CN108197326A - A kind of vehicle retrieval method and device, electronic equipment, storage medium - Google Patents
A kind of vehicle retrieval method and device, electronic equipment, storage medium Download PDFInfo
- Publication number
- CN108197326A CN108197326A CN201810119550.3A CN201810119550A CN108197326A CN 108197326 A CN108197326 A CN 108197326A CN 201810119550 A CN201810119550 A CN 201810119550A CN 108197326 A CN108197326 A CN 108197326A
- Authority
- CN
- China
- Prior art keywords
- vehicle
- sample
- image
- retrieved
- vehicle image
- 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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- 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
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Library & Information Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
The disclosure discloses a kind of vehicle retrieval method and device, electronic equipment, computer readable storage medium, and the program includes：Obtain vehicle image to be retrieved；The feature extraction of vehicle image to be retrieved is carried out by convolutional neural networks, obtains feature vector；Carry out the extraction of depth characteristic between attributive character extraction and relatively positive negative sample respectively to feature vector；Merge attributive character and depth characteristic, obtain the vehicle vision feature of vehicle image to be retrieved；The matching primitives of vehicle vision feature between image and vehicle image to be retrieved in vehicle retrieval database are carried out, obtain the target image with Car image matching to be retrieved.Technical solution provided by the invention, the accuracy higher of vehicle retrieval, since without manually participating in inquiry, vehicle retrieval is efficient.
Description
Technical field
This disclosure relates to image identification technical field, more particularly to a kind of vehicle retrieval method and device, electronic equipment, meter
Calculation machine readable storage medium storing program for executing.
Background technology
Vehicle to be checked is retrieved in magnanimity monitoring video, there is important application in public security system.Usual monitoring system meeting
Hundreds and thousands of or even up to ten thousand monitoring videos is preserved, vehicle to be checked is retrieved in so huge video recording, needs largely to transport
It calculates, longer time.Required operand and required time, with the increase of monitoring video quantity and the expansion of retrieval time range
Greatly, it increases sharply.
Following two schemes may be used in the vehicle retrieval in multiple monitoring videos in the prior art：
Scheme one, the mode of manual retrieval.Check whether vehicle to be checked occurs by human eye, the standard of this method retrieval
True rate is high but less efficient.When the monitoring video that need to be retrieved is less, this method may be used, when the monitoring record that need to be retrieved
As it is more when then need to take the plenty of time.
Scheme two, by vehicle recongnition technique, vehicle retrieval is carried out or based on artificially defined mark based on license board information
Object carries out vehicle retrieval including one or more in annual test mark, goods of furniture for display rather than for use, ornament, according to these mark object features.It is but suspicious
The possibility of vehicle missing license board information is larger, while there is likely to be deck phenomenons, when lacking license board information, will be unable to
Into vehicle retrieval.Characteristic body based on Manual definition, which carries out retrieval, has limitation, for example the goods of furniture for display rather than for use ornament of vehicle is often to become
Dynamic, the accuracy that vehicle retrieval is carried out based on these local features is not also high.
To sum up, the vehicle retrieval scheme provided in the prior art, recall precision is relatively low, and retrieval accuracy is not high.
Invention content
The problem of in order to solve the vehicle retrieval scheme of prior art offer, recall precision is relatively low, and retrieval accuracy is not high,
Present disclose provides a kind of vehicle retrieval method and devices.
On the one hand, the present invention provides a kind of vehicle retrieval method, the method includes：
Obtain vehicle image to be retrieved；
The feature extraction of the vehicle image to be retrieved is carried out by convolutional neural networks, obtains feature vector；
Carry out the extraction of depth characteristic between attributive character extraction and relatively positive negative sample respectively to described eigenvector；
Merge the attributive character and depth characteristic, obtain the vehicle vision feature of the vehicle image to be retrieved；
Carry out the matching meter of vehicle vision feature between image and the vehicle image to be retrieved in vehicle retrieval database
It calculates, obtains the target image with the Car image matching to be retrieved.
It is described to carry out attributive character extraction and relatively positive and negative respectively to described eigenvector in a kind of exemplary embodiment
The extraction of depth characteristic between sample, including：
The class prediction branch built by sample vehicle image and sample vehicle image institute flag attribute classification
Model carries out described eigenvector the attribute class prediction of the vehicle image to be retrieved, obtains attributive character.
It is described to carry out attributive character extraction and relatively positive and negative respectively to described eigenvector in a kind of exemplary embodiment
The extraction of depth characteristic between sample, including：
The positive and negative constraint branch model built by sample vehicle image and corresponding positive negative sample, to the feature to
Amount carries out depth characteristic extraction.
It is described to carry out attributive character extraction and relatively positive and negative respectively to described eigenvector in a kind of exemplary embodiment
Between sample before the extraction of depth characteristic, the method further includes：
Obtain the attribute classification that sample vehicle image and the sample vehicle image are marked；
Neural network learning is carried out by the sample vehicle image and institute's flag attribute classification, obtains and carries out the category
The class prediction branch model of property feature extraction.
It is described that god is carried out by the sample vehicle image and institute's flag attribute classification in a kind of exemplary embodiment
Through e-learning, the class prediction branch model for carrying out the attributive character extraction is obtained, including：
The sampling feature vectors of the sample vehicle image are extracted by the convolutional neural networks；
By build the full articulamentum of multilayer to the sampling feature vectors carry out convolutional calculation, extract the sample characteristics to
Measure corresponding sample attribute feature；
According to the attribute classification that the sample vehicle image is marked, optimize the weight parameter of the full articulamentum, make institute
The difference that sample attribute feature is stated between the attribute classification that is marked is minimum, and the full articulamentum after optimization forms the class
Other predicted branches model.
It is described to carry out attributive character extraction and relatively positive and negative respectively to described eigenvector in a kind of exemplary embodiment
Between sample before the extraction of depth characteristic, the method further includes：
Sample vehicle image and the corresponding positive negative sample of the sample vehicle image are obtained, forms sample triple；
The nerve net of relatively positive negative sample is carried out using the loss function for being adapted to matching primitives to the sample triple
Network learns, and obtains the positive and negative constraint branch model for carrying out the depth characteristic extraction.
In a kind of exemplary embodiment, the sample triple is carried out using the loss function for being adapted to matching primitives
The neural network learning of relatively positive negative sample obtains the positive and negative constraint branch model for carrying out the depth characteristic extraction, including：
The neural network learning of the depth characteristic extraction is carried out to the sample triple, by optimizing sample vehicle figure
Picture and the characteristic distance between positive negative sample, obtain the positive and negative constraint branch model for carrying out the depth characteristic extraction.
In a kind of exemplary embodiment, image and the vehicle image to be retrieved in the progress vehicle retrieval database
Between vehicle vision feature matching primitives, obtain with the target image of the Car image matching to be retrieved, including：
Image in vehicle retrieval database is carried out to the comparison of vehicle vision feature with the vehicle image to be retrieved；
It is searched in the vehicle retrieval database with the vehicle vision characteristic similarity of the vehicle image to be retrieved most
High image obtains the target image with the Car image matching to be retrieved.
On the other hand, the present invention also provides a kind of vehicle retrieval device, described device includes：
Image collection module, for obtaining vehicle image to be retrieved；
Vector obtains module, for passing through the feature extraction that convolutional neural networks carry out the vehicle image to be retrieved, obtains
Obtain feature vector；
Characteristic extracting module, for being carried out respectively between attributive character extraction and relatively positive negative sample to described eigenvector
The extraction of depth characteristic；
Feature merging module for merging the attributive character and depth characteristic, obtains the vehicle image to be retrieved
Vehicle vision feature；
Vehicle retrieval module, for carrying out in vehicle retrieval database vehicle between image and the vehicle image to be retrieved
The matching primitives of visual signature obtain the target image with the Car image matching to be retrieved.
