CN108021693A - A kind of image search method and device - Google Patents
A kind of image search method and device Download PDFInfo
- Publication number
- CN108021693A CN108021693A CN201711364358.2A CN201711364358A CN108021693A CN 108021693 A CN108021693 A CN 108021693A CN 201711364358 A CN201711364358 A CN 201711364358A CN 108021693 A CN108021693 A CN 108021693A
- Authority
- CN
- China
- Prior art keywords
- image
- depth characteristic
- matrix
- retrieval
- depth
- 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.)
- Pending
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
Abstract
This application provides a kind of image search method and device, this method includes：Obtain retrieval image to be retrieved；Extract the depth characteristic of the retrieval image；Based on preset dimensionality reduction matrix, the dimension of the depth characteristic is reduced, obtains the depth characteristic by dimensionality reduction, wherein, the dimensionality reduction matrix is the formed matrix of the preceding specified row of the projection matrix for the linear discriminant analysis that training obtains in advance；From preset image library, retrieval and the matched at least width target image of the depth characteristic by dimensionality reduction.This method and device can improve the precision of image retrieval.
Description
Technical field
This application involves data analysis technique field, more particularly to a kind of image search method and device.
Background technology
With the continuous development of Internet technology, traditional text based retrieval mode can not meet the inspection of people
Rope demand, for this reason, there has been proposed the mode that information retrieval is carried out based on image.Such as, in some electric business websites or video network
Stand, user using picture as input, from database extract and associated with picture presence by analyzing picture by website
Commodity or video content etc., and the content retrieved is returned into user.
In image retrieval technologies, the expression way of characteristics of image is most important, can directly influence image retrieval
Precision and speed.At present, based on image carry out information retrieval during, typically using the color of image, texture,
The feature such as shape or edge is matched.Yet with object in the picture in database there are larger attitudes vibration, and
Angle, contrast and illumination of picture etc. are there is also many changes, according to the spy extracted in existing image retrieval mode
Sign can not adapt to many changes of picture in database, cause the accuracy of image retrieval low.
The content of the invention
In view of this, this application provides a kind of image search method and device, with improve the precision of image retrieval and
Efficiency.
To achieve the above object, the application provides following technical solution：A kind of image search method, including：
Obtain retrieval image to be retrieved；
Extract the depth characteristic of the retrieval image；
Based on preset dimensionality reduction matrix, the dimension of the depth characteristic is reduced, obtains the depth characteristic by dimensionality reduction, its
In, the dimensionality reduction matrix is the formed matrix of the preceding specified row of the projection matrix for the linear discriminant analysis that training obtains in advance；
From preset image library, retrieval and the matched at least width target image of the depth characteristic by dimensionality reduction.
Preferably, the depth characteristic of the extraction retrieval image, including：
The convolutional neural networks model obtained based on advance training, extracts the depth characteristic of the retrieval image.
Preferably, the depth characteristic of the extraction retrieval image, including：
Extract the vector of the depth characteristic of the retrieval image；
It is described that the dimension of the depth characteristic is reduced based on preset dimensionality reduction matrix, including：
By the vectorial and preset dimensionality reduction matrix multiple of the depth characteristic, to obtain the depth characteristic by dimensionality reduction
Vector.
Preferably, it is described from preset image library, retrieval and the depth characteristic by dimensionality reduction matched at least one
Width target image, including：
Determine that the depth between each feature in the depth characteristic and preset feature database by dimensionality reduction is special respectively
Levy matching degree；
According to the order of the depth characteristic matching degree from high to low, from the feature database, it is special to match the depth
The forward default quantity target signature of matching degree sequence is levied, the feature database includes being used for the multiple features for matching image；
From preset image library, at least width target to match with the default quantity target signature is retrieved respectively
Image.
Preferably, the projection matrix of the linear discriminant analysis obtains in the following way：
The data set for including m image pattern is obtained, wherein, the data set includes c classification sample set, and c is
Natural number more than or equal to 2, and m is the natural number more than or equal to 2；
The convolutional neural networks model obtained using advance training, extracts the depth of all image patterns in the data set
Feature, obtains the depth characteristic vector of each described image sample；
Depth characteristic vector based on each described image sample, calculates the depth of each classification sample set respectively
Feature average value, and the overall average of the depth characteristic of all categories sample set；
Utilize the depth characteristic average value of each classification sample set and the depth of all categories sample set
The overall average of feature is spent, calculates the inter _ class relationship matrix and within class scatter matrix of the data set；
Discriminating criterion based on linear discriminant analysis, the inter _ class relationship matrix and within class scatter matrix, calculate
Projection matrix.
