CN106250555B - Vehicle retrieval method and device based on big data - Google Patents
Vehicle retrieval method and device based on big data Download PDFInfo
- Publication number
- CN106250555B CN106250555B CN201610671729.0A CN201610671729A CN106250555B CN 106250555 B CN106250555 B CN 106250555B CN 201610671729 A CN201610671729 A CN 201610671729A CN 106250555 B CN106250555 B CN 106250555B
- Authority
- CN
- China
- Prior art keywords
- target vehicle
- marker
- image
- vehicle image
- probability
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Abstract
The vehicle retrieval method and device based on big data that the present invention provides a kind of, this method comprises: extracting the brand identity of target vehicle in target vehicle image；Determine that each pixel in target vehicle image corresponds to the probability of each marker, wherein marker includes one of annual test mark, goods of furniture for display rather than for use, ornament or a variety of；The probability and the corresponding probability threshold value of each marker that each marker is corresponded to according to pixel each in target vehicle image, determine position of each marker in target vehicle image；The characteristics of image of each marker is extracted according to position of each marker in target vehicle image；According to the brand identity of the characteristics of image of marker each in target vehicle image and target vehicle in multiple vehicle images to be retrieved searched targets vehicle.It can accurately be retrieved in magnanimity vehicle image through the invention and obtain target vehicle, improve the accuracy of target vehicle retrieval.
Description
Technical field
The present invention relates to technical field of image processing, in particular to a kind of vehicle retrieval method based on big data
And device.
Background technique
With the rapid development of our country's economy, city size it is continuous expand, vehicle fleet size increase substantially and society
The raising of awareness of safety, more and more occasion needs are retrieved in magnanimity vehicle image according to the image of target vehicle,
Target vehicle is obtained, for example, public security big data system has the function of to refer to provide user vehicle figure to scheme to search vehicle
As the image as target vehicle, the traveling record of target vehicle is searched in bayonet magnanimity record.
There are many vehicle retrieval methods for prior art offer, for the image according to target vehicle in magnanimity vehicle image
Lookup obtains target vehicle.The detailed process of vehicle retrieval method in the prior art are as follows: utilize sift, surf, dog scheduling algorithm
The entire image feature for extracting target vehicle image extracts database using identical algorithm using this feature as target signature
In each vehicle image entire image feature, using this feature as feature to be matched, calculate target signature with it is each to be matched
Euclidean distance between feature, using the nearest corresponding vehicle of feature to be matched of Euclidean distance as target vehicle.
Inventor has found under study for action, since above-mentioned entire image feature includes the feature of driver, works as vehicle
Do not become, driver change when, entire image feature will change, hence for same vehicle, when target vehicle image
In driver and database in vehicle image in driver it is inconsistent when, will lead to that it fails to match, influence to retrieve
As a result accuracy, therefore, vehicle retrieval method in the prior art have the problem of retrieval accuracy difference.
Summary of the invention
In view of this, the purpose of the present invention is to provide a kind of vehicle retrieval method and device based on big data, it can
Accurately retrieval obtains target vehicle in magnanimity vehicle image, improves the accuracy of target vehicle retrieval.
In a first aspect, the embodiment of the invention provides a kind of vehicle retrieval method based on big data, which comprises
Extract the brand identity of target vehicle in target vehicle image；Determine that each pixel corresponds to each in the target vehicle image
The probability of marker, wherein the marker includes one of annual test mark, goods of furniture for display rather than for use, ornament or a variety of；According to the target carriage
Each pixel corresponds to the probability and the corresponding probability threshold value of each marker of each marker in image, really
Fixed position of each marker in the target vehicle image；According to each marker in the target vehicle figure
Extract the characteristics of image of each marker in position as in；According to the marker each in the target vehicle image
The brand identity of characteristics of image and the target vehicle retrieves the target vehicle in multiple vehicle images to be retrieved.
With reference to first aspect, the first possible embodiment the embodiment of the invention provides first aspect, wherein described
Extract the brand identity of target vehicle described in the target vehicle image, comprising: carry out to the target vehicle image cycle
Process of convolution, nonlinear transformation and pondization processing, obtain the first extraction result；Result is extracted to described first to carry out at dimensionality reduction
Reason, and nonlinear transformation is carried out to the result after dimensionality reduction, the brand for obtaining target vehicle described in the target vehicle image is special
Sign.
With reference to first aspect, the embodiment of the invention provides second of first aspect possible embodiments, wherein described
Determine that each pixel in the target vehicle image corresponds to the probability of each marker, comprising: to the target vehicle
Image cycle carries out process of convolution, nonlinear transformation and pondization processing, obtains the first definitive result；To first definitive result
Circulation carries out process of convolution and nonlinear transformation, obtains the second definitive result；Probability calculation is carried out to second definitive result,
Obtain the probability that each pixel in the target vehicle image corresponds to each marker.
With reference to first aspect, the third possible embodiment the embodiment of the invention provides first aspect, wherein described
The probability and each marker pair of each marker are corresponded to according to each pixel in the target vehicle image
The probability threshold value answered determines position of each marker in the target vehicle image, comprising: in the target vehicle
Each pixel that the probability of corresponding current flag object is greater than the corresponding probability threshold value of the current flag object is determined in image
For the corresponding each pixel of the current flag object；By the connected region of the corresponding each pixel composition of the current flag object
Position of the position in domain as the current flag object in the target vehicle image.
With reference to first aspect, the embodiment of the invention provides the 4th kind of possible embodiments of first aspect, wherein described
The characteristics of image of each marker is extracted according to position of each marker in the target vehicle image, is wrapped
It includes: obtaining the characteristic pattern of the target vehicle image；According to the size of the characteristic pattern, the size of the target vehicle image,
Position of each marker in the target vehicle image determines the position of the marker each in the characteristic pattern
It sets；The corresponding feature in position of the marker each in the characteristic pattern is mentioned as the characteristics of image of each marker
It takes out.