In a kind of exemplary embodiment, the characteristic extracting module includes：
Attributive character extraction unit, for passing through sample vehicle image and sample vehicle image institute flag attribute class
The class prediction branch model not built carries out described eigenvector the attribute class prediction of the vehicle image to be retrieved,
Obtain attributive character.
In a kind of exemplary embodiment, the characteristic extracting module includes：
Depth characteristic extraction unit, for passing through the positive and negative constraint of sample vehicle image and corresponding positive negative sample structure
Branch model carries out depth characteristic extraction to described eigenvector.
In a kind of exemplary embodiment, described device further includes：
Sample image acquisition module, for obtaining the attribute that sample vehicle image and the sample vehicle image are marked
Classification；
Classification branch training module, for carrying out nerve net by the sample vehicle image and institute's flag attribute classification
Network learns, and obtains the class prediction branch model for carrying out the attributive character extraction.
In a kind of exemplary embodiment, classification branch training module includes：
Sample vector extraction unit, for passing through the sample spy that the convolutional neural networks extract the sample vehicle image
Sign vector；
Sample characteristics extraction unit, by being carried out based on convolution to the sampling feature vectors by building the full articulamentum of multilayer
It calculates, extracts the corresponding sample attribute feature of the sampling feature vectors；
Parameter optimization unit for the attribute classification marked according to the sample vehicle image, optimizes the full connection
The weight parameter of layer makes the difference between the sample attribute feature and the attribute classification marked minimum, described after optimization
Full articulamentum forms the class prediction branch model.
In a kind of exemplary embodiment, described device further includes：
Triple acquisition module, for obtaining sample vehicle image and the corresponding positive and negative sample of the sample vehicle image
This, forms sample triple；
Positive and negative bound branch training module, for using the loss function for being adapted to matching primitives to the sample triple
The neural network learning of relatively positive negative sample is carried out, obtains the positive and negative constraint branch model for carrying out the depth characteristic extraction.
In a kind of exemplary embodiment, the positive and negative bound branch training module includes：
Characteristic distance optimizes unit, for carrying out the Neural Network Science of the depth characteristic extraction to the sample triple
It practises, by optimizing the characteristic distance between sample vehicle image and positive negative sample, obtains and carrying out the depth characteristic extraction just
It breaks a promise beam branch model.
In a kind of exemplary embodiment, vehicle retrieval module includes：
Characteristic Contrast unit regards for image in vehicle retrieval database to be carried out vehicle with the vehicle image to be retrieved
Feel the comparison of feature；
Image searching unit, for searching the vehicle with the vehicle image to be retrieved in the vehicle retrieval database
The highest image of visual signature similarity obtains the target image with the Car image matching to be retrieved.
On the other hand, the present invention also provides a kind of electronic equipment, the electronic equipment includes：
Processor；
For storing the memory of processor-executable instruction；
Wherein, the processor is configured as performing above-mentioned vehicle retrieval method described in any one.
On the other hand, the present invention also provides a kind of computer readable storage medium, the computer readable storage mediums
It is stored with computer program, the computer program can be performed as processor completes above-mentioned vehicle retrieval side described in any one
Method.
The technical scheme provided by this disclosed embodiment can include the following benefits：
Technical solution provided by the invention, by extracting the depth characteristic and attributive character of vehicle image to be retrieved, and base
It is searched from vehicle retrieval database and the vehicle to be retrieved in the vehicle vision feature comprising the depth characteristic and attributive character
The target image of images match.Compared with the prior art, which contains the more features of vehicle image to be retrieved
Information, therefore the discrimination between other vehicle images is stronger, it is particularly possible to distinguish the different vehicle figure of same alike result feature
Picture is searched and the approximate target image of vehicle image to be retrieved, accuracy higher, due to need not based on the vehicle vision feature
It is artificial to participate in inquiry, improve the efficiency of vehicle retrieval.
It should be understood that above general description and following detailed description is only exemplary, this can not be limited
It is open.
Description of the drawings
Attached drawing herein is incorporated into specification and forms the part of this specification, shows the implementation for meeting the present invention
Example, and in specification together principle for explaining the present invention.
Fig. 1 is the schematic diagram according to the implementation environment involved by the disclosure；
Fig. 2 is the block diagram according to a kind of server shown in an exemplary embodiment；
Fig. 3 is the flow chart according to a kind of vehicle retrieval method shown in an exemplary embodiment；
Fig. 4 is a kind of vehicle retrieval method shown in another exemplary embodiment on the basis of Fig. 3 corresponding embodiments
Flow chart；
Fig. 5 is the details flow chart of the step 402 of Fig. 4 corresponding embodiments；
Fig. 6 is a kind of model framework schematic diagram of the carry out vehicle retrieval shown in exemplary embodiment；
Fig. 7 is a kind of vehicle retrieval side shown according to another exemplary embodiment on the basis of Fig. 3 corresponding embodiments
The flow chart of method；
Fig. 8 is a kind of model framework schematic diagram of the carry out vehicle retrieval shown in exemplary embodiment；
Fig. 9 is the details flow chart of the step 350 of Fig. 3 corresponding embodiments；
Figure 10 is adopted as vehicle retrieval method provided by the invention and vehicle to be retrieved is searched from vehicle retrieval database
Effect diagram；
Figure 11 is the block diagram according to a kind of vehicle retrieval device shown in an exemplary embodiment；
Figure 12 is a kind of vehicle retrieval device shown in another exemplary embodiment on the basis of Figure 11 corresponding embodiments
Block diagram；
Figure 13 is the details block diagram of classification branch training module in Figure 12 corresponding embodiments；
Figure 14 is a kind of vehicle retrieval device shown in another exemplary embodiment on the basis of Figure 11 corresponding embodiments
Block diagram；
Figure 15 is the details block diagram of vehicle retrieval model in Figure 11 corresponding embodiments.
Specific embodiment
Here explanation will be performed to exemplary embodiment in detail, example is illustrated in the accompanying drawings.Following description is related to
During attached drawing, unless otherwise indicated, the same numbers in different attached drawings represent the same or similar element.Following exemplary embodiment
Described in embodiment do not represent and the consistent all embodiments of the present invention.On the contrary, they be only with it is such as appended
The example of the consistent device and method of some aspects being described in detail in claims, of the invention.
Fig. 1 is according to the implementation environment schematic diagram according to the present invention shown in an exemplary embodiment.Involved by the present invention
And implementation environment include server 110.Vehicle retrieval database can be configured in server 110, hence for a certain to be checked
Vehicle retrieval method provided by the invention may be used in rope vehicle image, server 110, found out from vehicle retrieval database with
The image of Car image matching to be retrieved.
In a particular application, vehicle retrieval method provided by the invention may be used to find out from magnanimity road monitoring picture
With the picture of Car image matching to be retrieved, so as to carry out the track following of suspected vehicles.
As needed, which will also include providing data, i.e., vehicle image and treated in vehicle retrieval database
Retrieve the data source of vehicle image.Specifically, in this implementation environment, data source can be camera 120 or carry
The mobile terminal 130 of camera function.Server 110, by wireless network connection, is obtained with camera 120 or mobile terminal 130
The vehicle image that camera 120 or mobile terminal 130 acquire, forms vehicle retrieval database.Also, for camera 120 or
Scheme provided by the invention may be used from vehicle retrieval number in the vehicle image to be retrieved that mobile terminal 130 acquires, server 110
According to the vehicle image found out in library with Car image matching to be retrieved.
It should be noted that vehicle retrieval method provided by the invention, be not limited to dispose corresponding place in server 110
Logic is managed, can also be the processing logic being deployed in other machines.For example, having in the middle part of the terminal device of computing capability
Affix one's name to processing logic of vehicle retrieval etc..
