CN108734209A - Feature recognition based on more images and equipment - Google Patents
Feature recognition based on more images and equipment Download PDFInfo
- Publication number
- CN108734209A CN108734209A CN201810470392.6A CN201810470392A CN108734209A CN 108734209 A CN108734209 A CN 108734209A CN 201810470392 A CN201810470392 A CN 201810470392A CN 108734209 A CN108734209 A CN 108734209A
- Authority
- CN
- China
- Prior art keywords
- image
- feature vector
- combination
- machine learning
- characteristic information
- 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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/757—Matching configurations of points or features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of feature recognition and equipment based on more images, the method includes:Multiple images are obtained, described multiple images are the image sequences with relevance for determining at least one identical characteristic information;The feature vector of described multiple images is obtained respectively;Multiple combinations are carried out to the feature vector of acquisition;The combination of the corresponding feature vector of described multiple images is identified to obtain image feature information using the first machine learning model, the combination of the corresponding feature vector of wherein different images differs, and first machine learning model is trained to obtain using sampling feature vectors combination and corresponding characteristic information.
Description
Technical field
The present invention relates to artificial intelligence image processing fields, and in particular to a kind of feature recognition based on more images and sets
It is standby.
Background technology
Using the means of the artificial intelligence such as machine learning, deep learning and neural network algorithm, to image carry out classification and
Identification is the more universal analysis method of industrial quarters and academia.Above-mentioned recognition methods needs first instruct initial model
Practice, trained mode can be led to using a large amount of single image and its corresponding label (characteristic information) input model, model
It crosses the weight learnt and one prediction is provided to input picture.After thinking that Model Identification accuracy reaches target, you can
Classification and Identification is carried out to the image of unknown characteristics information using the model, to determine its characteristic information.This method is to model
Input requirements are not high, can be adapted for most problems.
In some application fields, there is the images with stronger related information, such as eye fundus image.In general environment
Under, the left eye and eye image of detected person includes for same person, left eye and right eye it is necessary to be acquired
Exception (characteristic information) be that maximum probability is similar.Using this image, one can be obtained in a manner of existing model training
Identification model, but the model is to be identified respectively according to two images, two characteristic informations is finally obtained, with regard to identification process
There is no any relevances for speech.Although by the training of great amount of samples data, the accuracy of Model Identification can be improved, it is accurate
The bottleneck of true property is difficult to be broken, and in actual use, which is still likely to left-eye image and right eye for same person
Image recognition goes out different characteristic informations.
Since the existing characteristics of image identification method based on artificial intelligence does not account for the relevance of multiple images itself,
Therefore its accuracy when facing the problem of identifying more images is still to be improved.
Invention content
In view of this, the present invention provides a kind of characteristic recognition methods based on more images, including:
Obtain multiple images, described multiple images be for determine at least one identical characteristic information have relevance
Image sequence;
The feature vector of described multiple images is obtained respectively;
Multiple combinations are carried out to the feature vector of acquisition;
The combination of the corresponding feature vector of described multiple images is identified to obtain using the first machine learning model
Image feature information, wherein the combination of the corresponding feature vector of different images differs, first machine learning model is
It is trained using sampling feature vectors combination and corresponding characteristic information.
Optionally, the feature vector for obtaining described multiple images respectively, including:
Described multiple images are inputted into the second machine learning model respectively, make second machine learning model respectively to institute
It states multiple images to be identified, wherein second machine learning model is carried out using sample image and corresponding characteristic information
What training obtained;
Second machine learning model is obtained when identifying each described image, the feature of one of middle layer output
Vector.
Optionally, the middle layer is last layer in the middle layer of second machine learning model.
Optionally, the combination of described eigenvector is that different at least two feature vectors progress is end to end.
Optionally, each combination of eigenvectors of generation separately includes all feature vectors.
Optionally, the feature vector of described pair of acquisition carries out multiple combinations, including:
According to the correspondence of scheduled combination of eigenvectors and image sequence, from whole modes of combination of eigenvectors
Selected identical as described image quantity and corresponding combination of eigenvectors mode;
Feature vector is combined according to selected combination of eigenvectors mode.
