>q</mi>
<mi>t</mi>
</msup>
<mo>=</mo>
<mi>s</mi>
<mi>o</mi>
<mi>f</mi>
<mi>t</mi>
<mi>max</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>m</mi>
<mi>t</mi>
</msup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<msup>
<mi>e</mi>
<mrow>
<msup>
<mi>&theta;</mi>
<mi>T</mi>
</msup>
<msup>
<mi>m</mi>
<mi>t</mi>
</msup>
</mrow>
</msup>
<mrow>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<msup>
<mi>c</mi>
<mi>t</mi>
</msup>
</msubsup>
<msup>
<mi>e</mi>
<mrow>
<msup>
<mi>&theta;</mi>
<mi>T</mi>
</msup>
<msubsup>
<mi>m</mi>
<mi>i</mi>
<mi>t</mi>
</msubsup>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>7</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, θ represents random hyperplane, mtRepresent the full connection output vector of each task, ctRepresent existing under each task
Classification, qtRepresent full articulamentum output；
To encourage the output of the binary system of segment encoding module, threshold segmentation function binaryzation is reused：
<mrow>
<msup>
<mi>H</mi>
<mi>t</mi>
</msup>
<mo>=</mo>
<mi>H</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>q</mi>
<mi>t</mi>
</msup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mn>1</mn>
</mtd>
<mtd>
<mtable>
<mtr>
<mtd>
<mrow>
<mi>i</mi>
<mi>f</mi>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>max</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>q</mi>
<mi>t</mi>
</msup>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
<mtd>
<mrow>
<mi>o</mi>
<mi>t</mi>
<mi>h</mi>
<mi>e</mi>
<mi>r</mi>
<mi>w</mi>
<mi>i</mi>
<mi>s</mi>
<mi>e</mi>
<mo>,</mo>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>8</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, qtRepresent full articulamentum output, HtRepresent the binary system output of excitation segment encoding module；
Finally by HtFusion turns into the compact Hash codes vector f of vehicle subsectionA：
fA=[α1H1；α2H2；...；αtHt] (9)
Wherein, fARepresent the compact Hash code vector of vehicle subsection, αtCoefficient is represented, is calculated with formula (10), HtRepresent excitation
The binary system output of segment encoding module, t ∈ (1, T)：
<mrow>
<msup>
<mi>&alpha;</mi>
<mi>t</mi>
</msup>
<mo>=</mo>
<mn>1</mn>
<mo>-</mo>
<mfrac>
<mrow>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<msup>
<mi>c</mi>
<mi>t</mi>
</msup>
</msubsup>
<msubsup>
<mi>c</mi>
<mi>j</mi>
<mi>t</mi>
</msubsup>
</mrow>
<mrow>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>T</mi>
</msubsup>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<msup>
<mi>c</mi>
<mi>t</mi>
</msup>
</msubsup>
<msubsup>
<mi>c</mi>
<mi>j</mi>
<mi>i</mi>
</msubsup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>10</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, αtCoefficient is represented, factor alpha is multiplied by before H vectorstFor make up between different task because classify it is uneven caused by by mistake
Difference.
6. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 5, it is characterised in that：It is described
In 3rd step, the characteristic vector for retrieval is to be merged the compact feature and example aspects of the compact Hash codes of vehicle subsection
Obtain, process is as follows：
{ 4 are selected respectively for conv2_x to conv5_x bottommost layer2,82,162,162Output Size as characteristic pattern；It is right
It is h × w in given input picture I size, convolution convx_x is activated as three-dimensional vector T, and size is h ' × w ' × d, comprising
A series of two dimensional character figure S=S { Sn, n ∈ (1, d), SnSize correspond to the big of n-th channel characteristics figure for h ' × w '
It is small；T to be sent into pyramid pond layer again and obtains three-dimensional vector T ', size is l × l × d, still comprising series of features figure S '=
S′{S′n, n ∈ (1, d), S 'nSize be l × l, to each S 'nIt is that k × k sliding windows carry out traversal selection most using size
Big value, S 'nIt is changed into l/k × l/k, then the S ' to each passagenMerged to obtain one-dimensional vector, phase is carried out to d passage successively
Biconditional operation, the personal characteristics vector f finally obtainedBSize is (1, l/k × d)；Final searching characteristic vector f, computational methods
As shown in formula (11)；
F=[fA；fB] (11)
In formula, f is used for the characteristic vector of vehicle retrieval, fBExample aspects vector, i.e. personal characteristics vector, fARepresent vehicle subsection
Compact Hash code vector.
7. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 6, it is characterised in that：It is described
In 3rd step, similar sample is mapped in the same same bucket with high probability；The Hash letter of local sensitivity Hash
Number h () meets following condition：
s{h(fAq)=h (fA)=sim (fAq,fA) (12)
In formula, sim (fAq,fA) represent fAqWith fASimilarity, h (fA) represent fAHash function, h (fAq) represent fAqHash
Function, similarity measurement therein and a distance function σ direct correlation, are calculated with formula (13),
<mrow>
<mi>s</mi>
<mi>i</mi>
<mi>m</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>f</mi>
<mrow>
<mi>A</mi>
<mi>q</mi>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>f</mi>
<mi>A</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mi>exp</mi>
<mo>{</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>f</mi>
<mrow>
<mi>A</mi>
<mi>q</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mi>A</mi>
</msub>
<mo>|</mo>
<mo>|</mo>
</mrow>
<msup>
<mi>&sigma;</mi>
<mn>2</mn>
</msup>
</mfrac>
<mo>}</mo>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>13</mn>
<mo>)</mo>
</mrow>
</mrow>
The typical classification of local sensitivity hash function is provided by accidental projection and threshold value, is calculated with formula (14),
h(fA)=sign (WfA+b) (14)
In formula, W is a random hyperplane vector, and b is a random intercept.
8. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 7, it is characterised in that：It is described
In 3rd step, the example aspects knot of image is recycled after query image is mapped in similar bucket by the compact Hash codes of segmentation
Formula (15) is closed to sort to the image returned in bucket again；Shown in the computational methods that sort again such as formula (15)：
In formula, k represents k-th of image in bucket,Represent penalty factor andCos represents COS distance formula, and y represents to reflect
F before penetratingAqWithIt is whether equal；Y is 1 if equal, is otherwise 0,Represent k-th of compact Hash of image vehicle subsection
Code vector, fAqRepresent the compact Hash code vector of vehicle subsection after sorting againAllow retrieval result error image and input inquiry figure
The distance between picture is farther；Dis is smaller to show that similarity is higher.
9. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 1 or 2, it is characterised in that：
In 4th step, cross-module state search method is that image and text data are passed through into feature by building one group of deep neural network
The mode of study maps to common semantic space, to realize that the semanteme to different modalities data couples；Using depth convolution god
The semantic feature of image modalities is extracted directly from input picture through network, represents text by the way of term vector, use is one-dimensional
Convolutional neural networks extract the semantic feature of text modality from term vector expression；First, moved by depth convolutional neural networks
State generates the compact Hash codes f of segmentation of vehicleA；Then, by text generation searching characteristic vector, so both each self-generatings
Characteristic vector using same searching system with regard to can be retrieved.