Fig. 2 is the block diagram according to a kind of server 110 shown in an exemplary embodiment.The server 110 can be because of configuration
Or performance is different and generate bigger difference, can include at least one central processing unit (central processing
Units, CPU) 222 (for example, at least one processors) and memory 232, at least one storage application program 242 or data
244 storage medium 230 (a for example, at least mass memory unit).Wherein, memory 232 and storage medium 230 can be
Of short duration storage or persistent storage.At least one module (diagram is not shown) can be included by being stored in the program of storage medium 230, often
A module can include operating the series of instructions in server.Further, central processing unit 222 could be provided as with
Storage medium 230 communicates, and the series of instructions operation in storage medium 230 is performed on server 110.Server 110 may be used also
To include at least one power supply 226, at least one wired or wireless network interface 250, at least one input/output interface 258,
And/or at least one operating system 241, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM,
FreeBSDTM etc..Following Fig. 3, Fig. 4, Fig. 5, Fig. 7, described in embodiment illustrated in fig. 9 as the step performed by server 110
Suddenly 110 structure of server shown in Fig. 2 can be based on.
Fig. 3 is the flow chart according to a kind of vehicle retrieval method shown in an exemplary embodiment.The vehicle retrieval method
Vehicle retrieval method shown in Fig. 3 can be performed in the server 110 of implementation environment shown in Fig. 1.As shown in figure 3, the vehicle
Search method may comprise steps of.
In the step 310, vehicle image to be retrieved is obtained.
Wherein, vehicle image to be retrieved is sent to server 110 after being acquired by camera 120, can also deposit in advance
Storage is in 110 local data base of server.Vehicle image to be retrieved refers to the image of a certain vehicle, and the present invention is directed to from a large amount of vehicles
Vehicle image highest with this vehicle similarity is retrieved in image rather than is only limitted to find out same vehicle or the vehicle with color
Image.Offer scheme of the present invention can be used for traffic control department, inquire what suspected vehicles occurred in various scenes from monitoring video
Picture, the driving trace that auxiliary carries out suspected vehicles judge.
In step 320, the feature extraction of the vehicle image to be retrieved is carried out by convolutional neural networks, obtains feature
Vector.
Wherein, convolutional neural networks can be the depth convolutional neural networks mould for removing sort module trained in advance
Type, such as InceptionV3.Wherein, the weight parameter of the depth convolutional neural networks model can be utilized through Imagenet (mesh
The maximum database-name of former world epigraph identification) the trained weight of data set, it subsequently can no longer carry out the part power
The update of weight.As needed, which can also be other trained picture classification networks, such as
RestNet, VGG, DenseNet.
Feature vector refers to by what is generated after vehicle image progress convolution algorithm to be retrieved, including the cable car to be checked
The vector of the characteristic information of image.
Specifically, according to the pixel value of pixel each in vehicle image to be retrieved, trained convolution god in advance is utilized
Weight parameter through network carries out convolution algorithm to the pixel value of each pixel in the vehicle image to be retrieved.Convolution god
It can include input layer, convolutional layer and pond layer through network, for inputting vehicle image to be retrieved, convolutional layer is used to make input layer
Convolutional calculation is carried out to vehicle image to be retrieved with weight parameter, the characteristics of image that pond layer is used to export convolutional layer drops
Dimension.In an embodiment of the present invention, the vehicle image to be retrieved is exported in the last one pond layer of the convolutional neural networks
The feature vector of 2048 dimensions.
In a step 330, depth between attributive character extraction and relatively positive negative sample is carried out respectively to described eigenvector
The extraction of feature.
Wherein, vehicle face of the attributive character including vehicle to be retrieved, vehicle window, vehicle, body color, car plate bottom plate color, year
Feature in inspection mark, pendant and pose etc..In a kind of exemplary embodiment, attributive character can be vehicle feature and vehicle
Body color characteristic.Vehicle can be the types such as car, offroad vehicle, station wagon, motorbus, buggy, truck, tank car.
As an example it is assumed that share 5 possible body colors：Red, silver color, grey, black, white.The vehicle of vehicle image to be retrieved
Body color is silver color, then can extract vehicle body face from the feature vector of 2048 dimensions according to trained mapping ruler in advance
Color characteristic can be [0,1,0,0,0].Assuming that the vehicle station wagon of vehicle image to be retrieved, then can from 2048 dimensional features to
Vehicle is extracted in amount to be characterized as [0,0,1,0,0,0,0].By the way that body color feature and vehicle feature are merged
Obtain attributive character [0,1,0,0,0,0,0,1,0,0,0,0].
Wherein, positive negative sample refers to carry out used positive sample and negative sample during deep learning, and positive sample expression has
The sample vehicle image of similitude, negative sample represent the sample vehicle image of dissimilarity.The depth of vehicle image to be retrieved is special
Sign can embody the similitude between the positive negative sample provided when the vehicle image to be retrieved and deep learning stage.It is if to be checked
Euclidean distance between the depth characteristic of rope vehicle image and the depth characteristic of some sample vehicle image is very close, and representing should
Vehicle image to be retrieved is similar to the sample vehicle image, thus the depth characteristic of vehicle image to be retrieved may be considered it is to be checked
Similarity feature when rope vehicle image and deep learning between sample vehicle image.
Specifically, trained parameter when the extraction of depth characteristic can be according to positive negative sample progress deep learning, right
Feature vector carries out convolution algorithm, extracts the depth characteristic of vehicle image to be retrieved.In a kind of exemplary embodiment, rolling up
The last one pond layer of product neural network is followed by three layers of full articulamentum, and the feature vector tieed up to 2048 is successively according to trained
The network parameter and bias progress convolutional calculation of every layer of full articulamentum, the depth characteristic that final output 256 is tieed up.
In step 340, merge attributive character and depth characteristic, obtain the vehicle vision feature of vehicle image to be retrieved.
Wherein, it can splice attributive character and depth characteristic to merge attributive character and depth characteristic.For example, belong to
Property be characterized as [1,0,0,0], depth characteristic be [1,2,1,2], it can be formed that attributive character and depth characteristic, which are spliced,
[1,0,0,0,1,2,1,2] or [1,2,1,2,1,0,0,0].As needed, merging attributive character and depth characteristic can be with
Be between row vector addition, subtract each other.
Vehicle vision feature refers to the feature generated after merging attributive character and depth characteristic, in a kind of exemplary implementation
In example, the attributive character of 1024 dimensions is extracted from 2048 dimensional feature vectors by attribute forecast branch model, by plus or minus about
Beam branch model extracts 256 dimension depth characteristics from 2048 dimensional feature vectors, after attributive character and depth characteristic are merged
To the vehicle vision feature of 1280 dimensions.By the vehicle vision feature of image in the vehicle vision feature and vehicle retrieval database into
Row matching primitives.Compared with the prior art, which contains the more features information of vehicle image to be retrieved, therefore
Discrimination between the vehicle vision feature of other vehicle images is stronger, and then by comparing vehicle vision feature, so that it may accurate
Really distinguish different vehicle images.
It should be noted that it is based purely on the body color feature of vehicle image to be retrieved or vehicle feature progress target figure
The lookup of picture is typically only capable to find out the vehicle pictures identical with vehicle image body color to be retrieved or find out and cable car to be checked
Identical vehicle pictures of image vehicle, and can not accurately find out vehicle image of the designated vehicle under different scenes.Citing comes
It says, it is assumed that the body color of vehicle to be retrieved is black, and vehicle is car, then is carried out by vehicle feature and body color feature
Vehicle image is searched, and can find out the car image of all black, it is seen that the vehicle image data volume found out is still very big, and can not
Accurately find out image of certain suspected vehicles under other scenes.
In step 350, vehicle vision between image and the vehicle image to be retrieved in vehicle retrieval database is carried out
The matching primitives of feature obtain the target image with the Car image matching to be retrieved.
Wherein, it is stored with great amount of images in advance in vehicle retrieval database, which can be road monitoring camera 120
The monitored picture of acquisition, road monitoring camera can shoot the vehicle pictures for largely passing through the section, and store to server
In 110 vehicle retrieval database.