Optionally, described multiple images are eye fundus image, and the quantity of described multiple images is 2, respectively left eye eyeground
Image and right eye eye fundus image, described image characteristic information are lesion information.
The present invention also provides a kind of more characteristics of image identification model training methods, including:
It obtains multiple images and its corresponding characteristic information, wherein described multiple images corresponds to identical characteristic information;
Obtain the feature vector of described multiple images;
Multiple combinations are carried out to the feature vector of acquisition;
Determining sample data, the sample data is the corresponding combination of eigenvectors of described image and the characteristic information,
The combination of the corresponding feature vector of wherein different images differs;
Machine learning model is trained using multiple sample datas, so that the machine learning model is according to institute
It states the corresponding combination of eigenvectors of image and identifies the characteristic information.
The present invention also provides another more characteristics of image identification model training methods, including:
Multiple sample datas are obtained, each sample data includes multiple images and its corresponding characteristic information, and
And the multiple images in each sample data correspond to identical characteristic information;
The multiple sample data is input in the first sort module of machine learning model, the first classification mould is made
Root tuber identifies the characteristic information according to described multiple images;
First sort module is obtained when identifying each described image, the feature of one of middle layer output to
Amount;
Multiple combinations are carried out to the feature vector of acquisition;
The corresponding feature vector of described multiple images is input in the second sort module of the machine learning model, is made
Second sort module identifies the characteristic information according to the combination of the corresponding feature vector of described multiple images.
The present invention also provides a kind of characteristic identificating equipments based on more images, including:At least one processor;And with
The memory of at least one processor communication connection;Wherein, the memory, which is stored with, to be held by one processor
Capable instruction, described instruction are executed by least one processor, so that at least one processor executes above-mentioned be based on
The characteristic recognition method of more images.
According to characteristic recognition method and equipment provided by the invention based on more images, the figure with relevance is obtained first
As sequence, and then obtain the feature vector of each image respectively, the feature of each image reflected with feature vector, by feature to
Amount is combined, so as to get combination of eigenvectors synthesis multiple images the characteristics of;By the different spies corresponding to each image
Sign vector combination input machine learning model, makes machine learning model be known according to the combination of eigenvectors of each associated image
The semantic feature that other image sequence is included so that identification process is not identified further according to single image, and utilize feature to
The combination of amount is established multiple images and is contacted, and so that machine learning model is considered the relevance of image sequence itself, it is possible thereby to carry
The accuracy of high recognition result.
Description of the drawings
It, below will be to specific in order to illustrate more clearly of the specific embodiment of the invention or technical solution in the prior art
Embodiment or attached drawing needed to be used in the description of the prior art are briefly described, it should be apparent that, in being described below
Attached drawing is some embodiments of the present invention, for those of ordinary skill in the art, before not making the creative labor
It puts, other drawings may also be obtained based on these drawings.
Fig. 1 is the flow chart of the characteristic recognition method based on more images in the embodiment of the present invention;
Fig. 2 is the flow chart of the characteristic recognition method based on double eye fundus images in the embodiment of the present invention;
Fig. 3 is a kind of flow chart of more characteristics of image identification model training methods in the embodiment of the present invention;
Fig. 4 is the flow chart of the more characteristics of image identification model training methods of another kind in the embodiment of the present invention.
Specific implementation mode
Technical scheme of the present invention is clearly and completely described below in conjunction with attached drawing, it is clear that described implementation
Example is a part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, ordinary skill
The every other embodiment that personnel are obtained without making creative work, shall fall within the protection scope of the present invention.
As long as in addition, technical characteristic involved in invention described below different embodiments non-structure each other
It can be combined with each other at conflict.
The embodiment of the present invention provides a kind of characteristic recognition method based on more images, and this method can be by computer, service
The electronic equipments such as device execute, and this method comprises the following steps as shown in Figure 1:
S1, obtain multiple images, these images be for determine at least one identical characteristic information have relevance
Image sequence.Illustrate as an example, it is assumed that there are one subjects there are one or more features, can respectively from
Different angles shoots the object to obtain multiple images, these images are above-mentioned image sequence;As another
Kind illustrates, it is assumed that and there are multiple reference objects, themselves there are one or more identical features, it can be respectively to these
Object is shot, and obtained image is above-mentioned image sequence.It is possible thereby to learn, the reference object of image in the present invention
It can be the same target, target can also be different, and these images itself serve as and determine consistent feature.Characteristic information
It is a kind of semantic information, such as the subject in expression picture is what object or what shape subject is,
Or subject is abnormal etc. there are what.