On the other hand, present invention also provides a kind of image retrieving apparatus, including：
Image acquisition unit, for obtaining retrieval image to be retrieved；
Feature extraction unit, for extracting the depth characteristic of the retrieval image；
Feature Dimension Reduction unit, for based on preset dimensionality reduction matrix, reducing the dimension of the depth characteristic, obtaining by drop
The depth characteristic of dimension, wherein, the dimensionality reduction matrix is specified for the preceding of the projection matrix for the linear discriminant analysis that training obtains in advance
The formed matrix of row；
Image retrieval unit, for from preset image library, retrieval with it is described matched by the depth characteristic of dimensionality reduction
An at least width target image.
Preferably, the feature extraction unit, including：
Feature extraction subelement, for the convolutional neural networks model obtained based on advance training, extracts the retrieval
The depth characteristic of image.
Preferably, the feature extraction unit is specifically used for, and extracts the vector of the depth characteristic of the retrieval image；
The Feature Dimension Reduction matrix, including：
Feature Dimension Reduction submatrix, for by the vectorial and preset dimensionality reduction matrix multiple of the depth characteristic, with
To the depth characteristic vector by dimensionality reduction.
Preferably, described image retrieval unit, including：
Matching primitives subelement, for determine respectively the depth characteristic by dimensionality reduction with it is every in preset feature database
Depth characteristic matching degree between a feature；
Characteristic matching subelement, according to the order of the depth characteristic matching degree from high to low, from the feature database,
The forward default quantity target signature of the depth characteristic matching degree sequence is allotted, the feature database includes being used to match image
Multiple features；
Image retrieval subelement, for from preset image library, retrieving and the default quantity target signature respectively
At least width target image to match.
Preferably, further include：
Projection matrix determination unit, for the projection matrix to be calculated in the following way：
The data set for including m image pattern is obtained, wherein, the data set includes c classification sample set, its
In, c is the natural number more than or equal to 2, and m is the natural number more than or equal to 2；
The convolutional neural networks model obtained using advance training, extracts the depth of all image patterns in the data set
Feature, obtains the depth characteristic vector of each described image sample；
Depth characteristic vector based on each described image sample, calculates the depth of each classification sample set respectively
Feature average value, and the overall average of the depth characteristic of all categories sample set；
Utilize the depth characteristic average value of each classification sample set and the depth of all categories sample set
The overall average of feature is spent, calculates the inter _ class relationship matrix and within class scatter matrix of the data set；
Discriminating criterion based on linear discriminant analysis, the inter _ class relationship matrix and within class scatter matrix, calculate
Projection matrix.
It can be seen via above technical scheme that the application can extract retrieval after retrieval image to be retrieved is got
The depth characteristic of image, and before image retrieval is carried out using depth characteristic, preset dimensionality reduction matrix can be utilized to extracting
Depth characteristic carry out dimensionality reduction, since dimensionality reduction matrix for being specified before the projection matrix of linear discriminant analysis arranges the square that forms
Battle array, therefore, by the dimensionality reduction matrix reduce extract depth characteristic dimension while, can also reduce depth characteristic it
Between correlation, in this way, using by dimensionality reduction depth characteristic carry out image retrieval, can both reduce due to depth characteristic it
Between influence of the correlation to image retrieval precision, and the recall precision of image retrieval can be improved.
Brief description of the drawings
In order to illustrate more clearly of the technical solution of the embodiment of the present application, required use in being described below to embodiment
Attached drawing be briefly described, it should be apparent that, drawings in the following description are only embodiments herein, for this area
For those of ordinary skill, without creative efforts, it can also be obtained according to the attached drawing of offer other attached
Figure.
Fig. 1 shows a kind of flow diagram of image search method one embodiment of the application；
Fig. 2 shows a kind of flow diagram of another embodiment of image search method of the application；
Fig. 3 shows the flow diagram that projection matrix is calculated in a kind of image search method of the application；
Fig. 4 shows a kind of composition structure diagram of image retrieving apparatus one embodiment of the application.
Embodiment
The image search method and device of the application can be applied to possess data processing can any computer in, such as
The program can be applied in the server with image retrieval or other computer equipments.
Below in conjunction with the attached drawing in the embodiment of the present application, the technical solution in the embodiment of the present application is carried out clear, complete
Site preparation describes, it is clear that described embodiments are only a part of embodiments of the present application, instead of all the embodiments.It is based on
Embodiment in the application, those of ordinary skill in the art are obtained every other without making creative work
Embodiment, shall fall in the protection scope of this application.
Such as, referring to Fig. 1, it illustrates a kind of flow diagram of image search method one embodiment of the application, this reality
Applying the method for example can include：
S101, obtains retrieval image to be retrieved.