Matching primitives refer to every image in the vehicle vision feature of vehicle image to be retrieved and vehicle retrieval database
Vehicle vision feature carry out similitude comparison.And then vehicle with vehicle image to be retrieved is found out from vehicle retrieval database
The highest image of similitude between visual signature.For example, it can be regarded by calculating the vehicle of vehicle image to be retrieved
Feel the Euclidean distance between the vehicle vision feature of every image in feature and vehicle retrieval database, determine vehicle figure to be retrieved
Picture and the similarity between image in vehicle retrieval database.
Target image refers to the image closest with vehicle image to be retrieved found out from vehicle retrieval database.
Specifically, the corresponding vehicle vision feature of every image in vehicle retrieval database can be calculated in advance and is stored
In the vehicle retrieval database.When needing retrieval, by the vehicle vision feature of vehicle image to be retrieved and vehicle retrieval number
It compares, therefrom finds out highest with vehicle image similitude to be retrieved according to the vehicle vision feature of every image in library and similitude
Image is as target image.
The technical solution that the above exemplary embodiments of the present invention provide, by the depth characteristic for extracting vehicle image to be retrieved
And attributive character, and searched from vehicle retrieval database based on the vehicle vision feature comprising the depth characteristic and attributive character
With the target image of the Car image matching to be retrieved.Compared with the prior art, which contains vehicle to be retrieved
The more characteristic informations of image, therefore the discrimination between other vehicle images is stronger, it is particularly possible to distinguish same alike result spy
The different vehicle image of sign, based on the vehicle vision feature search with the approximate target image of vehicle image to be retrieved, accurately
Property higher, due to without manually participating in inquiry, improving the efficiency of vehicle retrieval.
In a kind of exemplary embodiment, above-mentioned steps 330 specifically include：
The class prediction branch built by sample vehicle image and sample vehicle image institute flag attribute classification
Model carries out described eigenvector the attribute class prediction of the vehicle image to be retrieved, obtains attributive character.
Wherein, sample vehicle image refer to for carry out model training known attribute classification vehicle image.Sample vehicle
It can also be other vehicle images that image, which can be the vehicle image stored in vehicle retrieval database,.Sample vehicle image
It is stored in server 110 after can also being acquired by road monitoring camera 120.Attribute classification includes vehicle in sample vehicle image
Body color classification (one kind in such as red, silver color, black), vehicle classification (one in such as car, lorry, car
Kind) or other classifications.
Class prediction branch model is the network model to continue after convolutional neural networks in step 320.It specifically can root
According to multiple sample vehicle images and its corresponding attribute classification, the neural network learning of attributive character extraction is carried out, generates classification
Predicted branches model.After the feature vector input category predicted branches model of vehicle image to be retrieved, it is to be retrieved to export this
The attributive character of vehicle image.
In a kind of exemplary embodiment, above-mentioned steps 330 can also include：
The positive and negative constraint branch model built by sample vehicle image and corresponding positive negative sample, to the feature to
Amount carries out depth characteristic extraction.
Wherein, positive negative sample includes positive sample and negative sample, and positive sample refers to include same vehicle with sample vehicle image
Vehicle image, same vehicle refer to specific a certain vehicle rather than refer to body color or the identical a kind of vehicle of vehicle.Negative sample
Originally refer to the vehicle image for including different vehicle with sample vehicle image.Such as it is included with sample vehicle image with money vehicle system but face
The different vehicle image of color includes same color but the different vehicle image of vehicle, other vehicles and face from sample vehicle image
Vehicle pictures of color etc..It should be noted that the labeled other positive sample of Attribute class and negative sample can also be used as sample vehicle
Image carries out the training of class prediction branch model.
Positive and negative constraint branch model is the network model to continue after convolutional neural networks in step 320, positive and negative constraint point
Branch model, which belongs to class prediction branch model after convolutional neural networks, carries out multi-task learning, the Liang Ge branches built.Just
Beam branch model of breaking a promise can parallel be built with class prediction branch model, can also class prediction branch model structure after,
Build positive and negative constraint branch model again, and can based on the positive negative sample newly increased to the parameter of class prediction branch model into one
Step optimizes.
It it should be noted that can be according to related between sample vehicle image and positive sample, sample vehicle image and negative sample
It is uncorrelated between this, study is ranked up to sample vehicle image, positive sample and negative sample, builds positive and negative constraint branch model.
The feature vector of vehicle image to be retrieved is inputted into the positive and negative depth spy for constraining branch model and exporting the vehicle image to be retrieved
Sign.The depth characteristic is embodied with participating in the similarity relationships between trained sample during positive and negative constraint branch model structure.It is false
If depth characteristic is 256 dimensions, then the depth characteristic of each sample vehicle image can map in the feature space of 256 dimensions
To a point, the depth characteristic spy of vehicle image to be retrieved can also map out corresponding point, therefore special in this feature space
The distance between depth characteristic corresponding points can embody the similitude between corresponding vehicle image in sign space.
In a kind of exemplary embodiment, as shown in figure 4, before step 330, method provided by the invention is also possible to wrap
Include following steps：
In step 401, the attribute classification that sample vehicle image and the sample vehicle image are marked is obtained.
Sample vehicle image refer to for carry out class prediction branch model training known attribute classification vehicle image.
The attribute classification that sample vehicle image is marked can be obtained from the label information entrained by sample vehicle image.
In step 402, neural network learning is carried out by the sample vehicle image and institute's flag attribute classification, obtained
It must carry out the class prediction branch model of the attributive character extraction.
Wherein, neural network learning is carried out to refer to according to sample vehicle image and its corresponding attribute classification, by building
The study attributive character extraction of neural network framework, difference is minimum between making the attributive character and its attribute classification that study is arrived, at this time
The network parameter of neural network framework according to building can obtain class prediction branch model.It should be noted that the nerve
The network architecture includes trunk convolutional neural networks (input layer, convolutional layer and pond layer) and the full articulamentum of multilayer.After optimization
The full articulamentum of multilayer be class prediction branch model.
In a kind of exemplary embodiment, as shown in figure 5, above-mentioned steps 402 specifically include：
In step 501, the sampling feature vectors of the sample vehicle image are extracted by the convolutional neural networks.
Wherein, which refers to remove the depth convolutional neural networks model of sort module.I.e. it is above-mentioned at least
Trunk convolutional neural networks including input layer, convolutional layer and pond layer.The network parameter of the convolutional neural networks can shift to an earlier date
It trains to obtain using Imagenet data sets.Characteristic vector pickup is being carried out or to sample vehicle image to vehicle image to be retrieved
Can be used during sampling feature vectors extraction the parameter of the trained convolutional neural networks.
The sampling feature vectors of sample vehicle image are similar with the feature vector of vehicle image to be retrieved, comprising corresponding vehicle
The characteristic information of image.The sample attribute feature and sample of sample vehicle image can also be extracted from sampling feature vectors
Depth characteristic.
In step 502, convolutional calculation is carried out to the sampling feature vectors by building multilayer full articulamentum, extracts institute
State the corresponding sample attribute feature of sampling feature vectors.
As shown in fig. 6, the sample vehicle image to single input extracts feature vector by above-mentioned convolutional neural networks,
The sampling feature vectors of last layer of 2048 dimension of pond layer output of convolutional neural networks.In the follow-up access two of this feature vector
Branch is class prediction branch model and positive and negative constraint branch model respectively.
Class prediction branch model can include multiple full articulamentums, as shown in fig. 6, after 2048 dimensional feature vectors connect entirely
Continue 1024 neurons, be then divided into Liang Ge branches, vehicle predicted branches 601 and body color predicted branches 602.In vehicle
601,1024 neurons of predicted branches connect follow-up n neuron entirely, and are classified using softmax active coatings.N is represented
Total vehicle class number.Such as car, offroad vehicle, station wagon, motorbus, buggy, truck, tank car totally 7 type
Type.In 602,1024 neurons of body color predicted branches follow-up m neuron is connected entirely, and use softmax active coatings
Classify, m represents body color classification number.Such as red, grey, silver color, black, white, totally 5 kinds of body color classification
Number.