Above-mentioned embodiment with reference to first aspect, the embodiment of the invention provides the 5th kind of possible implementations of first aspect
Mode, wherein the characteristics of image and the target vehicle according to the marker each in the target vehicle image
Brand identity retrieve the target vehicle in multiple vehicle images to be retrieved, comprising: calculate separately the target vehicle figure
It is similar between the characteristics of image of each marker and the characteristics of image of each marker in current vehicle image to be retrieved as in
Degree, wherein each marker in each marker and the current vehicle image to be retrieved in the target vehicle image
It corresponds；Calculate current vehicle to be retrieved in the brand identity and the current vehicle image to be retrieved of the target vehicle
Similarity between brand identity；Summation is weighted to each similarity being calculated, obtains the target vehicle and institute
State the matching degree between current vehicle to be retrieved；Using the highest vehicle to be retrieved of matching degree as the target vehicle.
Second aspect, the embodiment of the invention provides a kind of vehicle retrieval device based on big data, described device include:
Brand extraction module, for extracting the brand identity of target vehicle in target vehicle image；Probability determination module, for determining
State the probability that each pixel in target vehicle image corresponds to each marker, wherein the marker includes annual test mark, pendulum
One of part, ornament are a variety of；Position determination module, for corresponding each according to pixel each in the target vehicle image
The probability of a marker and the corresponding probability threshold value of each marker, determine each marker in the mesh
Mark the position in vehicle image；Marker feature extraction module is used for according to each marker in the target vehicle figure
Extract the characteristics of image of each marker in position as in；Target retrieval module, for according to the target vehicle image
In the characteristics of image of each marker and the brand identity of the target vehicle examined in multiple vehicle images to be retrieved
Suo Suoshu target vehicle.
In conjunction with second aspect, the first possible embodiment the embodiment of the invention provides second aspect, wherein described
Brand extraction module includes: the first extracting sub-module, for carrying out process of convolution, non-linear to the target vehicle image cycle
Transformation and pondization processing, obtain the first extraction result；Second extracting sub-module, for being dropped to the first extraction result
Dimension processing, and nonlinear transformation is carried out to the result after dimensionality reduction, obtain the product of target vehicle described in the target vehicle image
Board feature.
In conjunction with second aspect, the embodiment of the invention provides second of second aspect possible embodiments, wherein described
Probability determination module includes: first to determine submodule, for carrying out process of convolution, non-linear to the target vehicle image cycle
Transformation and pondization processing, obtain the first definitive result；Second determines submodule, carries out for recycling to first definitive result
Process of convolution and nonlinear transformation obtain the second definitive result；Third determines submodule, for second definitive result into
Row probability calculation obtains the probability that each pixel in the target vehicle image corresponds to each marker.
In conjunction with second aspect, the third possible embodiment the embodiment of the invention provides second aspect, wherein described
Position determination module, comprising: pixel determines submodule, for will correspond to current flag object in the target vehicle image
It is corresponding each that each pixel that probability is greater than the corresponding probability threshold value of the current flag object is determined as the current flag object
A pixel；Position determination submodule, for by the connected region of the corresponding each pixel composition of the current flag object
Position of the position as the current flag object in the target vehicle image.
In conjunction with second aspect, the embodiment of the invention provides the 4th kind of possible embodiments of second aspect, wherein described
Marker feature extraction module includes: characteristic pattern acquisition submodule, for obtaining the characteristic pattern of the target vehicle image；Mark
Object location determines submodule, for size, the size of the target vehicle image, each mark according to the characteristic pattern
Position of the object in the target vehicle image determines the position of the marker each in the characteristic pattern；Feature extraction
Module, for using the corresponding feature in position of the marker each in the characteristic pattern as the image of each marker
Feature extraction comes out.
In conjunction with the above-mentioned embodiment of second aspect, the embodiment of the invention provides the 5th kind of possible implementations of second aspect
Mode, wherein the target retrieval module includes: the first computational submodule, for calculating separately in the target vehicle image
Similarity in the characteristics of image of each marker and current vehicle image to be retrieved between the characteristics of image of each marker,
In, each marker one in each marker and the current vehicle image to be retrieved in the target vehicle image is a pair of
It answers；Second computational submodule, in the brand identity and the current vehicle image to be retrieved for calculating the target vehicle when
Similarity between the brand identity of preceding vehicle to be retrieved；Third computational submodule, for each similarity being calculated
It is weighted summation, obtains the matching degree between the target vehicle and the current vehicle to be retrieved；Target determines submodule,
For using the highest vehicle to be retrieved of matching degree as the target vehicle.
In the embodiment of the present invention, using the brand identity of annual test mark, goods of furniture for display rather than for use, ornament and vehicle as retrieval foundation, in magnanimity
Searched targets vehicle in image.Since annual test mark, goods of furniture for display rather than for use, ornament are what the personal habits based on driver were arranged, the product of vehicle
Board feature is fixed and invariable, thus through the embodiment of the present invention in the vehicle retrieval method and device based on big data, will
Four features combine, and can accurately retrieve target vehicle according to the personal habits of driver in mass data, from
And improve the accuracy of target vehicle retrieval.
To enable the above objects, features and advantages of the present invention to be clearer and more comprehensible, preferred embodiment is cited below particularly, and cooperate
Appended attached drawing, is described in detail below.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached
Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair
The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this
A little attached drawings obtain other relevant attached drawings.
Fig. 1 shows the process signal of the vehicle retrieval method based on big data provided by first embodiment of the invention
Figure；
Fig. 2 shows the deep learnings of the brand identity provided by first embodiment of the invention for extracting target vehicle
The schematic diagram of network structure；
Fig. 3 shows the training schematic diagram of the deep learning network structure in Fig. 2 provided by the embodiment of the present invention；
Fig. 4 shows showing for the deep learning network structure for being used to determine pixel probability provided by the embodiment of the present invention
It is intended to；
Fig. 5 shows the training schematic diagram of the deep learning network structure in Fig. 4 provided by the embodiment of the present invention；
Fig. 6 shows the process signal of the vehicle retrieval device based on big data provided by first embodiment of the invention
Figure.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
Middle attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is only
It is a part of the embodiment of the present invention, instead of all the embodiments.The present invention being usually described and illustrated herein in the accompanying drawings is real
The component for applying example can be arranged and be designed with a variety of different configurations.Therefore, of the invention to what is provided in the accompanying drawings below
The detailed description of embodiment is not intended to limit the range of claimed invention, but is merely representative of selected reality of the invention
Apply example.Based on the embodiment of the present invention, those skilled in the art institute obtained without making creative work
There are other embodiments, shall fall within the protection scope of the present invention.