Wherein, in order to be distinguished with the image that later retrieval goes out, it will need what is retrieved in the embodiment of the present application
Image is known as retrieving image.
Such as, computer (e.g., retrieval server) receives the image that is inputted by search engine of user, and using the image as
It is currently needed for the image retrieved.
For example, when user needs the picture search Related product according to certain product, then can be by this product
Image as retrieval image be input in search engine.
S102, extracts the depth characteristic of the retrieval image.
Wherein, the depth characteristic of image refers to learn the feature by way of deep learning.
It is understood that depth characteristic can be the vector of default dimension N.Wherein, presetting dimension N can be as needed
Setting.
Wherein, extract the depth characteristic mode of the retrieval image can have it is a variety of.Such as, can be rolled up by training in advance
Product neural network model, and the depth characteristic of the retrieval image is extracted by the convolutional neural networks model.
S103, based on preset dimensionality reduction matrix, reduces the dimension of the depth characteristic, obtains the depth characteristic by dimensionality reduction.
Wherein, linear discriminant analysis (the Linear Discriminant that the dimensionality reduction matrix obtains for advance training
Analysis, LDA) projection matrix the formed matrix of preceding specified row.
Wherein, this specifies the columns of row less than total columns of the projection matrix.
Since in multiple dimensions of depth characteristic, the depth characteristic corresponding to more forward dimension is more important, therefore, this Shen
The preceding specified row that please choose projection matrix are used as dimensionality reduction matrix, and reduce the dimension of depth characteristic using the dimensionality reduction matrix, both
It can realize the purpose for the dimension for reducing depth characteristic, and can the characteristic of maximum limitation holding depth characteristic in itself.
It is understood that in the embodiment of the present application by extracting the depth characteristic of retrieval image, to utilize retrieval figure
The depth characteristic of picture carries out image retrieval, and can reducing feature, can not to adapt to the posture of image, illumination etc. in database all changeable
The situation of change, is conducive to improve the precision of image retrieval.But there is certain correlation between the vector of depth characteristic,
Also the accuracy rate of retrieval can be had an impact；And the dimension of depth characteristic is also higher, so that influence whether the speed of retrieval, because
This, in order to further improve the precision of retrieval and recall precision, needs using dimensionality reduction matrix to depth in the embodiment of the present application
Spend feature and carry out dimensionality reduction.
Wherein, the projection matrix of linear discriminant analysis is：Determined previously according to the discriminating criterion of linear discriminant analysis,
So that the training sample after projection has the square corresponding to the projecting direction of maximum inter _ class relationship and minimum within-cluster variance
Battle array.Wherein, which is for calculating training sample included in the data set of projection matrix, the number of the training sample
Measure to be multiple, and the plurality of training sample can be divided at least two classifications.
, can be with it is understood that handled using the projection matrix of linear discriminant analysis the depth characteristic of image
The correlation of the depth characteristic of image is reduced, and the dimensionality reduction matrix is the preceding specified row of the projection matrix, therefore, passes through the dimensionality reduction
Matrix reduces the dimension of the depth characteristic of the retrieval image, can reduce depth while the dimension of depth characteristic is reduced
The correlation between feature is spent, in this way, depth characteristic of the later use after dimensionality reduction carries out image retrieval, is conducive to improve inspection
Rope precision and recall precision.
It is understood that since depth characteristic is also a vector in itself, it is special to the depth using dimensionality reduction matrix
Sign carries out dimensionality reduction, by the dimensionality reduction matrix and the multiplication of vectors of the depth characteristic, so as to obtain special by the depth of dimensionality reduction
The vector of sign.
Wherein, the projection matrix that linear discriminant analysis is calculated may refer to the related introduction of Examples hereinafter,
S104, from preset image library, the matched at least width target figure of depth characteristic of dimensionality reduction is passed through in retrieval with this
Picture.
Wherein, obtaining retrieving required depth characteristic, i.e., this is after the depth characteristic of dimensionality reduction, using this by dropping
The depth characteristic of dimension, the process for the target image that retrieval matches with the retrieval image can have a variety of implementations, the application
The retrieving of embodiment pair is not any limitation as.
As it can be seen that in the embodiment of the present application, after retrieval image to be retrieved is got, the depth of retrieval image can be extracted
Feature is spent, and before image retrieval is carried out using depth characteristic, preset dimensionality reduction matrix can be utilized special to the depth extracted
Sign carries out dimensionality reduction, since dimensionality reduction matrix is by specifying the matrix for arranging and forming before the projection matrix of linear discriminant analysis, lead to
The dimensionality reduction matrix is crossed while the dimension of the depth characteristic extracted is reduced, the correlation between depth characteristic can also be reduced
Property, in this way, carrying out image retrieval using the depth characteristic by dimensionality reduction, can both reduce due to the correlation between depth characteristic
Influence of the property to image retrieval precision, and the recall precision of image retrieval can be improved.