Convolutional calculation is carried out by the network parameter of the full articulamentum of multilayer to the sampling feature vectors of sample vehicle image, is carried
This attributive character is sampled, which can include vehicle characteristics and the vehicle body face that vehicle predicted branches 601 export
The body color feature that color predicted branches 602 export.
In step 503, the attribute classification marked according to the sample vehicle image optimizes the power of the full articulamentum
Weight parameter makes the difference between the sample attribute feature and the attribute classification marked minimum, the full connection after optimization
Layer forms the class prediction branch model.
Specifically, can be according to the attribute classification that sample vehicle image is marked, such as body color classification can be included
With vehicle classification, the weight parameter of the full articulamentum of above-mentioned multilayer is adjusted, the sample attribute feature so as to make finally to export is with being marked
Difference between the attribute classification of note is minimum.The full articulamentum of multilayer and softmax active coatings after adjustment form classification together
Predicted branches model.
Wherein, the loss function of class prediction branch model can be cross entropy,C is classification number
Mesh, such as above-mentioned 7 kinds of vehicle class numbers, body color class number in 5,To predict that all samples belong to a certain classification k
Probability, ykBelong to the probability of classification k for samples all under truth, by reducing between prediction result and legitimate reading
Error is trained, i.e., makes J as small as possible in training.
In a kind of exemplary embodiment, as shown in fig. 7, before above-mentioned steps 330, method provided by the invention may be used also
It can include the following steps.
In step 701, sample vehicle image and the corresponding positive negative sample of the sample vehicle image are obtained, forms sample
This triple；
Wherein, for positive negative sample is relatively a certain sample vehicle image, same is included with the sample vehicle image
Vehicle is known as positive sample, and the referred to as negative sample of different vehicle is included with sample vehicle image.It is emphasized that the present invention is signified
Same vehicle refer to specific a certain vehicle rather than refer to body color or the identical a kind of vehicle of vehicle.Positive negative sample is one
A relative concept, when building attribute forecast branch model, the positive negative sample of marked attributive classification can also serve as sample
Vehicle image carries out the training of attribute forecast branch model.
It should be noted that in order to ensure to have enough positive negative samples, the method for sampling can first take a vehicle figure
Piece (anchor) then ensures there is at least two color, the corresponding same money vehicle systems of anchor in the same a vehicle system belonging to anchor
2 are no less than with the vehicle fleet size of same color, obtains negative sample as much as possible.And positive sample can be vehicle in anchor
In the picture taken by other scenes.Positive sample, negative sample and sample vehicle image form sample triple.To sample three
Can be 299 pixels by all image scalings to suitable size, such as width before tuple carries out the convolution algorithm of convolutional neural networks,
A height of 299 pixel, the scaling size of image can be determined according to follow-up convolutional neural networks model.
In a step 702, the sample triple is carried out using the loss function for being adapted to matching primitives relatively positive and negative
The neural network learning of sample obtains the positive and negative constraint branch model for carrying out the depth characteristic extraction.
Wherein, the loss function for being adapted to matching primitives refers to that loss function and matching primitives are all based on similar image
The approximate principle of feature.For example, loss function can be the depth of the depth characteristic and positive sample that calculate sample vehicle image
The distance between feature is spent, makes this apart from as small as possible, the depth characteristic of calculating sample vehicle image and the depth spy of negative sample
The distance between sign, makes the distance as big as possible.And matching primitives can be calculate vehicle retrieval database in image with it is to be checked
The distance between rope vehicle image vehicle vision feature finds the minimum image of distance as with treating in vehicle retrieval database
Retrieve the highest target image of vehicle image similarity.At this time it is considered that damaging out function is adapted to matching primitives.
Specifically, as shown in figure 8, to sample vehicle image 801, positive sample 802 and the negative sample in sample triple
803 carry out convolution algorithm by convolutional neural networks, the corresponding sampling feature vectors of each sample (2048 dimension) are exported, in convolution
Neural network subsequently accesses three layers of full articulamentum, carries out convolution algorithm to sampling feature vectors by three layers of full articulamentum, successively
Obtain the feature vector of 1024 dimensions, 1024 dimensions, 256 dimensions.By adjusting the weight parameter of the full articulamentum of multilayer, make sample vehicle figure
Difference between 256 dimensional feature vectors belonging to 256 dimensional feature vectors and positive sample 802 as belonging to 801 is as small as possible, makes sample vehicle
Difference between 256 dimensional feature vectors belonging to 256 dimensional feature vectors belonging to image 801 and negative sample 903 is as big as possible.Pass through
The full articulamentum of multilayer after optimization can obtain the positive and negative constraint branch model.256 dimensional feature vectors of final output are pair
Answer the depth characteristic of input sample.
In a kind of exemplary embodiment, above-mentioned steps 701 specifically include：
The neural network learning of the depth characteristic extraction is carried out to the sample triple, by optimizing sample vehicle figure
Picture and the characteristic distance between positive negative sample, obtain the positive and negative constraint branch model for carrying out the depth characteristic extraction.
Wherein, characteristic distance refers to the depth characteristic that sample vehicle image learns and the depth characteristic that positive sample learns
Between the depth characteristic that learns of the depth characteristic that learns of the first distance and sample vehicle image and negative sample between
Second distance, when being optimized to neural network, needing to make this, second distance may be big, ability first apart from as small as possible
Meet positive sample and sample vehicle image similarity is higher, negative sample and the relatively low characteristic of sample vehicle image similarity.It is a kind of
In exemplary embodiment, characteristic distance can be that Euclidean distance can also be COS distance.
In a kind of exemplary embodiment, the loss function of positive and negative constraint branch model can beWhereinFor sample vehicle image,For positive and negative constraint point
The depth characteristic of sample vehicle image that branch model finally exports,For the depth characteristic of positive sample,For negative sample
Depth characteristic；For sample vehicle image and Euclidean distance of the positive sample in the depth characteristic acquired,
For threshold value, which ensures that the characteristic distance of sample vehicle image and positive sample is less than sample vehicle image and negative sample
This characteristic distance, i.e. positive sample are adjusted the distance as far as possible closely, and negative sample is adjusted the distance as far as possible, to reach the mesh of sequence study
's.The present invention is by adjusting sample triple is carried out the network parameter of the neural network learning of depth characteristic extraction, specifically
By adjusting the weight parameter of the full articulamentum of multilayer that positive and negative constraint branch model is built, above-mentioned loss function is made to reach minimum
Value, obtains the positive and negative constraint branch model being made of the full articulamentum of multilayer after optimizing.
In a kind of exemplary embodiment, as shown in figure 9, above-mentioned steps 350 specifically include：
In step 351, image in vehicle retrieval database and the vehicle image to be retrieved are subjected to vehicle vision spy
The comparison of sign.
Specifically, the vehicle for being referred to image in above-mentioned steps 320-340 extractions vehicle retrieval database of the present invention regards
Feel feature.Then by the vehicle vision feature of image in vehicle retrieval database and the vehicle vision feature of vehicle image to be retrieved
It is compared.The mode of comparison can be by the Euclidean distance between calculating comparison vehicle vision feature, apart from smaller expression
Similarity is higher, closer with vehicle image to be retrieved.
In step 352, the vehicle vision with the vehicle image to be retrieved is searched in the vehicle retrieval database
The highest image of characteristic similarity obtains the target image with the Car image matching to be retrieved.
Specifically, vehicle vision is special between image and vehicle image to be retrieved in vehicle retrieval database by comparing one by one
The distance of sign filters out image in the minimum vehicle retrieval database of distance, which is exactly similar to vehicle image to be retrieved
Highest image is spent, is exactly the target image with Car image matching to be retrieved.