10. it is based on the quick Hash vehicle retrieval method of multitask deep learning as claimed in claim 9, it is characterised in that：Institute
The semantic feature for the text modality stated is that characteristic vector is extracted from text, as the extraction algorithm first step firstly the need of fractionation
Text；Entry of the characteristic vector of text from text, specific steps：
Input：Text O；Output：One group of rough similar image；
STEP1：Initialization：(1) text file analysis is vectorial into entry；(2) small word, repetitor are removed；(3) check that entry ensures
The correctness of parsing；
STEP2：The minimum vectorial R=(r of entry of random combine are taken out from O1,r2,...,rn)；
STEP3：To R and fAOrder and the compact Hash codes of segmentation are integrated, and obtain text attribute featureF nowATxt
Dimension is less than R dimension；
STEP4：The hash algorithm that sorted again using local sensitivity is retrieved；
STEP5：Return to similar image group I；
Wherein, text attribute characteristic functionIt is expressed as with formula (16)：
In formula, ATThe transposed matrix of the compact Hash codes of vehicle subsection is expressed as, R is expressed as the minimum vector of entry of random combine,For text attribute characteristic function, it is sign function that sign, which is represented,；
In formula,Diagonal matrix is sought in diag expressions,Expression is that characteristic vector is extracted from text, vehicle point
The compact Hash codes A of sectionTIt is initialized as complete 1 vector of (1 × c).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710857318.5A CN107885764B (en) | 2017-09-21 | 2017-09-21 | Rapid Hash vehicle retrieval method based on multitask deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710857318.5A CN107885764B (en) | 2017-09-21 | 2017-09-21 | Rapid Hash vehicle retrieval method based on multitask deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107885764A true CN107885764A (en) | 2018-04-06 |
| CN107885764B CN107885764B (en) | 2020-12-18 |
Family
ID=61780800
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710857318.5A Active CN107885764B (en) | 2017-09-21 | 2017-09-21 | Rapid Hash vehicle retrieval method based on multitask deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107885764B (en) |
Cited By (34)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108629414A (en) * | 2018-05-09 | 2018-10-09 | 清华大学 | depth hash learning method and device |
| CN109033172A (en) * | 2018-06-21 | 2018-12-18 | 西安理工大学 | A kind of image search method of deep learning and approximate target positioning |
| CN109034245A (en) * | 2018-07-27 | 2018-12-18 | 燕山大学 | A kind of object detection method merged using characteristic pattern |
| CN109035267A (en) * | 2018-06-22 | 2018-12-18 | 华东师范大学 | A kind of image object based on deep learning takes method |
| CN109086866A (en) * | 2018-07-02 | 2018-12-25 | 重庆大学 | A kind of part two-value convolution method suitable for embedded device |
| CN109145798A (en) * | 2018-08-13 | 2019-01-04 | 浙江零跑科技有限公司 | A kind of Driving Scene target identification and travelable region segmentation integrated approach |
| CN109583305A (en) * | 2018-10-30 | 2019-04-05 | 南昌大学 | A kind of advanced method that the vehicle based on critical component identification and fine grit classification identifies again |
| CN109766884A (en) * | 2018-12-26 | 2019-05-17 | 哈尔滨工程大学 | A kind of airfield runway foreign matter detecting method based on Faster-RCNN |
| CN109857887A (en) * | 2018-12-10 | 2019-06-07 | 福州大学 | A kind of method that photograph album personage classifies automatically |
| CN109886286A (en) * | 2019-01-03 | 2019-06-14 | 武汉精测电子集团股份有限公司 | Object detection method, target detection model and system based on cascade detectors |