For the ease of understanding the scheme of the embodiment of the present application, the image retrieval side with reference to Fig. 2 to the embodiment of the present application
Method is introduced.
Such as Fig. 2, it illustrates a kind of flow diagram of another embodiment of image search method of the application, the present embodiment
Method can include：
S201, obtains retrieval image to be retrieved.
S202, utilizes the depth characteristic vector f of the preset convolutional neural networks model extraction retrieval imagec。
Wherein, the depth characteristic vector fcDimension be default dimension N.
Wherein, which advances with multiple images training sample and trains to obtain, and uses
In the convolutional neural networks model of the depth level characteristics of extraction image.
Wherein, train to obtain the convolution god for extracting depth characteristic by the use of multiple images sample as image training sample
Mode through network model can have a variety of, which kind of mode no matter train to obtain the volume of the depth characteristic for extracting image using
Product neural network model is suitable for the embodiment of the present application, is not any limitation as herein.
Such as, the model that can be used extracts the feature conduct of pool5 layers of Resnet50 models for Resnet50 disaggregated models
Feature vector, the dimension of feature vector is 2048 dimensions.It is understood that convolutional neural networks model in the embodiment of the present application
Can be Resnet50 models, but it is understood that, in practical applications, pass through such as visual geometric group
(VisualGeometry Group, VGG) etc. classifies network model to extract the depth characteristic of the retrieval image of input similarly
Suitable for the embodiment of the present application, it is other network models outside other convolutional neural networks models to extract the inspection that can also have
The depth characteristic of rope image.
S203, using preset dimensionality reduction matrix W ', with the depth characteristic vector fcIt is multiplied, obtains the depth after dimensionality reduction degree
Spend feature vector ft。
Wherein, ft=W'*fc(formula one)；
Wherein, relative to depth characteristic vector fc, depth characteristic vector ftDimension and correlation reduce.
Wherein, dimensionality reduction matrix W ' the projection matrix W of LDA that is previously obtained for extractionoptForward P arranges formed matrix,
Wherein, P is more than 1 and is less than projection matrix WoptTotal columns.
In order to make it easy to understand, below to projection matrix W is calculatedoptProcess be introduced.
Such as, referring to Fig. 3, it is the stream for the projection matrix that the LDA is calculated in the embodiment of the present application using training sample
Journey schematic diagram, the process may include steps of S301 to S305：
S301, obtains the data set for including m image pattern；
Wherein, which includes c classification sample set, and c is the natural number more than or equal to 2, and m be more than or equal to
2 natural number.Assuming that the number of image pattern is n in i classification sample setsi, i is the natural number from 1 to c, then m=n1+n2
+…ni+…+nc。
Such as, the multiple images sample for belonging to identical money product can be included in the data set, and this multiple images sample can
To be the image pattern from multiclass product, in this way, the data set can include the image pattern of the identical money product of multiclass, and it is every
A classification sample set can include the multiple images sample of same class product.
S302, the convolutional neural networks model obtained using advance training, extracts all image samples in the data set respectively
This depth characteristic, obtains the depth characteristic vector of each image pattern；
Such as, the depth characteristic vector set that the depth characteristic vector of the corresponding all image patterns of the data set is formed can
To be expressed as F={ f1, f2... ... fj……fm, wherein, fjRepresent the depth characteristic vector of image pattern j, j be from 1 to
The natural number of m.
S303, the depth characteristic vector based on image pattern, the depth characteristic for calculating each classification sample set i respectively are put down
Average ui, and the overall average u of the depth characteristic of all categories sample set；
Wherein, the depth characteristic average value u of classification sample set iiIt can be calculated by equation below two：
Wherein, niFor the number of image pattern in classification sample set i, i is the natural number from 1 to c, dk∈ classi tables
Show image pattern dkIt is from 1 to n to belong to classification sample set i, kiNatural number, fjRepresent image pattern dkDepth characteristic to
Amount.
Wherein, the overall average u of the depth characteristic of all categories sample set i can be calculated by equation below three
Arrive：
Wherein, m be data set in image pattern total number, fjRepresent the depth characteristic vector of image pattern j.
S304, utilizes the depth characteristic average value u of each classification sample setiAnd the depth of all categories sample set
The overall average u of feature, calculates the inter _ class relationship matrix S of the data setbAnd within class scatter matrix Sw；
Wherein, inter _ class relationship matrix SbIt can be calculated by equation below four：
Wherein, njRepresent the number of image pattern in classification sample set j, ujRepresent that classification sample set j depth characteristics are put down
Average；U is the overall average of the depth characteristic of all categories sample set.