Figure 10 is adopted as vehicle retrieval method provided by the invention, and cable car to be checked is searched from vehicle retrieval database 1002
1001 effect diagram.As shown in Figure 10, vehicle image 1001 to be retrieved is found out from vehicle retrieval database 1002 to include
The picture of same vehicle, labeled as " √ ".Conversely, labeled as "”.Server 110 is examined by using vehicle provided by the invention
Suo Fangfa can find out designated vehicle from the magnanimity monitored picture that camera 120 provides, and realize to scheme to search vehicle, and then carry out
The track following of designated vehicle reduces the workload of traffic control personnel, improves the efficiency that suspected vehicles are searched.
To sum up, vehicle retrieval method provided by the invention, retrieval accuracy is high, and traffic control personnel's quick search can be assisted to obtain
To designated vehicle in the picture of other scenes, user need to only input a vehicle pictures to be retrieved, it is possible to from vehicle retrieval number
According to approximate vehicle is found out in library, recall precision is high, reduces human cost.It, can be in addition, by follow-up ever-increasing data
Further model is optimized, further improves vehicle retrieval accuracy.
Following is embodiment of the present disclosure, can be used for performing the vehicle retrieval that the above-mentioned server 110 of the disclosure performs
Embodiment of the method.For the details not disclosed in embodiment of the present disclosure, disclosure vehicle retrieval embodiment of the method is please referred to.
Figure 11 is according to a kind of block diagram of vehicle retrieval device shown in an exemplary embodiment, which can
For in the server 110 of implementation environment shown in Fig. 1, performing any shown vehicle retrieval of Fig. 3, Fig. 4, Fig. 5, Fig. 7, Fig. 9
The all or part of step of method.As shown in figure 11, which includes but not limited to：Image collection module
1110th, vector obtains module 1120, characteristic extracting module 1130, feature merging module 1140 and vehicle retrieval module 1150.
Image collection module 1110, for obtaining vehicle image to be retrieved；
Vector obtains module 1120, and the feature that the vehicle image to be retrieved is carried out for passing through convolutional neural networks carries
It takes, obtains feature vector；
Characteristic extracting module 1130, for carrying out attributive character extraction and relatively positive negative sample respectively to described eigenvector
Between depth characteristic extraction；
Feature merging module 1140 for merging the attributive character and depth characteristic, obtains the vehicle figure to be retrieved
The vehicle vision feature of picture；
Vehicle retrieval module 1150, for carrying out in vehicle retrieval database between image and the vehicle image to be retrieved
The matching primitives of vehicle vision feature obtain the target image with the Car image matching to be retrieved.
The function of modules and the realization process of effect specifically refer to right in above-mentioned vehicle retrieval method in above device
The realization process of step is answered, details are not described herein.
Image collection module 1110 such as can be some physical arrangement wire/radio network interface 250 in Fig. 2.
Vector obtains module 1120, characteristic extracting module 1130, feature merging module 1140 and vehicle retrieval module
1150 can also be function module, for performing the correspondence step in above-mentioned vehicle retrieval method.It is appreciated that these modules can
With by hardware, software, or a combination of both realize.When realizing in hardware, these modules may be embodied as one or
Multiple hardware modules, such as one or more application-specific integrated circuits.When being realized with software mode, these modules may be embodied as
The one or more computer programs performed on the one or more processors, such as performed by the central processing unit 222 of Fig. 2
The program being stored in memory 232.
In a kind of exemplary embodiment, the characteristic extracting module 1130 includes：
Attributive character extraction unit, for passing through sample vehicle image and sample vehicle image institute flag attribute class
The class prediction branch model not built carries out described eigenvector the attribute class prediction of the vehicle image to be retrieved,
Obtain attributive character.
In a kind of exemplary embodiment, the characteristic extracting module 1130 includes：
Depth characteristic extraction unit, for passing through the positive and negative constraint of sample vehicle image and corresponding positive negative sample structure
Branch model carries out depth characteristic extraction to described eigenvector.
In a kind of exemplary embodiment, as shown in figure 12, described device further includes：
Sample image acquisition module 1210, for obtaining sample vehicle image and the sample vehicle image marked
Attribute classification；
Classification branch training module 1220, for carrying out god by the sample vehicle image and institute's flag attribute classification
Through e-learning, the class prediction branch model for carrying out the attributive character extraction is obtained.
In a kind of exemplary embodiment, as shown in figure 13, classification branch training module 1220 includes：
Sample vector extraction unit 1221, for passing through the sample that the convolutional neural networks extract the sample vehicle image
Eigen vector；
Sample characteristics extraction unit 1222, for being rolled up by building the full articulamentum of multilayer to the sampling feature vectors
Product calculates, and extracts the corresponding sample attribute feature of the sampling feature vectors；
Parameter optimization unit 1223 for the attribute classification marked according to the sample vehicle image, optimizes described complete
The weight parameter of articulamentum makes the difference between the sample attribute feature and the attribute classification marked minimum, after optimization
The full articulamentum forms the class prediction branch model.
In a kind of exemplary embodiment, as shown in figure 14, described device further includes：
Triple acquisition module 1410, it is corresponding positive and negative for obtaining sample vehicle image and the sample vehicle image
Sample forms sample triple；
Positive and negative bound branch training module 1420, for using the loss for being adapted to matching primitives to the sample triple
Function carries out the neural network learning of relatively positive negative sample, obtains the positive and negative bound branch mould for carrying out the depth characteristic extraction
Type.
In a kind of exemplary embodiment, the positive and negative bound branch training module 1420 includes：
Characteristic distance optimizes unit, for carrying out the Neural Network Science of the depth characteristic extraction to the sample triple
It practises, by optimizing the characteristic distance between sample vehicle image and positive negative sample, obtains and carrying out the depth characteristic extraction just
It breaks a promise beam branch model.
In a kind of exemplary embodiment, as shown in figure 15, vehicle retrieval module 1150 includes：
Characteristic Contrast unit 1151, for by image in vehicle retrieval database and the vehicle image to be retrieved into driving
The comparison of visual signature；
Image searching unit 1152, for being searched in the vehicle retrieval database and the vehicle image to be retrieved
The highest image of vehicle vision characteristic similarity obtains the target image with the Car image matching to be retrieved.
Optionally, the disclosure also provides a kind of electronic equipment, which can be used for the clothes of implementation environment shown in Fig. 1
It is engaged in device 110, performs all or part of step of any shown vehicle retrieval method of Fig. 3, Fig. 4, Fig. 5, Fig. 7, Fig. 9.Institute
Electronic equipment is stated to include：
Processor；
For storing the memory of processor-executable instruction；
Wherein, the processor is configured as performing the vehicle retrieval method described in the above exemplary embodiments.
The processor of electronic equipment performs the concrete mode of operation in the related vehicle retrieval method in the embodiment
Embodiment in perform detailed description, explanation will be not set forth in detail herein.
In the exemplary embodiment, a kind of storage medium is additionally provided, which is computer readable storage medium,
Such as can be the provisional and non-transitorycomputer readable storage medium for including instruction.The storage medium is stored with computer
Program, the computer program can be performed to complete above-mentioned vehicle retrieval method by the central processing unit 222 of server 110.
It should be understood that the invention is not limited in the precision architecture for being described above and being shown in the drawings, and
And various modifications and change can be being performed without departing from the scope.The scope of the present invention is only limited by appended claim.
Claims (15)
- A kind of 1. vehicle retrieval method, which is characterized in that the method includes：Obtain vehicle image to be retrieved；The feature extraction of the vehicle image to be retrieved is carried out by convolutional neural networks, obtains feature vector；Carry out the extraction of depth characteristic between attributive character extraction and relatively positive negative sample respectively to described eigenvector；Merge the attributive character and depth characteristic, obtain the vehicle vision feature of the vehicle image to be retrieved；The matching primitives of vehicle vision feature between image and the vehicle image to be retrieved in vehicle retrieval database are carried out, are obtained Obtain the target image with the Car image matching to be retrieved.