| CN109885716A (en) * | 2019-02-18 | 2019-06-14 | 成都快眼科技有限公司 | The image search method of discrete Hash is supervised based on heterogeneous multi-task learning depth |
| CN109933682A (en) * | 2019-01-11 | 2019-06-25 | 上海交通大学 | A kind of image Hash search method and system based on semanteme in conjunction with content information |
| CN109961057A (en) * | 2019-04-03 | 2019-07-02 | 罗克佳华科技集团股份有限公司 | A kind of vehicle location preparation method and device |
| CN109994201A (en) * | 2019-03-18 | 2019-07-09 | 浙江大学 | A kind of diabetes based on deep learning and hypertension method for calculating probability |
| CN110019652A (en) * | 2019-03-14 | 2019-07-16 | 九江学院 | A kind of cross-module state Hash search method based on deep learning |
| CN110163106A (en) * | 2019-04-19 | 2019-08-23 | 中国科学院计算技术研究所 | Integral type is tatooed detection and recognition methods and system |
| CN110222140A (en) * | 2019-04-22 | 2019-09-10 | 中国科学院信息工程研究所 | A kind of cross-module state search method based on confrontation study and asymmetric Hash |
| CN110298404A (en) * | 2019-07-02 | 2019-10-01 | 西南交通大学 | A kind of method for tracking target based on triple twin Hash e-learnings |
| CN110442741A (en) * | 2019-07-22 | 2019-11-12 | 成都澳海川科技有限公司 | A kind of mutual search method of cross-module state picture and text for merging and reordering based on tensor |
| CN110516640A (en) * | 2019-08-30 | 2019-11-29 | 华侨大学 | It is a kind of to combine the vehicle indicated discrimination method again based on feature pyramid |
| CN110532456A (en) * | 2019-06-14 | 2019-12-03 | 平安科技（深圳）有限公司 | Case querying method, device, computer equipment and storage medium |
| CN110765281A (en) * | 2019-11-04 | 2020-02-07 | 山东浪潮人工智能研究院有限公司 | Multi-semantic depth supervision cross-modal Hash retrieval method |
| CN111639240A (en) * | 2020-05-14 | 2020-09-08 | 山东大学 | Cross-modal Hash retrieval method and system based on attention awareness mechanism |
| CN112115291A (en) * | 2020-08-12 | 2020-12-22 | 南京止善智能科技研究院有限公司 | Three-dimensional indoor model retrieval method based on deep learning |
| CN112307241A (en) * | 2019-07-26 | 2021-02-02 | 丰田自动车株式会社 | Search device, learning device, search system, and recording medium |
| CN112418168A (en) * | 2020-12-10 | 2021-02-26 | 深圳云天励飞技术股份有限公司 | Vehicle identification method, device, system, electronic equipment and storage medium |
| CN112580569A (en) * | 2020-12-25 | 2021-03-30 | 山东旗帜信息有限公司 | Vehicle weight identification method and device based on multi-dimensional features |
| CN112614187A (en) * | 2020-12-31 | 2021-04-06 | 深圳市优必选科技股份有限公司 | Loop detection method, device, terminal equipment and readable storage medium |
| CN114220053A (en) * | 2021-12-15 | 2022-03-22 | 北京建筑大学 | Unmanned aerial vehicle video vehicle retrieval method based on vehicle feature matching |
| CN114332745A (en) * | 2022-03-11 | 2022-04-12 | 西安科技大学 | Near-repetitive video big data cleaning method based on deep neural network |
| CN114494818A (en) * | 2022-01-26 | 2022-05-13 | 北京百度网讯科技有限公司 | Image processing method, model training method, related device and electronic equipment |
| US11429820B2 (en) * | 2018-03-13 | 2022-08-30 | Recogni Inc. | Methods for inter-camera recognition of individuals and their properties |
| CN115512154A (en) * | 2022-09-21 | 2022-12-23 | 东南大学 | Highway vehicle image retrieval method based on deep learning neural network |
| CN115686868A (en) * | 2022-12-28 | 2023-02-03 | 中南大学 | Cross-node-oriented multi-mode retrieval method based on federated hash learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106227851A (en) * | 2016-07-29 | 2016-12-14 | 汤平 | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end |
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN106528662A (en) * | 2016-10-20 | 2017-03-22 | 中山大学 | Quick retrieval method and system of vehicle image on the basis of feature geometric constraint |
-
2017