In view of vehicle retrieval method in the prior art has the problem of retrieval accuracy difference, the embodiment of the present invention is provided
A kind of vehicle retrieval method and device based on big data, is described below by embodiment.
Embodiment one
Fig. 1 shows the process signal of the vehicle retrieval method based on big data provided by first embodiment of the invention
Figure, as shown, method includes the following steps:
Step S101 extracts the brand identity of target vehicle in target vehicle image.
Target vehicle image is the image comprising target vehicle, and target vehicle is vehicle to be found, the mesh of the present embodiment
Be that target vehicle is found according to target vehicle image in magnanimity vehicle image, to realize the retrieval of target vehicle.
In view of deep learning network structure has the advantages that the model accuracy learnt automatically and study obtains is high, this
In step, the brand identity of target vehicle in target vehicle image is extracted based on trained deep learning network structure.
The brand identity for extracting target vehicle in target vehicle image based on trained deep learning network structure is specific
It is acted including following two: (1) process of convolution, nonlinear transformation and pondization being carried out to target vehicle image cycle and handled, obtained
To the first extraction as a result, (2), which extract result to first, carries out dimension-reduction treatment, and nonlinear transformation is carried out to the result after dimensionality reduction,
Obtain the brand identity of target vehicle in target vehicle image.
Fig. 2 is the deep learning network knot for the brand identity for extracting target vehicle that first embodiment of the invention provides
The schematic diagram of structure, as shown in Fig. 2, the deep learning network structure includes five layers, wherein first layer is to the 4th layer including convolution
Layer, active coating and pond layer, convolutional layer, active coating and the sequential connection of pond layer, layer 5 include full articulamentum and active coating, entirely
Articulamentum and active coating are linked in sequence.First layer is to all sequential connections of layer 5.
During the work time, target vehicle image is input to first layer to deep learning network structure in Fig. 2, in first layer
Convolutional layer process of convolution is carried out to target vehicle image, the result of process of convolution is input to the active coating in first layer, the
Active coating in one layer carries out nonlinear transformation to the result of process of convolution, and the result of nonlinear transformation is input in first layer
Pond layer, pond layer in first layer carries out pond processing to the result of nonlinear transformation, and pond result is input to the
Convolutional layer in two layers, convolutional layer, active coating and pond layer in the second layer according to first layer processing mode, to pond result
Process of convolution, nonlinear transformation and pondization processing are successively carried out, and pond result is input to the convolutional layer in third layer, with
This circulation, until the pond layer in the 4th layer exports final pond as a result, the final pond result is above-mentioned first
Extract result；First extraction result is input to the full articulamentum in layer 5 by pond layer in the 4th layer, and full articulamentum is to the
One, which extracts result, carries out dimension-reduction treatment, and the result after dimensionality reduction is input to the active coating in layer 5, the activation in layer 5
Layer carries out nonlinear transformation to the result after dimensionality reduction, obtains the brand identity of target vehicle.
When training the deep learning network structure in Fig. 2, set depth learning network structure includes convolutional layer, activation
Layer, pond layer, full articulamentum, triple loss function layer.In the training process, prepare a large amount of training image first as instruction
Practice data set, and the vehicle for marking vehicle sub-brand name of the same race belongs to same vehicle brand.
The training principle of deep learning network structure is to train otherness between the image of same vehicle brand to minimize,
Otherness between the image of different vehicle brand maximizes.There is triple loss function in triple loss function layer, it should
The principle of function is: selecting a sample at random from training dataset τ, which is known as Anchor, then randomly selects one again
The sample that a and Anchor (being denoted as x_a) belongs to the sample of same vehicle brand and one belongs to different vehicle brand, the two
Sample it is corresponding be known as Positive (being denoted as x_p) and Negative (being denoted as x_n), thus composition one (Anchor,
Positive, Negative) triple, the feature representation of triple is denoted as respectively:Triple loss
The purpose of function is exactly to pass through study, makes the distance between x_a and x_p feature representation as small as possible, and the feature of x_a and x_n
The distance between expression is as big as possible, and to allow and have one between the distance between x_a and x_n and the distance between x_a and x_p
A the smallest interval α, formulation are expressed as follows:
In formulaRefer to the square operation symbol of Euclidean distance, 2 be norm；τ refers to training dataset.
Corresponding objective function is obtained according to above-mentioned formula are as follows:
The purpose of deep learning network structure training is exactly that the distance between the image of same vehicle brand is allowed to be less than difference
The distance between image of vehicle brand, using a spacing distance, purpose makes them separate a minimum interval, if not
MeetThen illustrate that triple loss function has loss, needs to pass through training
Amendment loss, loss function is minimized.Wherein, the loss of triple loss function is measured with Euclidean distance ,+indicate in []
Value be greater than zero when, take the value for loss, when minus, loss zero.
Fig. 3 is the training schematic diagram of the deep learning network structure in Fig. 2 that first embodiment of the invention provides, such as Fig. 3
Shown, in training process, deep learning network structure includes seven layers, and first layer is to the 4th layer including sequentially connected convolution
Layer, active coating and pond layer, layer 5 include sequentially connected full articulamentum and active coating, and layer 6 is full articulamentum, and the 7th
Layer is triple loss function layer.When training the deep learning network structure in Fig. 2, a large amount of triple training datas are inputted
To first layer, process of convolution, nonlinear transformation and pondization are successively carried out to training data by first layer and handled, and result is defeated
Enter to the second layer, the second layer continues the output result to first layer and carries out process of convolution, nonlinear transformation and pondization processing, with this
It repeats, until the output result of the 4th layer of warp is input to layer 5, layer 5 carries out dimension-reduction treatment to the 4th layer of output result
And nonlinear transformation, and result is input to layer 6, layer 6 carries out dimension-reduction treatment to the output result of layer 5, by dimensionality reduction
As a result it is input to layer 7, layer 7 calculates loss using output result of the triple loss function to layer 6, and will calculate
Loss out is successively reversely passed in above-mentioned network structure, and the training parameter in network structure is adjusted, to guarantee that training obtains
The deep learning network structure of loss reduction.