S2 obtains the feature vector of the above-mentioned multiple images with relevance respectively.The algorithm of acquisition feature vector has more
Kind, for example, can based on the principle of machine vision, according to the physical features such as pixel value, lines in image calculate feature to
Amount, or the means of artificial intelligence can also be utilized, by training characteristics identification model, from the hidden layer in model export feature to
Amount.This n image of in assuming that there is i1 ..., respectively can be obtained each image contract m dimensional feature vectors:
Featurei1=[v11,v12,v13,…,v1(m-1),v1m]、
Featurei2=[v21,v22,v23,…,v2(m-1),v2m]、
……
Featurein=[vn1,vn2,vn3,…,vn(m-1),vnm], this n feature vector.
S3 carries out multiple combinations to the feature vector of acquisition.There are many modes of combination, such as can choose n feature
A part in vector is combined, and the whole that can also be chosen in n feature vector is combined, after thus being combined
Feature vector Featurei1&Featurei2Or Featurei1&Featurei2&……&FeatureinDeng herein by symbol
" " should not be construed as oeprator as illustrating a kind of illustrative expression way of combination of eigenvectors, will hereinafter make
Use Feature&1……Feature&nIndicate different combination of eigenvectors.There are many combinational algorithms of vector, e.g. vectorial
Addition, vector cross product, vector are subtracted each other, vector connects etc..For different application scenarios, such as the different targets that is taken,
The factors such as the quantity of image, the type of characteristic information or quantity test various combinations and algorithm in advance, to be optimal
Recognition effect, this step can according to scheduled combination and algorithm according to n feature vector determine at least n it is different
Combination of eigenvectors, each combination of eigenvectors are obtained according at least two feature vectors.
S4 is identified to obtain figure the combination of the corresponding feature vector of multiple images using the first machine learning model
As characteristic information, wherein the combination of the corresponding feature vector of different images differs, the first machine learning model is to utilize sample
What the combination of eigen vector and corresponding characteristic information were trained.
In identification process, used first machine learning model can be deep learning model, neural network model
Deng the model trained by sample data using SGD, RMSProp scheduling algorithm specifically will below for training process
It describes in detail.In the present embodiment, it needs to input n data, each data to trained first machine learning model
Include the corresponding combination of eigenvectors of each image, such as first input data is Feature&1(corresponding to image i1),
Second input data is Feature&2(corresponding to image i2) ..., n-th of input data was Feature&n(correspond to image
In), the first machine learning model is made to export one or more features information according to this n input data, namely according to n image
Corresponding feature vector determines the semantic information of input picture.
The characteristic recognition method based on more images provided according to embodiments of the present invention obtains the figure with relevance first
As sequence, and then obtain the feature vector of each image respectively, the feature of each image reflected with feature vector, by feature to
Amount is combined, so as to get combination of eigenvectors synthesis multiple images the characteristics of;By the different spies corresponding to each image
Sign vector combination input machine learning model, makes machine learning model be known according to the combination of eigenvectors of each associated image
The semantic feature that other image sequence is included so that identification process is not identified further according to single image, and utilize feature to
The combination of amount is established multiple images and is contacted, and so that machine learning model is considered the relevance of image sequence itself, it is possible thereby to carry
The accuracy of high recognition result.
As a preferred embodiment, the present embodiment obtains the feature vector of image using the means of artificial intelligence.
Specifically, above-mentioned steps S2 may include steps of:
Multiple images are inputted the second machine learning model respectively, make the second machine learning model respectively to multiple figures by S21
As being identified, wherein the second machine learning model is trained to obtain using sample image and corresponding characteristic information.
In identification process, used second machine learning model can be deep learning model, neural network model
Deng, by be not limited to conventional machines study in such as SVM (Support vector machine), the sorting algorithms such as XGBoost
The model trained by sample data.In the present embodiment, it needs in the input of trained second machine learning model
N image is stated, which will be respectively according to this n image come identification feature information.