- 2017-09-21 CN CN201710857318.5A patent/CN107885764B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106250812A (en) * | 2016-07-15 | 2016-12-21 | 汤平 | A kind of model recognizing method based on quick R CNN deep neural network |
| CN106227851A (en) * | 2016-07-29 | 2016-12-14 | 汤平 | Based on the image search method searched for by depth of seam division that degree of depth convolutional neural networks is end-to-end |
| CN106528662A (en) * | 2016-10-20 | 2017-03-22 | 中山大学 | Quick retrieval method and system of vehicle image on the basis of feature geometric constraint |
Non-Patent Citations (2)
| Title |
|---|
| PENG YISHU 等: "BINARY CODING-BASED VEHICLE IMAGE CLASSIFICATION", 《IEEE》 * |
| SHAOQING REN 等: "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks", 《IEEE》 * |
Cited By (53)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11429820B2 (en) * | 2018-03-13 | 2022-08-30 | Recogni Inc. | Methods for inter-camera recognition of individuals and their properties |
| CN108629414A (en) * | 2018-05-09 | 2018-10-09 | 清华大学 | depth hash learning method and device |
| CN109033172A (en) * | 2018-06-21 | 2018-12-18 | 西安理工大学 | A kind of image search method of deep learning and approximate target positioning |
| CN109033172B (en) * | 2018-06-21 | 2021-12-17 | 西安理工大学 | Image retrieval method for deep learning and approximate target positioning |
| CN109035267A (en) * | 2018-06-22 | 2018-12-18 | 华东师范大学 | A kind of image object based on deep learning takes method |
| CN109086866A (en) * | 2018-07-02 | 2018-12-25 | 重庆大学 | A kind of part two-value convolution method suitable for embedded device |
| CN109086866B (en) * | 2018-07-02 | 2021-07-30 | 重庆大学 | Partial binary convolution method suitable for embedded equipment |
| CN109034245A (en) * | 2018-07-27 | 2018-12-18 | 燕山大学 | A kind of object detection method merged using characteristic pattern |
| CN109145798A (en) * | 2018-08-13 | 2019-01-04 | 浙江零跑科技有限公司 | A kind of Driving Scene target identification and travelable region segmentation integrated approach |
| CN109145798B (en) * | 2018-08-13 | 2021-10-22 | 浙江零跑科技股份有限公司 | Driving scene target identification and travelable region segmentation integration method |
| CN109583305A (en) * | 2018-10-30 | 2019-04-05 | 南昌大学 | A kind of advanced method that the vehicle based on critical component identification and fine grit classification identifies again |
| CN109857887A (en) * | 2018-12-10 | 2019-06-07 | 福州大学 | A kind of method that photograph album personage classifies automatically |
| CN109766884A (en) * | 2018-12-26 | 2019-05-17 | 哈尔滨工程大学 | A kind of airfield runway foreign matter detecting method based on Faster-RCNN |
| CN109886286A (en) * | 2019-01-03 | 2019-06-14 | 武汉精测电子集团股份有限公司 | Object detection method, target detection model and system based on cascade detectors |
| CN109933682A (en) * | 2019-01-11 | 2019-06-25 | 上海交通大学 | A kind of image Hash search method and system based on semanteme in conjunction with content information |
| CN109885716B (en) * | 2019-02-18 | 2022-10-14 | 成都快眼科技有限公司 | Image retrieval method based on heterogeneous multi-task learning deep supervised discrete hashing |
| CN109885716A (en) * | 2019-02-18 | 2019-06-14 | 成都快眼科技有限公司 | The image search method of discrete Hash is supervised based on heterogeneous multi-task learning depth |
| CN110019652A (en) * | 2019-03-14 | 2019-07-16 | 九江学院 | A kind of cross-module state Hash search method based on deep learning |
| CN110019652B (en) * | 2019-03-14 | 2022-06-03 | 九江学院 | Cross-modal Hash retrieval method based on deep learning |
| CN109994201A (en) * | 2019-03-18 | 2019-07-09 | 浙江大学 | A kind of diabetes based on deep learning and hypertension method for calculating probability |
| CN109994201B (en) * | 2019-03-18 | 2021-06-11 | 浙江大学 | Diabetes and hypertension probability calculation system based on deep learning |