In the present embodiment, the output result of layer 5 in obtained deep learning network structure will be trained as the product of vehicle
Board feature, therefore compared with Fig. 3, the structure in Fig. 2 only has first five layer.In the present embodiment, additionally it is possible to the mesh obtained to extraction
The brand identity for marking vehicle carries out depth Hash coding, to reduce characteristic storage size and improve vehicle retrieval performance.
Step S102 determines that each pixel in target vehicle image corresponds to the probability of each marker, wherein marker
Including one of annual test mark, goods of furniture for display rather than for use, ornament or a variety of.
In view of deep learning network structure has the advantages that the model accuracy learnt automatically and study obtains is high, this
In step, determine that each pixel corresponds to each marker in target vehicle image based on trained deep learning network structure
Probability.
Determine that each pixel corresponds to each mark in target vehicle image based on trained deep learning network structure
The probability of object specifically includes following three movements: (1) carrying out process of convolution, nonlinear transformation and pond to target vehicle image cycle
Change processing, obtains the first definitive result, and (2) recycle the first definitive result and carry out process of convolution and nonlinear transformation, obtains the
Two definitive results, (3) to the second definitive result carry out probability calculation, obtain each pixel in target vehicle image correspond to it is each
The probability of marker.
Fig. 4 is the signal for the deep learning network structure for determining pixel probability that first embodiment of the invention provides
Figure, as shown in figure 4, the deep learning network structure includes eight layers, wherein first layer and the second layer include convolutional layer, active coating
With pond layer, convolutional layer, active coating and pond layer are linked in sequence, third layer to layer 7 include the convolutional layer being linked in sequence and
Active coating, the 8th layer includes probability calculation function.First layer is to the 8th layer of all sequential connection.
During the work time, target vehicle image is input to first layer to deep learning network structure in Fig. 4, in first layer
Convolutional layer process of convolution is carried out to target vehicle image, the result of process of convolution is input to the active coating in first layer, the
Active coating in one layer carries out nonlinear transformation to the result of process of convolution, and the result of nonlinear transformation is input in first layer
Pond layer, pond layer in first layer carries out pond processing to the result of nonlinear transformation, and pond result is input to the
Convolutional layer in two layers, convolutional layer, active coating and pond layer in the second layer according to first layer processing mode, to pond result
Process of convolution, nonlinear transformation and pondization processing are successively carried out, obtains the first definitive result, and the first definitive result is inputted
Convolutional layer into third layer, the convolutional layer in third layer carry out process of convolution to pond result, and convolution results are input to the
Active coating in three layers, the active coating in third layer carries out nonlinear transformation to the result of process of convolution, by nonlinear transformation
As a result it is input to the 4th layer, is recycled with this, until the pond layer in layer 7 exports the second definitive result, the 8th layer utilizes probability
It calculates function and probability calculation is carried out to the second definitive result, obtain each pixel in target vehicle image and correspond to each marker
Probability.
When training the deep learning network structure in Fig. 4, set depth learning network structure includes convolutional layer, activation
Layer, pond layer and recurrence loss function layer.In the training process, prepare a large amount of training image first as training data,
The pixel in the region of each marker and background is labeled as different labels on every training image, such as will
It is 1 that the pixel of annual test mark region, which sets value, and it is 2 that the pixel of goods of furniture for display rather than for use region, which is set value, by ornament location
It is 3 that the pixel in domain, which sets value, and it is 0 that the pixel of background region, which is set value, in this way, each pixel in every image
A class label is all had, to generate the corresponding label image of every vehicle pictures.
Further, deep learning network structure is trained, obtains optimal deep learning network structure, thus logical
It crosses the optimal deep learning network structure and determines that each pixel in target vehicle image corresponds to the probability of each marker.?
In training process, guarantee the corresponding label of each pixel calculated and label image as far as possible one using loss function layer is returned
It causes.
Fig. 5 is the training schematic diagram of the deep learning network structure in Fig. 4 that first embodiment of the invention provides, such as Fig. 5
It is shown, in training process, original vehicle image and training sample are input to jointly in deep learning network structure, when the 8th layer
In recurrence loss function loss result is calculated after, the adjustment of calculated loss result is successively reversely passed into network structure
In, the training parameter in network structure is adjusted, to guarantee that training obtains optimal deep learning network structure.
Step S103 corresponds to the probability and each mark of each marker according to pixel each in target vehicle image
The corresponding probability threshold value of object determines position of each marker in target vehicle image.
This step specifically includes following procedure: (1) being greater than the probability of corresponding current flag object in target vehicle image
Each pixel of the corresponding probability threshold value of current flag object is determined as the corresponding each pixel of current flag object；(2) will work as
The position of the connected region of the corresponding each pixel composition of preceding marker is as current flag object in target vehicle image
Position.
In the above process (1), the probability that each pixel in target vehicle image corresponds to each marker is analyzed, for working as
Each pixel that probability is greater than probability threshold value is determined as the corresponding each pixel of current flag object by preceding marker；It is above-mentioned
In process (2), determines the position of the connected region of the corresponding each pixel composition of current flag object, refer specifically to the connected region
Boundary rectangle position, the position using the position of the boundary rectangle as current flag object in target vehicle image.
During specific implementation, the probability of current flag object can be corresponded to according to pixel each in target vehicle image,
And probability e.g., is greater than or equal to current flag by target vehicle image binaryzation by the corresponding probability threshold value of current flag object
It is 255 that the pixel of the corresponding probability threshold value of object, which sets value, i.e., white, and probability is less than the corresponding probability threshold value of current flag object
It is 0, i.e. black that pixel, which sets value,.The connected domain in binary image is calculated, using the position of the boundary rectangle of connected domain as working as
The position of preceding marker.