S22 obtains the second machine learning model when identify each image, the feature that one of middle layer exports to
Amount.It will be appreciated by those skilled in the art that machine learning model is made of input layer, hidden layer, output layer, input layer is for connecing
The image in input data namely the present embodiment is received, output layer is for exporting recognition result, i.e. characteristic information.For hidden layer
For, different models has the hidden layer of different number, and what is transmitted between hidden layer is in feature vector and parameter and weight etc.
Hold, the feature vector that current layer obtains representing certain pictorial information can be extracted per layer network, and transmitted to next layer.
The feature vector that one of hidden layer is extracted can be chosen in the present embodiment to answer for different about the selection of hidden layer
With factors such as scenes, such as the different targets that is taken, the quantity of image, the type of characteristic information or quantity in advance to each layer
Extraction result tested, determine optimal recognition effect, to preset choose a hidden layer.
According to the principle of prototype network it is found that the semantic information that extracts of last layer in hidden layer is most, thus it is most
In the case of preferably use hidden layer in last layer network extract feature vector.
According to above-mentioned preferred embodiment, the second machine learning is equivalent to a feature vector extraction module, utilizes machine learning
The feature vector of model extraction image has higher computational efficiency, and since most models generally have multiple hidden layers, because
Multiple feature vectors can be obtained for each image in this, can be chosen for different application scenarios defeated from different hidden layers
The feature vector gone out expands the range of selection, improves the practicability of this programme.
As described above, there are many combinations of feature vector, as a preferred embodiment, above-mentioned steps S3 can be with
Include the following steps:
S31, according to the correspondence of scheduled combination of eigenvectors and image sequence, from the whole side of combination of eigenvectors
Identical as amount of images and corresponding combination of eigenvectors mode is selected in formula.As one for example, such as when image
When quantity is 4, Feature can be obtained by step S2i1、Featurei2、Featurei3、Featurei4This 4 features to
Amount.By taking end to end this combination as an example, there are 16 kinds of combinations according to 4 character vectors of permutation and combination,
And finally need to input the feature vector of the first machine learning model to be 4, it can preset herein corresponding to each image
Combination of eigenvectors, wherein 4 kinds are selected from 16 combinations.
S32 is combined feature vector according to selected combination of eigenvectors mode.It is found, is pressed by test performance
Carry out the feature of first image in representative image sequence, according to the elder generation of " 2134 " according to the sequencing splicing feature vector of " 1324 "
Sequential concatenation feature vector carrys out the feature of second image in representative image sequence, splices spy according to the sequencing of " 3124 " afterwards
Sign vector carrys out the feature of third image in representative image sequence, is represented according to the sequencing of " 4132 " splicing feature vector
The feature of the 4th image, can obtain best performance in sequence, and it is empirical per pictures then to press this in this step
Corresponding sequence is spliced and combined.
When facing different application scenarios, testing feature vector the correspondence with image sequence can be combined in advance,
So that it is determined that wherein optimal combination, is then combined according to preset mode in identification process, thus carry
Computationally efficient and identification accuracy.
Technical solution provided by the invention is illustrated with reference to a specific application scenarios.One of the present invention
Embodiment provides a kind of characteristic recognition method based on double eye fundus images, as shown in Fig. 2, this method comprises the following steps:
S ' 1 obtains left eye eye fundus image i1 and right eye eye fundus image i2 from same detected person, the two eyeground figures
As for determining at least one identical characteristic information.Characteristic information can be a kind of tag along sort information, and such as normal, eyeground goes out
The label informations such as blood, macula lutea are abnormal, optic disk is abnormal.For the human body, the case where above- mentioned information is reflected usually occurs simultaneously
In eyes eye fundus image, therefore the two images are the image sequences for having relevance.
Left eye eye fundus image i1 and right eye eye fundus image i2 are inputted machine learning model model1 respectively, make machine by S ' 2
Learning model model1 is respectively identified eye eye fundus image i1 and right eye eye fundus image i2, wherein machine learning model
Model1 is trained to obtain using sample image and corresponding characteristic information.Specifically, for training model1's
Sample image can be the eye fundus image from other detected persons, and with their corresponding characteristic informations, i.e., such as normal, eye
The label informations such as bottom bleeding, abnormal, the optic disk exception of macula lutea.