Wherein, within class scatter matrix SwIt can be calculated by equation below five：
Wherein, dkRepresent k-th of image pattern in classification sample set j, dk∈ classj represent image pattern dkBelong to class
Other sample set j, j are the natural number from 1 to c, wherein, k is from 1 to njNatural number.njFor image in classification sample set j
The number of sample.
S305, the discriminating criterion based on linear discriminant analysis, such scatter matrix and within class scatter matrix, meter
Calculate projection matrix.
Wherein, when calculating projection matrix, the discriminating criterion of linear discriminant analysis is quoted, also referred to as Fisher differentiates accurate
Then, wherein, Fisher differentiates that criterion is exactly to choose to causeThe vector for reaching maximum is used as projection side
To, wherein,For a column vector.Based on this, the vector expression W of projection matrix can be obtainedoptEquation below can be passed through
Five are calculated：
Wherein, by the way that to optimizing as above formula five, projection matrix W can be sought outoptIn each column vector w, so as to obtain
Projection matrix.
Wherein, on the premise of the dimension of the depth characteristic vector of image pattern is N, the dimension of the projection matrix can be
N*N, e.g., using the dimension of projection matrix obtained based on Resnet50 models as 2048*2048.
The calculating of projection matrix can be completed to step S305 by as above step S301.
S204, from preset feature database, matches and the depth characteristic vector f by dimensionality reductiontMatching degree is highest pre-
If quantity target signature.
Wherein, the highest default quantity target signature of matching degree refers to, specifically, can determine respectively described by drop
The depth characteristic matching degree between each feature in the depth characteristic of dimension and preset feature database；Then, according to the depth
The order of characteristic matching degree from high to low, from the feature database, match the depth characteristic matching degree sort it is forward pre-
If quantity target signature.Wherein, the feature database includes being used for the multiple features for matching image.Pass through the feature in feature database
Characteristic matching can be carried out with the image in preset image library.
Such as, the depth characteristic vector f after reducing dimension can be utilizedtWith the feature in preset depth characteristic storehouse into
Row matching, and find out and the depth characteristic vector ftMatching degree it is highest before default quantity target signature.
In one implementation, depth characteristic vector f can be calculatedtWith each feature in feature database it is vectorial away from
From, and using the distance as matching degree, distance value is smaller, then matching degree is higher.
Wherein, Euclidean distance, COS distance, related coefficient etc. can be used by calculating the distance between two feature vectors.
In order to make it easy to understand, said exemplified by calculating the matching degree between feature vector, X and feature vector Y using related coefficient
It is bright, matching degree (i.e. distance) ρ between feature vector, X and feature vector YX,YIt can represent as follows：
S205, from preset image library, retrieves at least width to match with the default quantity target signature respectively
Target image.
Optionally, when image is the image information of product or product, after the target image is retrieved, can also obtain
Take the corresponding product link of the target image.
A kind of image search method of corresponding the application, the embodiment of the present application additionally provide a kind of image retrieving apparatus.
Such as, referring to Fig. 4, it illustrates a kind of composition structure diagram of image retrieving apparatus one embodiment of the application.
As shown in Figure 4, which can include：Image acquisition unit 401, feature extraction unit 402, feature
Dimensionality reduction unit 403 and image retrieval unit 404.
Wherein, image acquisition unit 401, for obtaining retrieval image to be retrieved；
Feature extraction unit 402, for extracting the depth characteristic of the retrieval image；
Feature Dimension Reduction unit 403, for based on preset dimensionality reduction matrix, reducing the dimension of the depth characteristic, obtain through
The depth characteristic of dimensionality reduction is crossed, wherein, before the dimensionality reduction matrix is the projection matrix for the linear discriminant analysis that training obtains in advance
Specify the formed matrix of row；
Image retrieval unit 404, for from preset image library, retrieval to be matched with described by the depth characteristic of dimensionality reduction
An at least width target image.
As it can be seen that after getting retrieval image to be retrieved in image acquisition unit, feature extraction unit can extract retrieval
The depth characteristic of image, before being retrieved using depth characteristic, Feature Dimension Reduction unit can utilize preset dimensionality reduction matrix pair
The depth characteristic extracted carries out dimensionality reduction, since dimensionality reduction matrix before the projection matrix of linear discriminant analysis by specifying row to form
Matrix, therefore, by the dimensionality reduction matrix reduce extract depth characteristic dimension while, can also reduce depth spy
Correlation between sign, in this way, carrying out image retrieval using by the depth characteristic of dimensionality reduction in image retrieval unit, can both subtract
Less due to influence of the correlation between depth characteristic to image retrieval precision, and the recall precision of image retrieval can be improved.