Detailed process is as follows for connected domain in calculating binary image:
(1) white pixel (pixel value 255) continuous in every a line is formed a sequence and called by progressive scanning picture
One group (run), and write down it starting point start, it terminal end and the line number where it；
(2) for the group in all rows other than the first row, if it is not all overlapped with all groups in previous row
Its new label is then given in region, if it only has overlapping region with a group in lastrow, that by lastrow is rolled into a ball
Label be assigned to it, if there are overlapping region in 2 or more groups of it and lastrow, will have in the group of overlapping region most with it
Small label is as the label currently rolled into a ball, while these labels are considered of equal value right, illustrate that they belong to one kind；
(3) by it is all it is of equal value to be converted to equivalent sequence (be same class assuming that label 2 and label 3 are that equivalence is right,
Label 3 is also of equal value right with label 4, then the label in this sequence represents by equivalence to being converted into equivalent sequence { 2,3,4 }
Same class), the label of all groups is traversed, equivalent sequence is searched, since 1, new to the group one in same equivalent sequence
Mark the label as the equivalent sequence；
It (4) will be in the label filling image of the equivalent sequence of each group；
(5) according to the image filled in, the region that identical label is constituted is determined as a connected domain, each connected domain is corresponding
It is a marker, the position of the boundary rectangle of connected domain is determined as to the position of marker.
Step S104, the image for extracting each marker according to position of each marker in target vehicle image are special
Sign.
This step is implemented as follows:
(1) characteristic pattern of target vehicle image is obtained；The spy of target vehicle image is extracted by deep learning network structure
Sign figure.For example, extracting characteristic pattern of conv5 layers of the vehicle feature as target vehicle image by model Alexnet, wherein
Model Alexnet can also be model googlenet, and conv5 layers can also be conv3 or conv4.
(2) position according to the size of characteristic pattern, the size, each marker of target vehicle image in target vehicle image
Set the position for determining marker each in characteristic pattern.For example, if position of the marker in target vehicle image be (x1, y1,
X2, y2), the size of conv5 layers of characteristic pattern is [h1, w1], and the size of target vehicle image is [h, w], then characteristic pattern is opposite
In scaling kw=w1/w, kh=h1/h of vehicle image, then position of the marker on characteristic pattern be (kw*x1, kh*y1,
Kw*x2, kh*y2).
(3) the corresponding feature in position of marker each in characteristic pattern is gone out as the image characteristics extraction of each marker
Come.
In another preferred embodiment, the size for setting the characteristics of image of each marker is consistent, according to setting
Size extracts the characteristics of image of each marker.Specifically, the size of the characteristics of image of each marker is set as poolh*
The characteristics of image of each marker is divided into poolh*poolw sub-block, is maximized in each sub-block by poolw
Feature of that feature as the sub-block, the size of the feature finally formed are poolh*poolw.As the spy of target vehicle image
When sign figure there are multiple, the characteristics of image of the marker extracted in every characteristic pattern is together in series, as final mark
The feature of will object.
Above-mentioned movement (2) and (3) can sample the realization of (ROIpooling) technology by area-of-interest.The present embodiment
In, additionally it is possible to it is encoded using characteristics of image of the Hash coding techniques to the marker of extraction, so that it is big to reduce characteristic storage
It is small and improve vehicle retrieval performance.
In the present embodiment, mentioned by above-mentioned movement (1) to (3) according to position of each marker in target vehicle image
The characteristics of image for taking each marker can be improved the extraction rate of the characteristics of image of marker, accelerate the retrieval of target vehicle
Efficiency.
Step S105, according to the brand identity of the characteristics of image of marker each in target vehicle image and target vehicle
The searched targets vehicle in multiple vehicle images to be retrieved.
Specifically, it for current vehicle image to be retrieved, is obtained using above-mentioned steps S101 to step S104 current to be checked
The current character of vehicle to be retrieved and the characteristics of image of each marker, calculate separately target vehicle figure in rope vehicle image
It is similar between the characteristics of image of each marker and the characteristics of image of each marker in current vehicle image to be retrieved as in
Degree, wherein each marker in each marker and current vehicle image to be retrieved in target vehicle image corresponds,
For example, calculating annual test target characteristics of image and annual test target characteristics of image in current vehicle image to be retrieved in target vehicle image
Between similarity, similarity can be indicated with Euclidean distance here, calculate the brand identity of target vehicle with it is current to be retrieved
Similarity in vehicle image between the brand identity of current vehicle to be retrieved, similarity can be indicated with Euclidean distance here,
Summation is weighted to each similarity being calculated, obtains the matching degree between target vehicle and current vehicle to be retrieved,
For example, annual test target weight is 3, the weight of goods of furniture for display rather than for use is 3, and the weight of ornament is 4, and the weight of brand identity is 2, according to the weight
With each similarity calculation matching degree, using the highest vehicle to be retrieved of matching degree as target vehicle.
In the present embodiment, due in each deep learning network structure of training, needing using great amount of samples data, therefore
The present embodiment is substantially a kind of vehicle retrieval method based on big data.
In the embodiment of the present invention, using the brand identity of annual test mark, goods of furniture for display rather than for use, ornament and vehicle as retrieval foundation, in magnanimity
Searched targets vehicle in image.Since annual test mark, goods of furniture for display rather than for use, ornament are what the personal habits based on driver were arranged, the product of vehicle
Board feature is fixed and invariable, thus through the embodiment of the present invention in the vehicle retrieval method based on big data, by this four
Feature combines, and target vehicle can be accurately retrieved according to the personal habits of driver in mass data, to improve
The accuracy of target vehicle retrieval.
Embodiment two
The vehicle retrieval method based on big data in corresponding above-described embodiment one, second embodiment of the invention provide one
Vehicle retrieval device of the kind based on big data, as shown in fig. 6, the device includes:
Brand extraction module 61, for extracting the brand identity of target vehicle in target vehicle image；
Probability determination module 62, for determining that each pixel in target vehicle image corresponds to the probability of each marker,
Wherein, marker includes one of annual test mark, goods of furniture for display rather than for use, ornament or a variety of；
Position determination module 63, for according to pixel each in target vehicle image correspond to the probability of each marker with
And the corresponding probability threshold value of each marker, determine position of each marker in target vehicle image；
Marker feature extraction module 64, it is each for being extracted according to position of each marker in target vehicle image
The characteristics of image of marker；
Target retrieval module 65, for the characteristics of image and target vehicle according to marker each in target vehicle image
Brand identity in multiple vehicle images to be retrieved searched targets vehicle.