- 2. according to the method described in claim 1, it is characterized in that, the attributive character that carried out respectively to described eigenvector carries The extraction of the depth characteristic between relatively positive negative sample is taken, including：The class prediction branch model built by sample vehicle image and sample vehicle image institute flag attribute classification, The attribute class prediction of the vehicle image to be retrieved is carried out to described eigenvector, obtains attributive character.
- 3. according to the method described in claim 1, it is characterized in that, the attributive character that carried out respectively to described eigenvector carries The extraction of the depth characteristic between relatively positive negative sample is taken, including：The positive and negative constraint branch model built by sample vehicle image and corresponding positive negative sample, to described eigenvector into Row depth characteristic is extracted.
- 4. according to the method described in claim 1, it is characterized in that, the attributive character that carried out respectively to described eigenvector carries It takes between relatively positive negative sample before the extraction of depth characteristic, the method further includes：Obtain the attribute classification that sample vehicle image and the sample vehicle image are marked；Neural network learning is carried out by the sample vehicle image and institute's flag attribute classification, obtains and carries out the attribute spy Levy the class prediction branch model of extraction.
- 5. according to the method described in claim 4, described carried out by the sample vehicle image and institute's flag attribute classification Neural network learning obtains the class prediction branch model for carrying out the attributive character extraction, including：The sampling feature vectors of the sample vehicle image are extracted by the convolutional neural networks；Convolutional calculation is carried out to the sampling feature vectors by building multilayer full articulamentum, extracts the sampling feature vectors pair The sample attribute feature answered；According to the attribute classification that the sample vehicle image is marked, optimize the weight parameter of the full articulamentum, make the sample Difference between this attributive character and the attribute classification marked is minimum, and it is pre- that the full articulamentum after optimization forms the classification Survey branch model.
- 6. according to the method described in claim 1, it is characterized in that, the attributive character that carried out respectively to described eigenvector carries It takes between relatively positive negative sample before the extraction of depth characteristic, the method further includes：Sample vehicle image and the corresponding positive negative sample of the sample vehicle image are obtained, forms sample triple；The Neural Network Science of relatively positive negative sample is carried out using the loss function for being adapted to matching primitives to the sample triple It practises, obtains the positive and negative constraint branch model for carrying out the depth characteristic extraction.
- 7. according to the method described in claim 6, it is characterized in that, described count the sample triple using matching is adapted to The loss function of calculation carries out the neural network learning of relatively positive negative sample, obtains the positive and negative constraint for carrying out the depth characteristic extraction Branch model, including：The neural network learning of the depth characteristic extraction is carried out to the sample triple, by optimize sample vehicle image with Characteristic distance between positive negative sample, obtains the positive and negative constraint branch model for carrying out the depth characteristic extraction.
- 8. according to the method described in claim 1, it is characterized in that, image is treated with described in the progress vehicle retrieval database The matching primitives of vehicle vision feature, obtain the target figure with the Car image matching to be retrieved between retrieval vehicle image Picture, including：Image in vehicle retrieval database is carried out to the comparison of vehicle vision feature with the vehicle image to be retrieved；It is searched in the vehicle retrieval database highest with the vehicle vision characteristic similarity of the vehicle image to be retrieved Image obtains the target image with the Car image matching to be retrieved.
- 9. a kind of vehicle retrieval device, which is characterized in that described device includes：Image collection module, for obtaining vehicle image to be retrieved；Vector obtains module, for passing through the feature extraction that convolutional neural networks carry out the vehicle image to be retrieved, obtains special Sign vector；Characteristic extracting module, for carrying out depth between attributive character extraction and relatively positive negative sample respectively to described eigenvector The extraction of feature；Feature merging module for merging the attributive character and depth characteristic, obtains the vehicle of the vehicle image to be retrieved Visual signature；Vehicle retrieval module, for carrying out in vehicle retrieval database vehicle vision between image and the vehicle image to be retrieved The matching primitives of feature obtain the target image with the Car image matching to be retrieved.
- 10. device according to claim 9, which is characterized in that the characteristic extracting module includes：Attributive character extraction unit, for passing through sample vehicle image and the sample vehicle image institute flag attribute classification structure The class prediction branch model built carries out described eigenvector the attribute class prediction of the vehicle image to be retrieved, obtains Attributive character.
- 11. device according to claim 9, which is characterized in that the characteristic extracting module includes：Depth characteristic extraction unit, for passing through the positive and negative bound branch of sample vehicle image and corresponding positive negative sample structure Model carries out depth characteristic extraction to described eigenvector.
- 12. device according to claim 9, which is characterized in that described device further includes：Sample image acquisition module, for obtaining the Attribute class that sample vehicle image and the sample vehicle image are marked Not；Classification branch training module, for carrying out Neural Network Science by the sample vehicle image and institute's flag attribute classification It practises, obtains the class prediction branch model for carrying out the attributive character extraction.
- 13. device according to claim 12, classification branch training module includes：Sample vector extraction unit, for pass through the convolutional neural networks extract the sample characteristics of the sample vehicle image to Amount；Sample characteristics extraction unit, for carrying out convolutional calculation to the sampling feature vectors by building the full articulamentum of multilayer, Extract the corresponding sample attribute feature of the sampling feature vectors；Parameter optimization unit for the attribute classification marked according to the sample vehicle image, optimizes the full articulamentum Weight parameter makes difference between the sample attribute feature and the attribute classification marked minimum, and described after optimization connects entirely It connects layer and forms the class prediction branch model.
- 14. a kind of electronic equipment, which is characterized in that the electronic equipment includes：Processor；For storing the memory of processor-executable instruction；Wherein, the processor is configured as the vehicle retrieval method described in perform claim requirement 1-8 any one.