In a kind of possible implementation, the feature extraction unit, including：
Feature extraction subelement, for the convolutional neural networks model obtained based on advance training, extracts the retrieval
The depth characteristic of image.
In a kind of possible implementation, the Feature Dimension Reduction matrix, including：
Feature Dimension Reduction submatrix, for by the vectorial and preset dimensionality reduction matrix multiple of the depth characteristic, with
To the depth characteristic vector by dimensionality reduction.
In a kind of possible implementation, described image retrieval unit, including：
Matching primitives subelement, for determine respectively the depth characteristic by dimensionality reduction with it is every in preset feature database
Depth characteristic matching degree between a feature；
Characteristic matching subelement, for the order according to the depth characteristic matching degree from high to low, from the feature database
In, match the depth characteristic matching degree and sort forward default quantity target signature, the feature database include for
Multiple features with image；Image retrieval subelement, for from preset image library, retrieval and the default quantity to be a respectively
At least width target image that target signature matches.
In a kind of possible implementation, which can also include：
Projection matrix determination unit, for the projection matrix to be calculated in the following way：
The data set for including m image pattern is obtained, wherein, the data set includes c classification sample ni, wherein i
Natural number from 1 to c, c is the natural number more than or equal to 2, and m is the natural number more than or equal to 2；
The convolutional neural networks model obtained using advance training, extracts the depth of all image patterns in the data set
Feature, obtains the depth characteristic vector of each described image sample；
Depth characteristic vector based on each described image sample, calculates the depth characteristic of each classification sample respectively
Average value, and the overall average of the depth characteristic of all categories sample；
Utilize the depth characteristic average value of each classification sample and the depth characteristic of all categories sample
Overall average, calculates the inter _ class relationship matrix and within class scatter matrix of the data set；
Discriminating criterion based on linear discriminant analysis, the inter _ class relationship matrix and within class scatter matrix, calculate
Projection matrix.
For device embodiment, since it essentially corresponds to embodiment of the method, so related part is real referring to method
Apply the part explanation of example.Device embodiment described above is only schematical, wherein described be used as separating component
The unit of explanation may or may not be physically separate, can be as the component that unit is shown or can also
It is not physical location, you can with positioned at a place, or can also be distributed in multiple network unit.Can be according to reality
Need to select some or all of module therein to realize the purpose of this embodiment scheme.Those of ordinary skill in the art are not
In the case of making the creative labor, you can to understand and implement.
In several embodiments provided by the present invention, it should be understood that disclosed apparatus and method, are being not above
In spirit and scope, it can realize in other way.Current embodiment is a kind of exemplary example,
It should not be taken as limiting, given particular content should in no way limit the purpose of the application.For example, the unit or subelement
Division, is only a kind of division of logic function, can there is other dividing mode, such as multiple units or multiple when actually realizing
Subelement combines.In addition, multiple units can with or component can combine or be desirably integrated into another system, or some
Feature can be ignored, or not perform.
In addition, the schematic diagram of described system and method and different embodiments, can in without departing from scope of the present application
To combine or integrate with other systems, module, techniques or methods.It is another, shown or discussed mutual coupling or
Direct-coupling or communication connection can be by some interfaces, the INDIRECT COUPLING or communication connection of device or unit, can be electricity
Property, mechanical or other forms.
The above is only the embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications also should
It is considered as protection scope of the present invention.
Claims (10)
- A kind of 1. image search method, it is characterised in that including：Obtain retrieval image to be retrieved；Extract the depth characteristic of the retrieval image；Based on preset dimensionality reduction matrix, the dimension of the depth characteristic is reduced, obtains the depth characteristic by dimensionality reduction, wherein, institute Dimensionality reduction matrix is stated as the formed matrix of the preceding specified row of the projection matrix for the linear discriminant analysis that training obtains in advance；From preset image library, retrieval and the matched at least width target image of the depth characteristic by dimensionality reduction.
- 2. image search method according to claim 1, it is characterised in that the depth of the extraction retrieval image is special Sign, including：The convolutional neural networks model obtained based on advance training, extracts the depth characteristic of the retrieval image.
- 3. image search method according to claim 1, it is characterised in that the depth of the extraction retrieval image is special Sign, including：Extract the vector of the depth characteristic of the retrieval image；It is described that the dimension of the depth characteristic is reduced based on preset dimensionality reduction matrix, including：By the vectorial and preset dimensionality reduction matrix multiple of the depth characteristic, with obtain the depth characteristic by dimensionality reduction to Amount.