Above-mentioned brand extraction module 61 is equivalent to above-mentioned trained for extracting the depth of the brand identity of vehicle
Network structure is practised, brand extraction module 61 includes: the first extracting sub-module, for carrying out at convolution to target vehicle image cycle
Reason, nonlinear transformation and pondization processing, obtain the first extraction result；Second extracting sub-module, for extracting result to first
Dimension-reduction treatment is carried out, and nonlinear transformation is carried out to the result after dimensionality reduction, obtains the brand of target vehicle in target vehicle image
Feature.
Above-mentioned probability determination module 62 is equivalent to above-mentioned trained for determining the depth of the probability of each pixel
Learning network structure, probability determination module 62 includes: the first determining submodule, for carrying out convolution to target vehicle image cycle
Processing, nonlinear transformation and pondization processing, obtain the first definitive result；Second determines submodule, for the first definitive result
Circulation carries out process of convolution and nonlinear transformation, obtains the second definitive result；Third determines submodule, ties for determining to second
Fruit carries out probability calculation, obtains the probability that each pixel in target vehicle image corresponds to each marker.
Above-mentioned position determination module 63, comprising: pixel determines submodule, for will be to should in target vehicle image
It is corresponding that each pixel that the probability of preceding marker is greater than the corresponding probability threshold value of current flag object is determined as current flag object
Each pixel；Position determination submodule, the position of the connected region for forming the corresponding each pixel of current flag object
Set the position as current flag object in target vehicle image.
Above-mentioned marker feature extraction module 64 includes: characteristic pattern acquisition submodule, for obtaining target vehicle image
Characteristic pattern；Marker position determines submodule, for size, the size of target vehicle image, each mark according to characteristic pattern
Position of the object in target vehicle image determines the position of marker each in characteristic pattern；Feature extraction submodule, being used for will
The corresponding feature in the position of each marker comes out as the image characteristics extraction of each marker in characteristic pattern.
Above-mentioned target retrieval module 65 includes: the first computational submodule, each in target vehicle image for calculating separately
Similarity in the characteristics of image of marker and current vehicle image to be retrieved between the characteristics of image of each marker, wherein
Each marker in each marker and current vehicle image to be retrieved in target vehicle image corresponds；Second calculates
Submodule, for calculating brand identity and the brand spy of current vehicle to be retrieved in current vehicle image to be retrieved of target vehicle
Similarity between sign；Third computational submodule obtains target for being weighted summation to each similarity being calculated
Matching degree between vehicle and current vehicle to be retrieved；Target determines submodule, is used for the highest vehicle to be retrieved of matching degree
As target vehicle.
In the present embodiment, due in each deep learning network structure of training, needing using great amount of samples data, therefore
The present embodiment is substantially a kind of vehicle retrieval device based on big data.
In the embodiment of the present invention, using the brand identity of annual test mark, goods of furniture for display rather than for use, ornament and vehicle as retrieval foundation, in magnanimity
Searched targets vehicle in image.Since annual test mark, goods of furniture for display rather than for use, ornament are what the personal habits based on driver were arranged, the product of vehicle
Board feature is fixed and invariable, thus through the embodiment of the present invention in the vehicle retrieval device based on big data, by this four
Feature combines, and target vehicle can be accurately retrieved according to the personal habits of driver in mass data, to improve
The accuracy of target vehicle retrieval.
Vehicle retrieval device based on big data provided by the embodiment of the present invention can in equipment specific hardware or
Person is installed on software or firmware in equipment etc..The technology of device provided by the embodiment of the present invention, realization principle and generation
Effect is identical with preceding method embodiment, and to briefly describe, Installation practice part does not refer to place, can refer to preceding method reality
Apply corresponding contents in example.It is apparent to those skilled in the art that for convenience and simplicity of description, foregoing description
System, the specific work process of device and unit, the corresponding process during reference can be made to the above method embodiment, herein no longer
It repeats.
In embodiment provided by the present invention, it should be understood that disclosed device and method, it can be by others side
Formula is realized.The apparatus embodiments described above are merely exemplary, for example, the division of the unit, only one kind are patrolled
Function division is collected, there may be another division manner in actual implementation, in another example, multiple units or components can combine or can
To be integrated into another system, or some features can be ignored or not executed.Another point, shown or discussed is mutual
Coupling, direct-coupling or communication connection can be INDIRECT COUPLING or communication link by some communication interfaces, device or unit
It connects, can be electrical property, mechanical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme
's.
In addition, each functional unit in embodiment provided by the invention can integrate in one processing unit, it can also
To be that each unit physically exists alone, can also be integrated in one unit with two or more units.
It, can be with if the function is realized in the form of SFU software functional unit and when sold or used as an independent product
It is stored in a computer readable storage medium.Based on this understanding, technical solution of the present invention is substantially in other words
The part of the part that contributes to existing technology or the technical solution can be embodied in the form of software products, the meter
Calculation machine software product is stored in a storage medium, including some instructions are used so that a computer equipment (can be a
People's computer, server or network equipment etc.) it performs all or part of the steps of the method described in the various embodiments of the present invention.
And storage medium above-mentioned includes: that USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), arbitrary access are deposited
The various media that can store program code such as reservoir (RAM, Random Access Memory), magnetic or disk.
It should also be noted that similar label and letter indicate similar terms in following attached drawing, therefore, once a certain Xiang Yi
It is defined in a attached drawing, does not then need that it is further defined and explained in subsequent attached drawing, in addition, term " the
One ", " second ", " third " etc. are only used for distinguishing description, are not understood to indicate or imply relative importance.