S ' 3 obtains machine learning model model1 when identifying left eye eye fundus image i1 and right eye eye fundus image i2, wherein
The feature vector of one middle layer output.Such as the m dimensional feature vectors that wherein the last one hidden layer extracts are chosen, it obtains
Featurei1=[v11,v12,v13,…,v1(m-1),v1m] and Featurei2=[v21,v22,v23,…,v2(m-1),v2m]。
S ' 4 carries out multiple combinations to the feature vector of acquisition, and the combination of feature vector in the present embodiment is will be different
At least two feature vectors carry out it is end to end, and generate each combination of eigenvectors separately include all features to
Amount.In the present embodiment, since only there are two feature vector, most multipotency obtains two kinds of combinations:
With
Here, can preset left eye eye fundus image i1 corresponds to [v11…v1m, v21…v2m], right eye eye fundus image i2
Corresponding to [v21…v2m, v11…v1m]。
S ' 5, using machine learning model model2 to left eye eye fundus image i1 and the corresponding features of right eye eye fundus image i2
Vector combination is identified to obtain image feature information, and wherein machine learning model model2 is to utilize sampling feature vectors group
What conjunction and corresponding characteristic information were trained.Specifically, can come from for training the sample data of model2
The feature vector of the eye fundus image of other detected persons and corresponding characteristic information (label information).
In this step, the data of input machine learning model model2 are [v11…v1m, v21…v2m] and [v21…v2m,
v11…v1m], model2 exports final recognition result, such as whether left eye eye fundus image i1 and right eye eye fundus image i2 belong to just
Often, a certain kind in fundus hemorrhage, macula lutea exception, optic disk exception.
The characteristic recognition method based on double eye fundus images provided according to embodiments of the present invention obtains eyes eyeground figure first
Picture obtains the feature vector of two eye fundus images by machine learning model model1 respectively, is reflected with feature vector each
The feature of image;Then feature vector is combined, so as to get combination of eigenvectors integrate two eye fundus images the characteristics of;
By the two feature vectors combination input machine learning model model2 corresponding to two eye fundus images, make machine learning model
Model2 is combined according to two feature vectors and is identified the semantic feature that two eye fundus images are included so that identification process is no longer
It is identified according to simple eye eye fundus image, and two eye fundus images of combination pair of feature vector is utilized to establish contact, make engineering
The relevance that model considers two eye fundus images itself is practised, it is possible thereby to improve the accuracy of recognition result.
The embodiment of the present invention also provides a kind of characteristic identificating equipment based on more images, including:At least one processor;With
And the memory being connect at least one processor communication;Wherein, be stored with can be by one processing for the memory
The instruction that device executes, described instruction are executed by least one processor, so that at least one processor execution is above-mentioned
Characteristic recognition method based on more images.
One embodiment of the present of invention additionally provides a kind of more characteristics of image identification model training methods, as shown in figure 3, should
Method includes the following steps:
S01, obtains multiple images and its corresponding characteristic information, plurality of image correspond to identical characteristic information,
Such as image sequence i1 ... in, they correspond to a feature F or they both correspond to multiple feature F1 ... Fn;
S02 obtains the feature vector of multiple images, and for details, reference can be made to corresponding steps in above-mentioned recognition methods;
S03 carries out multiple combinations to the feature vector of acquisition, and for details, reference can be made to corresponding steps in above-mentioned recognition methods;
S04 determines that sample data, sample data are the corresponding combination of eigenvectors of image and characteristic information, wherein different
The combination of the corresponding feature vector of image differ.For a sample data, itself, which has comprising one, to close
Various combination of eigenvectors corresponding to the image sequence of connection property and one or more features information, such as label information.Assuming that
The characteristic information of a certain image sequence is unique, then a sample data itself includes then multiple combination of eigenvectors-labels
The correspondence group of information.
S05 is trained machine learning model using multiple sample datas, so that machine learning model is according to image pair
The combination of eigenvectors answered identifies characteristic information.Using multiple sample datas as described in step S04 to the same engineering
It practises model to be trained, makes the correspondence of its learning characteristic vector combination and label information, so as to according to multiple features
Vector combination identifies label information.