- 4. image search method according to claim 1, it is characterised in that it is described from preset image library, retrieval with The matched at least width target image of the depth characteristic by dimensionality reduction, including：The depth characteristic between each feature in the depth characteristic and preset feature database by dimensionality reduction is determined respectively With degree；According to the order of the depth characteristic matching degree from high to low, from the feature database, the depth characteristic is matched Sort forward default quantity target signature with degree, and the feature database includes being used for the multiple features for matching image；From preset image library, at least width target figure to match with the default quantity target signature is retrieved respectively Picture.
- 5. image search method according to claim 1, it is characterised in that the projection matrix of the linear discriminant analysis leads to Following manner is crossed to obtain：Obtain the data set comprising m image pattern, wherein, the data set includes c classification sample set, c for more than Natural number equal to 2, and m is the natural number more than or equal to 2；The convolutional neural networks model obtained using advance training, the depth for extracting all image patterns in the data set are special Sign, obtains the depth characteristic vector of each described image sample；Depth characteristic vector based on each described image sample, calculates the depth characteristic of each classification sample set respectively Average value, and the overall average of the depth characteristic of all categories sample set；It is special using the depth characteristic average value of each classification sample set and the depth of all categories sample set The overall average of sign, calculates the inter _ class relationship matrix and within class scatter matrix of the data set；Discriminating criterion based on linear discriminant analysis, the inter _ class relationship matrix and within class scatter matrix, calculate projection Matrix.
- A kind of 6. image retrieving apparatus, it is characterised in that including：Image acquisition unit, for obtaining retrieval image to be retrieved；Feature extraction unit, for extracting the depth characteristic of the retrieval image；Feature Dimension Reduction unit, for based on preset dimensionality reduction matrix, reducing the dimension of the depth characteristic, obtaining by dimensionality reduction Depth characteristic, wherein, the preceding specified row institute of the projection matrix for the linear discriminant analysis that the dimensionality reduction matrix obtains for advance training The matrix of composition；Image retrieval unit, for from preset image library, retrieval and the depth characteristic by dimensionality reduction to be matched at least One width target image.
- 7. image retrieving apparatus according to claim 6, it is characterised in that the feature extraction unit, including：Feature extraction subelement, for the convolutional neural networks model obtained based on advance training, extracts the retrieval image Depth characteristic.
- 8. image retrieving apparatus according to claim 6, it is characterised in that the feature extraction unit is specifically used for, and carries Take the vector of the depth characteristic of the retrieval image；The Feature Dimension Reduction matrix, including：Feature Dimension Reduction submatrix, for by the vectorial and preset dimensionality reduction matrix multiple of the depth characteristic, with obtain through Cross the depth characteristic vector of dimensionality reduction.
- 9. image retrieving apparatus according to claim 6, it is characterised in that described image retrieval unit, including：Matching primitives subelement, for determining the depth characteristic by dimensionality reduction and each spy in preset feature database respectively Depth characteristic matching degree between sign；Characteristic matching subelement, according to the order of the depth characteristic matching degree from high to low, from the feature database, matches The forward default quantity target signature of the depth characteristic matching degree sequence, the feature database include being used to match the more of image A feature；Image retrieval subelement, for from preset image library, retrieval and the default quantity target signature phase respectively At least width target image matched somebody with somebody.
- 10. image retrieving apparatus according to claim 6, it is characterised in that further include：Projection matrix determination unit, for the projection matrix to be calculated in the following way：The data set for including m image pattern is obtained, wherein, the data set includes c classification sample set, wherein, c is Natural number more than or equal to 2, and m is the natural number more than or equal to 2；The convolutional neural networks model obtained using advance training, the depth for extracting all image patterns in the data set are special Sign, obtains the depth characteristic vector of each described image sample；Depth characteristic vector based on each described image sample, calculates the depth characteristic of each classification sample set respectively Average value, and the overall average of the depth characteristic of all categories sample set；It is special using the depth characteristic average value of each classification sample set and the depth of all categories sample set The overall average of sign, calculates the inter _ class relationship matrix and within class scatter matrix of the data set；Discriminating criterion based on linear discriminant analysis, the inter _ class relationship matrix and within class scatter matrix, calculate projection Matrix.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711364358.2A CN108021693A (en) | 2017-12-18 | 2017-12-18 | A kind of image search method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711364358.2A CN108021693A (en) | 2017-12-18 | 2017-12-18 | A kind of image search method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108021693A true CN108021693A (en) | 2018-05-11 |