Finally, it should be noted that embodiment described above, only a specific embodiment of the invention, to illustrate the present invention
Technical solution, rather than its limitations, scope of protection of the present invention is not limited thereto, although with reference to the foregoing embodiments to this hair
It is bright to be described in detail, those skilled in the art should understand that: anyone skilled in the art
In the technical scope disclosed by the present invention, it can still modify to technical solution documented by previous embodiment or can be light
It is readily conceivable that variation or equivalent replacement of some of the technical features；And these modifications, variation or replacement, do not make
The essence of corresponding technical solution is detached from the spirit and scope of technical solution of the embodiment of the present invention.Should all it cover in protection of the invention
Within the scope of.Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (8)
1. a kind of vehicle retrieval method based on big data, which is characterized in that the described method includes:
Extract the brand identity of target vehicle in target vehicle image；
Determine that each pixel in the target vehicle image corresponds to the probability of each marker, wherein the marker includes
One of annual test mark, goods of furniture for display rather than for use, ornament are a variety of；
The probability and each mark of each marker are corresponded to according to each pixel in the target vehicle image
The corresponding probability threshold value of object determines position of each marker in the target vehicle image；
The characteristics of image of each marker is extracted according to position of each marker in the target vehicle image；
According to the brand identity of the characteristics of image of the marker each in the target vehicle image and the target vehicle
The target vehicle is retrieved in multiple vehicle images to be retrieved；
The brand identity for extracting target vehicle described in the target vehicle image, comprising:
Process of convolution, nonlinear transformation and pondization processing are carried out to the target vehicle image cycle, obtain the first extraction knot
Fruit；
Result is extracted to described first and carries out dimension-reduction treatment, and nonlinear transformation is carried out to the result after dimensionality reduction, obtains the mesh
Mark the brand identity of target vehicle described in vehicle image；
Each pixel corresponds to the probability of each marker in the determination target vehicle image, comprising:
Process of convolution, nonlinear transformation and pondization processing are carried out to the target vehicle image cycle, obtain the first definitive result；
First definitive result is recycled and carries out process of convolution and nonlinear transformation, obtains the second definitive result；
Probability calculation is carried out to second definitive result, each pixel in the target vehicle image is obtained and corresponds to each institute
State the probability of marker.
2. the method according to claim 1, wherein described according to each pixel in the target vehicle image
The probability and the corresponding probability threshold value of each marker of corresponding each marker, determine that each marker exists
Position in the target vehicle image, comprising:
The probability of corresponding current flag object is greater than the corresponding probability threshold of the current flag object in the target vehicle image
Each pixel of value is determined as the corresponding each pixel of the current flag object；
The position of the connected region of the corresponding each pixel composition of the current flag object is existed as the current flag object
Position in the target vehicle image.
3. the method according to claim 1, wherein it is described according to each marker in the target vehicle
Extract the characteristics of image of each marker in position in image, comprising:
Obtain the characteristic pattern of the target vehicle image；
According to the size of the characteristic pattern, the size of the target vehicle image, each marker in the target vehicle
Position in image determines the position of the marker each in the characteristic pattern；
Using the corresponding feature in position of the marker each in the characteristic pattern as the characteristics of image of each marker
It extracts.
4. method according to any one of claims 1 to 3, which is characterized in that described according in the target vehicle image
The characteristics of image of each marker and the brand identity of the target vehicle are retrieved in multiple vehicle images to be retrieved
The target vehicle, comprising:
Calculate separately in the target vehicle image characteristics of image of each marker with it is each in current vehicle image to be retrieved
Similarity between the characteristics of image of marker, wherein each marker in the target vehicle image and it is described currently to
The each marker retrieved in vehicle image corresponds；
Calculate the brand of current vehicle to be retrieved in the brand identity and the current vehicle image to be retrieved of the target vehicle
Similarity between feature；
Summation is weighted to each similarity being calculated, obtain the target vehicle and the current vehicle to be retrieved it
Between matching degree；
Using the highest vehicle to be retrieved of matching degree as the target vehicle.
5. a kind of vehicle retrieval device based on big data, which is characterized in that described device includes:
Brand extraction module, for extracting the brand identity of target vehicle in target vehicle image；
Probability determination module, for determining that each pixel in the target vehicle image corresponds to the probability of each marker,
In, the marker includes one of annual test mark, goods of furniture for display rather than for use, ornament or a variety of；
Position determination module, for corresponding to the probability of each marker according to each pixel in the target vehicle image
And the corresponding probability threshold value of each marker, determine position of each marker in the target vehicle image
It sets；
Marker feature extraction module, it is each for being extracted according to position of each marker in the target vehicle image
The characteristics of image of a marker；
Target retrieval module, for the characteristics of image and the mesh according to the marker each in the target vehicle image
The brand identity of mark vehicle retrieves the target vehicle in multiple vehicle images to be retrieved；
The brand extraction module includes:
First extracting sub-module, for carrying out process of convolution, nonlinear transformation and pond to the target vehicle image cycle
Processing, obtains the first extraction result；
Second extracting sub-module carries out dimension-reduction treatment for extracting result to described first, and carries out to the result after dimensionality reduction non-
Linear transformation obtains the brand identity of target vehicle described in the target vehicle image；
The probability determination module includes:
First determines submodule, for carrying out process of convolution, nonlinear transformation and Chi Huachu to the target vehicle image cycle
Reason, obtains the first definitive result；
Second determines submodule, carries out process of convolution and nonlinear transformation for recycling to first definitive result, obtains the
Two definitive results；
Third determines submodule, for carrying out probability calculation to second definitive result, obtains in the target vehicle image
Each pixel corresponds to the probability of each marker.
6. device according to claim 5, which is characterized in that the position determination module, comprising:
Pixel determines submodule, for the probability of corresponding current flag object to be greater than described work as in the target vehicle image
Each pixel of the corresponding probability threshold value of preceding marker is determined as the corresponding each pixel of the current flag object；
Position determination submodule, for making the position of the connected region of the corresponding each pixel composition of the current flag object
For position of the current flag object in the target vehicle image.
7. device according to claim 5, which is characterized in that the marker feature extraction module includes:
Characteristic pattern acquisition submodule, for obtaining the characteristic pattern of the target vehicle image；
Marker position determines submodule, for according to the size of the characteristic pattern, size of the target vehicle image, each
Position of the marker in the target vehicle image determines the position of the marker each in the characteristic pattern；
Feature extraction submodule, for using the corresponding feature in position of the marker each in the characteristic pattern as each institute
The image characteristics extraction for stating marker comes out.