The model training method provided according to embodiments of the present invention obtains the sample image sequence with relevance first,
And then the feature vector of each image is obtained respectively, the feature of each image is reflected with feature vector, and feature vector is carried out
Combination, so as to get combination of eigenvectors synthesis multiple images the characteristics of, as the training data of machine learning model, make
Machine learning model learning characteristic vector combines the correspondence with image sequence characteristic information, thus obtained machine learning mould
Type utilizes the data of this reaction image relevance of the combination of feature vector to know not further according to single image identification feature information
Other characteristic information, it is possible thereby to improve the accuracy of Model Identification result.
One embodiment of the present of invention additionally provides another more characteristics of image identification model training methods, as shown in figure 4,
This method comprises the following steps:
S ' 01, obtains multiple sample datas, and each sample data includes multiple images and its corresponding characteristic information, and
And the multiple images in each sample data correspond to identical characteristic information.For each sample data, have comprising image
Sequence i1 ... in, and they correspond to a feature F or they both correspond to multiple feature F1 ... Fn;
Multiple sample datas are input in the first sort module of machine learning model by S ' 02, make the first sort module
Characteristic information is identified according to multiple images.First sort module itself is a machine learning model, it learns sample respectively
The correspondence of each image and corresponding characteristic information in notebook data.
S ' 03 obtains the first sort module when identifying each image, the feature vector of one of middle layer output,
In study and identification process, the hidden layer of the first sort module can export feature vector, this step can be chosen last in hidden layer
The feature vector of one layer of output.
S ' 04 carries out multiple combinations to the feature vector of acquisition, and for details, reference can be made to corresponding steps in above-mentioned recognition methods;
The corresponding feature vector of multiple images and characteristic information are input to the second classification mould of machine learning model by S ' 05
In block, the second sort module is made to identify characteristic information according to the combination of multiple images and its corresponding feature vector.Second point
Generic module itself is a machine learning model, different from the first sort module, and the study of the second sort module is each image
Correspondence between the combination and characteristic information of corresponding feature vector.By learning to a large amount of sample data, from
And characteristic information (label information) can be identified according to multiple combination of eigenvectors.
The model training method provided according to embodiments of the present invention, by sample image sequence and its correspondence with relevance
Characteristic information form sample data, learn the correspondence of image and characteristic information using the first sort module, and learning
Extraction feature vector in the process;Then feature vector is combined, so as to get combination of eigenvectors synthesis multiple images
Feature makes the combination of the second sort module learning characteristic vector and image sequence as the input data of the second sort module
The correspondence of characteristic information, thus obtained machine learning model utilize not further according to single image identification feature information
The data identification feature information of this reaction image relevance of combination of feature vector, it is possible thereby to improve Model Identification result
Accuracy.And the above method trains two sort modules simultaneously using the training method of end-end, and there is higher training to imitate
Rate.
It should be understood by those skilled in the art that, the embodiment of the present invention can be provided as method, system or computer program
Product.Therefore, complete hardware embodiment, complete software embodiment or reality combining software and hardware aspects can be used in the present invention
Apply the form of example.Moreover, the present invention can be used in one or more wherein include computer usable program code computer
The computer program production implemented in usable storage medium (including but not limited to magnetic disk storage, CD-ROM, optical memory etc.)
The form of product.
The present invention be with reference to according to the method for the embodiment of the present invention, the flow of equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that can be realized by computer program instructions every first-class in flowchart and/or the block diagram
The combination of flow and/or box in journey and/or box and flowchart and/or the block diagram.These computer programs can be provided
Instruct the processor of all-purpose computer, special purpose computer, Embedded Processor or other programmable data processing devices to produce
A raw machine so that the instruction executed by computer or the processor of other programmable data processing devices is generated for real
The device for the function of being specified in present one flow of flow chart or one box of multiple flows and/or block diagram or multiple boxes.
These computer program instructions, which may also be stored in, can guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works so that instruction generation stored in the computer readable memory includes referring to
Enable the manufacture of device, the command device realize in one flow of flow chart or multiple flows and/or one box of block diagram or
The function of being specified in multiple boxes.