| CN109961057B (en) * | 2019-04-03 | 2021-09-03 | 罗克佳华科技集团股份有限公司 | Vehicle position obtaining method and device |
| CN109961057A (en) * | 2019-04-03 | 2019-07-02 | 罗克佳华科技集团股份有限公司 | A kind of vehicle location preparation method and device |
| CN110163106A (en) * | 2019-04-19 | 2019-08-23 | 中国科学院计算技术研究所 | Integral type is tatooed detection and recognition methods and system |
| CN110222140A (en) * | 2019-04-22 | 2019-09-10 | 中国科学院信息工程研究所 | A kind of cross-module state search method based on confrontation study and asymmetric Hash |
| CN110222140B (en) * | 2019-04-22 | 2021-07-13 | 中国科学院信息工程研究所 | Cross-modal retrieval method based on counterstudy and asymmetric hash |
| CN110532456B (en) * | 2019-06-14 | 2023-06-27 | 平安科技（深圳）有限公司 | Case query method, device, computer equipment and storage medium |
| CN110532456A (en) * | 2019-06-14 | 2019-12-03 | 平安科技（深圳）有限公司 | Case querying method, device, computer equipment and storage medium |
| CN110298404A (en) * | 2019-07-02 | 2019-10-01 | 西南交通大学 | A kind of method for tracking target based on triple twin Hash e-learnings |
| CN110442741B (en) * | 2019-07-22 | 2022-10-18 | 成都澳海川科技有限公司 | Tensor fusion and reordering-based cross-modal image-text mutual search method |
| CN110442741A (en) * | 2019-07-22 | 2019-11-12 | 成都澳海川科技有限公司 | A kind of mutual search method of cross-module state picture and text for merging and reordering based on tensor |
| CN112307241A (en) * | 2019-07-26 | 2021-02-02 | 丰田自动车株式会社 | Search device, learning device, search system, and recording medium |
| CN110516640B (en) * | 2019-08-30 | 2022-09-30 | 华侨大学 | Vehicle re-identification method based on feature pyramid joint representation |
| CN110516640A (en) * | 2019-08-30 | 2019-11-29 | 华侨大学 | It is a kind of to combine the vehicle indicated discrimination method again based on feature pyramid |
| CN110765281A (en) * | 2019-11-04 | 2020-02-07 | 山东浪潮人工智能研究院有限公司 | Multi-semantic depth supervision cross-modal Hash retrieval method |
| CN111639240B (en) * | 2020-05-14 | 2021-04-09 | 山东大学 | Cross-modal Hash retrieval method and system based on attention awareness mechanism |
| CN111639240A (en) * | 2020-05-14 | 2020-09-08 | 山东大学 | Cross-modal Hash retrieval method and system based on attention awareness mechanism |
| CN112115291A (en) * | 2020-08-12 | 2020-12-22 | 南京止善智能科技研究院有限公司 | Three-dimensional indoor model retrieval method based on deep learning |
| CN112115291B (en) * | 2020-08-12 | 2024-02-27 | 南京止善智能科技研究院有限公司 | Three-dimensional indoor model retrieval method based on deep learning |
| CN112418168B (en) * | 2020-12-10 | 2024-04-02 | 深圳云天励飞技术股份有限公司 | Vehicle identification method, device, system, electronic equipment and storage medium |
| CN112418168A (en) * | 2020-12-10 | 2021-02-26 | 深圳云天励飞技术股份有限公司 | Vehicle identification method, device, system, electronic equipment and storage medium |
| CN112580569A (en) * | 2020-12-25 | 2021-03-30 | 山东旗帜信息有限公司 | Vehicle weight identification method and device based on multi-dimensional features |
| CN112580569B (en) * | 2020-12-25 | 2023-06-09 | 山东旗帜信息有限公司 | Vehicle re-identification method and device based on multidimensional features |
| CN112614187A (en) * | 2020-12-31 | 2021-04-06 | 深圳市优必选科技股份有限公司 | Loop detection method, device, terminal equipment and readable storage medium |
| CN112614187B (en) * | 2020-12-31 | 2024-03-26 | 深圳市优必选科技股份有限公司 | Loop detection method, loop detection device, terminal equipment and readable storage medium |
| CN114220053A (en) * | 2021-12-15 | 2022-03-22 | 北京建筑大学 | Unmanned aerial vehicle video vehicle retrieval method based on vehicle feature matching |
| CN114494818A (en) * | 2022-01-26 | 2022-05-13 | 北京百度网讯科技有限公司 | Image processing method, model training method, related device and electronic equipment |