8. according to the described in any item devices of claim 5 to 7, which is characterized in that the target retrieval module includes:
First computational submodule, for calculate separately in the target vehicle image characteristics of image of each marker with currently to
Retrieve the similarity in vehicle image between the characteristics of image of each marker, wherein each in the target vehicle image
Each marker in marker and the current vehicle image to be retrieved corresponds；
Second computational submodule, in the brand identity and the current vehicle image to be retrieved for calculating the target vehicle when
Similarity between the brand identity of preceding vehicle to be retrieved；
Third computational submodule, for being weighted summation to each similarity being calculated, obtain the target vehicle with
Matching degree between the current vehicle to be retrieved；
Target determines submodule, for using the highest vehicle to be retrieved of matching degree as the target vehicle.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610671729.0A CN106250555B (en) | 2016-08-15 | 2016-08-15 | Vehicle retrieval method and device based on big data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610671729.0A CN106250555B (en) | 2016-08-15 | 2016-08-15 | Vehicle retrieval method and device based on big data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106250555A CN106250555A (en) | 2016-12-21 |
| CN106250555B true CN106250555B (en) | 2019-07-09 |
Family
ID=57592836
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610671729.0A Active CN106250555B (en) | 2016-08-15 | 2016-08-15 | Vehicle retrieval method and device based on big data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106250555B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107730905A (en) * | 2017-06-13 | 2018-02-23 | 银江股份有限公司 | Multitask fake license plate vehicle vision detection system and method based on depth convolutional neural networks |
| CN107871124B (en) * | 2017-11-15 | 2018-08-10 | 陕西师范大学 | A kind of Remote Sensing Target detection method based on deep neural network |
| CN108921019B (en) * | 2018-05-27 | 2022-03-08 | 北京工业大学 | Gait recognition method based on GEI and TripletLoss-DenseNet |
| 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 |
| CN109165620B (en) * | 2018-09-04 | 2022-10-11 | 张晓飞 | Detection method and system of electric vehicle and terminal equipment |
| CN110807415B (en) * | 2019-10-31 | 2023-04-07 | 南通大学 | Traffic checkpoint vehicle intelligent retrieval system and method based on annual inspection marks |
| CN112818736A (en) * | 2020-12-10 | 2021-05-18 | 西南交通大学 | Emergency command big data supporting platform |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1928892A (en) * | 2006-09-20 | 2007-03-14 | 王枚 | Method and device for license plate location recognition, vehicle-logo location recognition and vehicle type |
| CN103077384A (en) * | 2013-01-10 | 2013-05-01 | 北京万集科技股份有限公司 | Method and system for positioning and recognizing vehicle logo |
| CN103530366A (en) * | 2013-10-12 | 2014-01-22 | 湖北微模式科技发展有限公司 | Vehicle searching method and system based on user-defined features |
| CN105139011A (en) * | 2015-09-07 | 2015-12-09 | 浙江宇视科技有限公司 | Method and apparatus for identifying vehicle based on identification marker image |
| CN105447451A (en) * | 2015-11-13 | 2016-03-30 | 东方网力科技股份有限公司 | Method and device for retrieving object markers |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8254670B2 (en) * | 2009-02-25 | 2012-08-28 | Toyota Motor Engineering & Manufacturing North America, Inc. | Self-learning object detection and classification systems and methods |
-
2016
- 2016-08-15 CN CN201610671729.0A patent/CN106250555B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1928892A (en) * | 2006-09-20 | 2007-03-14 | 王枚 | Method and device for license plate location recognition, vehicle-logo location recognition and vehicle type |
| CN103077384A (en) * | 2013-01-10 | 2013-05-01 | 北京万集科技股份有限公司 | Method and system for positioning and recognizing vehicle logo |
| CN103530366A (en) * | 2013-10-12 | 2014-01-22 | 湖北微模式科技发展有限公司 | Vehicle searching method and system based on user-defined features |
| CN105139011A (en) * | 2015-09-07 | 2015-12-09 | 浙江宇视科技有限公司 | Method and apparatus for identifying vehicle based on identification marker image |
| CN105447451A (en) * | 2015-11-13 | 2016-03-30 | 东方网力科技股份有限公司 | Method and device for retrieving object markers |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106250555A (en) | 2016-12-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106250555B (en) | Vehicle retrieval method and device based on big data | |
| Li et al. | Building-a-nets: Robust building extraction from high-resolution remote sensing images with adversarial networks | |
| Dvornik et al. | On the importance of visual context for data augmentation in scene understanding | |
| CN109948425B (en) | Pedestrian searching method and device for structure-aware self-attention and online instance aggregation matching | |
| Peng et al. | Syn2real: A new benchmark forsynthetic-to-real visual domain adaptation | |
| CN108416266B (en) | Method for rapidly identifying video behaviors by extracting moving object through optical flow | |
| CN110738207A (en) | character detection method for fusing character area edge information in character image | |
| CN105205453B (en) | Human eye detection and localization method based on depth self-encoding encoder | |
| CN103578093B (en) | Method for registering images, device and augmented reality system | |
| Zhang et al. | A GANs-based deep learning framework for automatic subsurface object recognition from ground penetrating radar data | |
| CN112347995A (en) | Unsupervised pedestrian re-identification method based on fusion of pixel and feature transfer | |
| CN106844739A (en) | A kind of Remote Sensing Imagery Change information retrieval method based on neutral net coorinated training | |
| CN112329771A (en) | Building material sample identification method based on deep learning | |
| CN110399760A (en) | A kind of batch two dimensional code localization method, device, electronic equipment and storage medium | |
| Yu et al. | Exemplar-based recursive instance segmentation with application to plant image analysis | |
| CN110210480A (en) | Character recognition method, device, electronic equipment and computer readable storage medium | |
| CN112084860A (en) | Target object detection method and device and thermal power plant detection method and device | |
| CN112149526A (en) | Lane line detection method and system based on long-distance information fusion | |
| Zaech et al. | Texture underfitting for domain adaptation | |
| CN111259814B (en) | Living body detection method and system | |
| CN111209886A (en) | Rapid pedestrian re-identification method based on deep neural network | |
| CN115410081A (en) | Multi-scale aggregated cloud and cloud shadow identification method, system, equipment and storage medium | |
| KR101391667B1 (en) | A model learning and recognition method for object category recognition robust to scale changes | |
| CN110942056A (en) | Clothing key point positioning method and device, electronic equipment and medium | |
| Bourennane et al. | An enhanced visual object tracking approach based on combined features of neural networks, wavelet transforms, and histogram of oriented gradients |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| PP01 | Preservation of patent right | ||
| PP01 | Preservation of patent right |
Effective date of registration: 20220726 Granted publication date: 20190709 |