These computer program instructions also can be loaded onto a computer or other programmable data processing device so that count
Series of operation steps are executed on calculation machine or other programmable devices to generate computer implemented processing, in computer or
The instruction executed on other programmable devices is provided for realizing in one flow of flow chart or multiple flows and/or block diagram one
The step of function of being specified in a box or multiple boxes.
Obviously, the above embodiments are merely examples for clarifying the description, and does not limit the embodiments.It is right
For those of ordinary skill in the art, can also make on the basis of the above description it is other it is various forms of variation or
It changes.There is no necessity and possibility to exhaust all the enbodiments.And it is extended from this it is obvious variation or
It changes still within the protection scope of the invention.
Claims (10)
1. a kind of characteristic recognition method based on more images, which is characterized in that including:
Multiple images are obtained, described multiple images are the figures with relevance for determining at least one identical characteristic information
As sequence;
The feature vector of described multiple images is obtained respectively;
Multiple combinations are carried out to the feature vector of acquisition;
The combination of the corresponding feature vector of described multiple images is identified to obtain image using the first machine learning model
Characteristic information, wherein the combination of the corresponding feature vector of different images differs, first machine learning model is to utilize
What sampling feature vectors combination and corresponding characteristic information were trained.
2. according to the method described in claim 1, it is characterized in that, it is described respectively obtain described multiple images feature vector,
Including:
Described multiple images are inputted into the second machine learning model respectively, make second machine learning model respectively to described more
A image is identified, wherein second machine learning model is trained using sample image and corresponding characteristic information
It obtains;
Second machine learning model is obtained when identifying each described image, the feature of one of middle layer output to
Amount.
3. according to the method described in claim 2, it is characterized in that, the middle layer is in second machine learning model
Last layer in interbed.
4. according to the method described in claim 1, it is characterized in that, the combination of described eigenvector is by different at least two
Feature vector carries out end to end.
5. according to the method described in claim 4, it is characterized in that, generate each combination of eigenvectors separately include it is all
Feature vector.
6. according to the method described in claim 4, it is characterized in that, the feature vector of described pair of acquisition carries out multiple combinations, packet
It includes:
According to the correspondence of scheduled combination of eigenvectors and image sequence, selected from whole modes of combination of eigenvectors
Identical as described image quantity and corresponding combination of eigenvectors mode;
Feature vector is combined according to selected combination of eigenvectors mode.
7. according to the method described in any one of claim 1-6, which is characterized in that described multiple images are eye fundus image, institute
The quantity for stating multiple images is 2, respectively left eye eye fundus image and right eye eye fundus image, and described image characteristic information is lesion
Information.
8. a kind of more characteristics of image identification model training methods, which is characterized in that including:
It obtains multiple images and its corresponding characteristic information, wherein described multiple images corresponds to identical characteristic information;
Obtain the feature vector of described multiple images;
Multiple combinations are carried out to the feature vector of acquisition;
Determining sample data, the sample data is the corresponding combination of eigenvectors of described image and the characteristic information, wherein
The combination of the corresponding feature vector of different images differs;
Machine learning model is trained using multiple sample datas, so that the machine learning model is according to the figure
As corresponding combination of eigenvectors identifies the characteristic information.
9. a kind of more characteristics of image identification model training methods, which is characterized in that including:
Multiple sample datas are obtained, each sample data includes multiple images and its corresponding characteristic information, and every
Multiple images in a sample data correspond to identical characteristic information;
The multiple sample data is input in the first sort module of machine learning model, the first sort module root is made
The characteristic information is identified according to described multiple images;
First sort module is obtained when identifying each described image, the feature vector of one of middle layer output;
Multiple combinations are carried out to the feature vector of acquisition;
The corresponding feature vector of described multiple images is input in the second sort module of the machine learning model, is made described
Second sort module identifies the characteristic information according to the combination of the corresponding feature vector of described multiple images.