- 15. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has computer journey Sequence, the computer program can be performed the vehicle retrieval method described in completing claim 1-8 any one as processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810119550.3A CN108197326B (en) | 2018-02-06 | 2018-02-06 | Vehicle retrieval method and device, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810119550.3A CN108197326B (en) | 2018-02-06 | 2018-02-06 | Vehicle retrieval method and device, electronic equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108197326A true CN108197326A (en) | 2018-06-22 |
| CN108197326B CN108197326B (en) | 2023-05-26 |
Family
ID=62593045
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810119550.3A Active CN108197326B (en) | 2018-02-06 | 2018-02-06 | Vehicle retrieval method and device, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108197326B (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108984644A (en) * | 2018-06-25 | 2018-12-11 | 高新兴科技集团股份有限公司 | A kind of fake-licensed car search method and system |
| CN109035779A (en) * | 2018-08-30 | 2018-12-18 | 南京邮电大学 | Freeway traffic flow prediction technique based on DenseNet |
| CN109165589A (en) * | 2018-08-14 | 2019-01-08 | 北京颂泽科技有限公司 | Vehicle based on deep learning recognition methods and device again |
| CN109325420A (en) * | 2018-08-27 | 2019-02-12 | 广州烽火众智数字技术有限公司 | A kind of vehicle characteristics search method and system based on big data |
| CN109558823A (en) * | 2018-11-22 | 2019-04-02 | 北京市首都公路发展集团有限公司 | A kind of vehicle identification method and system to scheme to search figure |
| CN109711443A (en) * | 2018-12-14 | 2019-05-03 | 平安城市建设科技(深圳)有限公司 | Floor plan recognition methods, device, equipment and storage medium neural network based |
| CN109902732A (en) * | 2019-02-22 | 2019-06-18 | 哈尔滨工业大学（深圳） | Automobile automatic recognition method and relevant apparatus |
| CN110378237A (en) * | 2019-06-21 | 2019-10-25 | 浙江工商大学 | Human facial expression recognition method based on depth measure converged network |
| CN110766077A (en) * | 2019-10-24 | 2020-02-07 | 上海眼控科技股份有限公司 | Method, device and equipment for screening sketch in evidence chain image |
| CN110837572A (en) * | 2019-11-15 | 2020-02-25 | 北京推想科技有限公司 | Image retrieval method and device, readable storage medium and electronic equipment |
| CN111008294A (en) * | 2018-10-08 | 2020-04-14 | 阿里巴巴集团控股有限公司 | Traffic image processing and image retrieval method and device |
| CN111062400A (en) * | 2018-10-16 | 2020-04-24 | 浙江宇视科技有限公司 | Target matching method and device |
| CN111310844A (en) * | 2020-02-26 | 2020-06-19 | 广州华工邦元信息技术有限公司 | Vehicle identification model construction method and device and identification method and device |
| CN111611414A (en) * | 2019-02-22 | 2020-09-01 | 杭州海康威视数字技术股份有限公司 | Vehicle retrieval method, device and storage medium |
| CN112085034A (en) * | 2020-09-11 | 2020-12-15 | 北京埃福瑞科技有限公司 | Rail transit train positioning method and system based on machine vision |
| CN111062400B (en) * | 2018-10-16 | 2024-04-30 | 浙江宇视科技有限公司 | Target matching method and device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105354273A (en) * | 2015-10-29 | 2016-02-24 | 浙江高速信息工程技术有限公司 | Method for fast retrieving high-similarity image of highway fee evasion vehicle |
| CN106649688A (en) * | 2016-12-16 | 2017-05-10 | 深圳市华尊科技股份有限公司 | Image retrieval method and terminal |
| CN106776943A (en) * | 2016-12-01 | 2017-05-31 | 中科唯实科技（北京）有限公司 | A kind of vehicle retrieval method based on AutoEncoder and attribute tags |
-
2018
- 2018-02-06 CN CN201810119550.3A patent/CN108197326B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105354273A (en) * | 2015-10-29 | 2016-02-24 | 浙江高速信息工程技术有限公司 | Method for fast retrieving high-similarity image of highway fee evasion vehicle |
| CN106776943A (en) * | 2016-12-01 | 2017-05-31 | 中科唯实科技（北京）有限公司 | A kind of vehicle retrieval method based on AutoEncoder and attribute tags |
| CN106649688A (en) * | 2016-12-16 | 2017-05-10 | 深圳市华尊科技股份有限公司 | Image retrieval method and terminal |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108984644A (en) * | 2018-06-25 | 2018-12-11 | 高新兴科技集团股份有限公司 | A kind of fake-licensed car search method and system |
| CN108984644B (en) * | 2018-06-25 | 2021-04-23 | 高新兴科技集团股份有限公司 | Fake-licensed vehicle retrieval method and system |
| CN109165589A (en) * | 2018-08-14 | 2019-01-08 | 北京颂泽科技有限公司 | Vehicle based on deep learning recognition methods and device again |
| CN109325420A (en) * | 2018-08-27 | 2019-02-12 | 广州烽火众智数字技术有限公司 | A kind of vehicle characteristics search method and system based on big data |
| CN109035779A (en) * | 2018-08-30 | 2018-12-18 | 南京邮电大学 | Freeway traffic flow prediction technique based on DenseNet |
| CN109035779B (en) * | 2018-08-30 | 2021-01-19 | 南京邮电大学 | DenseNet-based expressway traffic flow prediction method |
| CN111008294A (en) * | 2018-10-08 | 2020-04-14 | 阿里巴巴集团控股有限公司 | Traffic image processing and image retrieval method and device |
| CN111008294B (en) * | 2018-10-08 | 2023-06-20 | 阿里巴巴集团控股有限公司 | Traffic image processing and image retrieval method and device |
| CN111062400B (en) * | 2018-10-16 | 2024-04-30 | 浙江宇视科技有限公司 | Target matching method and device |
| CN111062400A (en) * | 2018-10-16 | 2020-04-24 | 浙江宇视科技有限公司 | Target matching method and device |
| CN109558823A (en) * | 2018-11-22 | 2019-04-02 | 北京市首都公路发展集团有限公司 | A kind of vehicle identification method and system to scheme to search figure |
| CN109711443A (en) * | 2018-12-14 | 2019-05-03 | 平安城市建设科技(深圳)有限公司 | Floor plan recognition methods, device, equipment and storage medium neural network based |
| CN109902732A (en) * | 2019-02-22 | 2019-06-18 | 哈尔滨工业大学（深圳） | Automobile automatic recognition method and relevant apparatus |
| CN111611414A (en) * | 2019-02-22 | 2020-09-01 | 杭州海康威视数字技术股份有限公司 | Vehicle retrieval method, device and storage medium |
| CN109902732B (en) * | 2019-02-22 | 2021-08-27 | 哈尔滨工业大学（深圳） | Automatic vehicle classification method and related device |
| CN111611414B (en) * | 2019-02-22 | 2023-10-24 | 杭州海康威视数字技术股份有限公司 | Vehicle searching method, device and storage medium |
| CN110378237A (en) * | 2019-06-21 | 2019-10-25 | 浙江工商大学 | Human facial expression recognition method based on depth measure converged network |
| CN110378237B (en) * | 2019-06-21 | 2021-06-11 | 浙江工商大学 | Facial expression recognition method based on depth measurement fusion network |
| CN110766077A (en) * | 2019-10-24 | 2020-02-07 | 上海眼控科技股份有限公司 | Method, device and equipment for screening sketch in evidence chain image |
| CN110837572A (en) * | 2019-11-15 | 2020-02-25 | 北京推想科技有限公司 | Image retrieval method and device, readable storage medium and electronic equipment |
| CN111310844A (en) * | 2020-02-26 | 2020-06-19 | 广州华工邦元信息技术有限公司 | Vehicle identification model construction method and device and identification method and device |
| CN112085034A (en) * | 2020-09-11 | 2020-12-15 | 北京埃福瑞科技有限公司 | Rail transit train positioning method and system based on machine vision |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108197326B (en) | 2023-05-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108197326A (en) | A kind of vehicle retrieval method and device, electronic equipment, storage medium | |
| CN109584248B (en) | Infrared target instance segmentation method based on feature fusion and dense connection network | |
| Chen et al. | Progressive lidar adaptation for road detection | |
| Mery | Aluminum casting inspection using deep learning: a method based on convolutional neural networks | |
| CN109886222B (en) | Face recognition method, neural network training method, device and electronic equipment | |
| CN113486981B (en) | RGB image classification method based on multi-scale feature attention fusion network | |
| CN110852316A (en) | Image tampering detection and positioning method adopting convolution network with dense structure | |
| Türkyılmaz et al. | License plate recognition system using artificial neural networks | |
| CN111709311A (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| CN108875754B (en) | Vehicle re-identification method based on multi-depth feature fusion network | |
| CN112529005B (en) | Target detection method based on semantic feature consistency supervision pyramid network | |
| CN111291812B (en) | Method and device for acquiring attribute category, storage medium and electronic device | |
| Ni et al. | A Generative adversarial learning strategy for enhanced lightweight crack delineation networks | |
| CN111242144A (en) | Method and device for detecting abnormality of power grid equipment | |
| CN114495010A (en) | Cross-modal pedestrian re-identification method and system based on multi-feature learning | |
| CN110457992A (en) | Pedestrian based on Bayes's optimisation technique recognition methods, device and system again | |
| CN115966010A (en) | Expression recognition method based on attention and multi-scale feature fusion | |
| Liang et al. | Cross-scene foreground segmentation with supervised and unsupervised model communication | |
| Buenaposada et al. | Improving multi-class boosting-based object detection | |
| CN111709382A (en) | Human body trajectory processing method and device, computer storage medium and electronic equipment | |
| CN117197763A (en) | Road crack detection method and system based on cross attention guide feature alignment network | |
| CN116977249A (en) | Defect detection method, model training method and device | |
| CN114627493A (en) | Gait feature-based identity recognition method and system | |
| CN114187476A (en) | Vehicle insurance information checking method, device, equipment and medium based on image analysis | |
| Liu et al. | Component detection for power line inspection using a graph-based relation guiding network |
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 |