| CN114494818B (en) * | 2022-01-26 | 2023-07-25 | 北京百度网讯科技有限公司 | Image processing method, model training method, related device and electronic equipment |
| CN114332745A (en) * | 2022-03-11 | 2022-04-12 | 西安科技大学 | Near-repetitive video big data cleaning method based on deep neural network |
| CN114332745B (en) * | 2022-03-11 | 2022-05-31 | 西安科技大学 | Near-repetitive video big data cleaning method based on deep neural network |
| CN115512154A (en) * | 2022-09-21 | 2022-12-23 | 东南大学 | Highway vehicle image retrieval method based on deep learning neural network |
| CN115686868B (en) * | 2022-12-28 | 2023-04-07 | 中南大学 | Cross-node-oriented multi-mode retrieval method based on federated hash learning |
| CN115686868A (en) * | 2022-12-28 | 2023-02-03 | 中南大学 | Cross-node-oriented multi-mode retrieval method based on federated hash learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107885764B (en) | 2020-12-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107885764A (en) | Based on the quick Hash vehicle retrieval method of multitask deep learning | |
| CN108108657B (en) | Method for correcting locality sensitive Hash vehicle retrieval based on multitask deep learning | |
| CN106227851B (en) | The image search method of depth of seam division search based on depth convolutional neural networks | |
| EP3327583B1 (en) | Method and device for searching a target in an image | |
| CN109784203B (en) | Method for inspecting contraband in weak supervision X-ray image based on layered propagation and activation | |
| CN106250812B (en) | A kind of model recognizing method based on quick R-CNN deep neural network | |
| Ni et al. | An improved deep network-based scene classification method for self-driving cars | |
| CN105138973B (en) | The method and apparatus of face authentication | |
| Chang et al. | Automatic license plate recognition | |
| CN108460328A (en) | A kind of fake-licensed car detection method based on multitask convolutional neural networks | |
| CN108764228A (en) | Word object detection method in a kind of image | |
| CN107729801A (en) | A kind of vehicle color identifying system based on multitask depth convolutional neural networks | |
| CN108595636A (en) | The image search method of cartographical sketching based on depth cross-module state correlation study | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN107330396A (en) | A kind of pedestrian's recognition methods again based on many attributes and many strategy fusion study | |
| CN108171136A (en) | A kind of multitask bayonet vehicle is to scheme to search the system and method for figure | |
| CN110033002A (en) | Detection method of license plate based on multitask concatenated convolutional neural network | |
| CN110097044A (en) | Stage car plate detection recognition methods based on deep learning | |
| CN110321862B (en) | Pedestrian re-identification method based on compact ternary loss | |
| CN110211127B (en) | Image partition method based on bicoherence network | |
| CN110008899B (en) | Method for extracting and classifying candidate targets of visible light remote sensing image | |
| Pei et al. | Localized traffic sign detection with multi-scale deconvolution networks | |
| Guo et al. | Traffic-sign spotting in the wild via deep features | |
| CN105404858A (en) | Vehicle type recognition method based on deep Fisher network | |
| CN106529391A (en) | Robust speed-limit traffic sign detection and recognition method |
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 | ||
| CP01 | Change in the name or title of a patent holder |
Address after: 310012 1st floor, building 1, 223 Yile Road, Hangzhou City, Zhejiang Province Patentee after: Yinjiang Technology Co.,Ltd. Address before: 310012 1st floor, building 1, 223 Yile Road, Hangzhou City, Zhejiang Province Patentee before: ENJOYOR Co.,Ltd. |
|
| CP01 | Change in the name or title of a patent holder |