Family
ID=62073846
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711364358.2A Pending CN108021693A (en) | 2017-12-18 | 2017-12-18 | A kind of image search method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108021693A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108960412A (en) * | 2018-06-29 | 2018-12-07 | 北京京东尚科信息技术有限公司 | Image-recognizing method, device and computer readable storage medium |
| CN110069643A (en) * | 2019-03-08 | 2019-07-30 | 杭州晶彩数字科技有限公司 | The method and system of similar building picture searching |
| CN110647796A (en) * | 2019-08-02 | 2020-01-03 | 中山市奥珀金属制品有限公司 | Iris identification method and device |
| CN110659665A (en) * | 2019-08-02 | 2020-01-07 | 深圳力维智联技术有限公司 | Model construction method of different-dimensional features and image identification method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8818103B2 (en) * | 2009-03-04 | 2014-08-26 | Osaka Prefecture University Public Corporation | Image retrieval method, image retrieval program, and image registration method |
| CN105095435A (en) * | 2015-07-23 | 2015-11-25 | 北京京东尚科信息技术有限公司 | Similarity comparison method and device for high-dimensional image features |
| CN106326288A (en) * | 2015-06-30 | 2017-01-11 | 阿里巴巴集团控股有限公司 | Image search method and apparatus |
| CN106886599A (en) * | 2017-02-28 | 2017-06-23 | 北京京东尚科信息技术有限公司 | Image search method and device |
-
2017
- 2017-12-18 CN CN201711364358.2A patent/CN108021693A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8818103B2 (en) * | 2009-03-04 | 2014-08-26 | Osaka Prefecture University Public Corporation | Image retrieval method, image retrieval program, and image registration method |
| CN106326288A (en) * | 2015-06-30 | 2017-01-11 | 阿里巴巴集团控股有限公司 | Image search method and apparatus |
| CN105095435A (en) * | 2015-07-23 | 2015-11-25 | 北京京东尚科信息技术有限公司 | Similarity comparison method and device for high-dimensional image features |
| CN106886599A (en) * | 2017-02-28 | 2017-06-23 | 北京京东尚科信息技术有限公司 | Image search method and device |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108960412A (en) * | 2018-06-29 | 2018-12-07 | 北京京东尚科信息技术有限公司 | Image-recognizing method, device and computer readable storage medium |
| CN110069643A (en) * | 2019-03-08 | 2019-07-30 | 杭州晶彩数字科技有限公司 | The method and system of similar building picture searching |
| CN110647796A (en) * | 2019-08-02 | 2020-01-03 | 中山市奥珀金属制品有限公司 | Iris identification method and device |
| CN110659665A (en) * | 2019-08-02 | 2020-01-07 | 深圳力维智联技术有限公司 | Model construction method of different-dimensional features and image identification method and device |
| CN110659665B (en) * | 2019-08-02 | 2023-09-29 | 深圳力维智联技术有限公司 | Model construction method of different-dimension characteristics and image recognition method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108197532B (en) | The method, apparatus and computer installation of recognition of face | |
| You et al. | Relative CNN-RNN: Learning relative atmospheric visibility from images | |
| CN104715023B (en) | Method of Commodity Recommendation based on video content and system | |
| US20190108599A1 (en) | Facilitating Like-Minded User Pooling | |
| CN103810299B (en) | Image retrieval method on basis of multi-feature fusion | |
| CN108021693A (en) | A kind of image search method and device | |
| CN108230337A (en) | A kind of method that semantic SLAM systems based on mobile terminal are realized | |
| CN107862270A (en) | Face classification device training method, method for detecting human face and device, electronic equipment | |
| CN105243154B (en) | Remote sensing image retrieval method based on notable point feature and sparse own coding and system | |
| CN107844753A (en) | Pedestrian in video image recognition methods, device, storage medium and processor again | |
| CN109190643A (en) | Based on the recognition methods of convolutional neural networks Chinese medicine and electronic equipment | |
| CN109558902A (en) | A kind of fast target detection method | |
| CN106156693A (en) | The robust error correction method represented based on multi-model for facial recognition | |
| WO2020134102A1 (en) | Article recognition method and device, vending system, and storage medium | |
| CN110490444A (en) | Mark method for allocating tasks, device, system and storage medium | |
| CN107679539B (en) | Single convolution neural network local information and global information integration method based on local perception field | |
| CN109871762B (en) | Face recognition model evaluation method and device | |
| CN108984555A (en) | User Status is excavated and information recommendation method, device and equipment | |
| CN110472652A (en) | A small amount of sample classification method based on semanteme guidance | |
| CN109241379A (en) | A method of across Modal detection network navy | |
| CN110163567A (en) | Classroom roll calling system based on multitask concatenated convolutional neural network | |
| CN110322418A (en) | A kind of super-resolution image generates the training method and device of confrontation network | |
| CN109544262A (en) | Item recommendation method, device, electronic equipment, system and readable storage medium storing program for executing | |
| CN111680678A (en) | Target area identification method, device, equipment and readable storage medium | |
| CN110503099A (en) | Information identifying method and relevant device based on deep learning |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180511 |