10. a kind of characteristic identificating equipment based on more images, which is characterized in that including:At least one processor;And with it is described
The memory of at least one processor communication connection;Wherein, the memory is stored with and can be executed by one processor
Instruction, described instruction are executed by least one processor, so that at least one processor executes such as claim 1-7
Any one of described in the characteristic recognition method based on more images.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810470392.6A CN108734209A (en) | 2018-05-16 | 2018-05-16 | Feature recognition based on more images and equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810470392.6A CN108734209A (en) | 2018-05-16 | 2018-05-16 | Feature recognition based on more images and equipment |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108734209A true CN108734209A (en) | 2018-11-02 |
Family
ID=63938382
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810470392.6A Pending CN108734209A (en) | 2018-05-16 | 2018-05-16 | Feature recognition based on more images and equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108734209A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111143693A (en) * | 2019-12-31 | 2020-05-12 | 腾讯科技(深圳)有限公司 | Artificial intelligence-based feature processing model training method and device |
WO2020093563A1 (en) * | 2018-11-07 | 2020-05-14 | 哈尔滨工业大学(深圳) | Medical image processing method, system, device, and storage medium |
CN111180011A (en) * | 2019-12-31 | 2020-05-19 | 上海依智医疗技术有限公司 | Method and device for detecting gene mutation of focus |
CN111382623A (en) * | 2018-12-28 | 2020-07-07 | 广州市百果园信息技术有限公司 | Live broadcast auditing method, device, server and storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7697752B2 (en) * | 2005-12-20 | 2010-04-13 | General Instrument Corporation | Method and apparatus for performing object detection |
CN105893936A (en) * | 2016-03-28 | 2016-08-24 | 浙江工业大学 | Behavior identification method base on fusion of HOIRM and local feature |
CN106203448A (en) * | 2016-07-08 | 2016-12-07 | 南京信息工程大学 | A kind of scene classification method based on Nonlinear Scale Space Theory |
CN106599804A (en) * | 2016-11-30 | 2017-04-26 | 哈尔滨工业大学 | Retina fovea centralis detection method based on multi-feature model |
-
2018
- 2018-05-16 CN CN201810470392.6A patent/CN108734209A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7697752B2 (en) * | 2005-12-20 | 2010-04-13 | General Instrument Corporation | Method and apparatus for performing object detection |
CN105893936A (en) * | 2016-03-28 | 2016-08-24 | 浙江工业大学 | Behavior identification method base on fusion of HOIRM and local feature |
CN106203448A (en) * | 2016-07-08 | 2016-12-07 | 南京信息工程大学 | A kind of scene classification method based on Nonlinear Scale Space Theory |
CN106599804A (en) * | 2016-11-30 | 2017-04-26 | 哈尔滨工业大学 | Retina fovea centralis detection method based on multi-feature model |
Non-Patent Citations (2)
Title |
---|
罗院生: "从视网膜图像到糖尿病视网膜病变诊断", 《中国优秀硕士学位论文全文数据库 医药卫生科技辑》 * |
黄潇 等: "人工智能糖网眼底图像识别在真实世界的应用", 《情报工程》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020093563A1 (en) * | 2018-11-07 | 2020-05-14 | 哈尔滨工业大学(深圳) | Medical image processing method, system, device, and storage medium |
CN111382623A (en) * | 2018-12-28 | 2020-07-07 | 广州市百果园信息技术有限公司 | Live broadcast auditing method, device, server and storage medium |
CN111143693A (en) * | 2019-12-31 | 2020-05-12 | 腾讯科技(深圳)有限公司 | Artificial intelligence-based feature processing model training method and device |
CN111180011A (en) * | 2019-12-31 | 2020-05-19 | 上海依智医疗技术有限公司 | Method and device for detecting gene mutation of focus |
CN111143693B (en) * | 2019-12-31 | 2024-04-30 | 腾讯科技(深圳)有限公司 | Training method and device for feature processing model based on artificial intelligence |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Yang et al. | An emotion recognition model based on facial recognition in virtual learning environment | |
CN108734209A (en) | Feature recognition based on more images and equipment | |
US10853623B2 (en) | Method and apparatus for generating information | |
Richard et al. | Weakly supervised action learning with rnn based fine-to-coarse modeling | |
CN108090470B (en) | Face alignment method and device | |
CN109740413A (en) | Pedestrian recognition methods, device, computer equipment and computer storage medium again | |
EP3327628A1 (en) | Method and apparatus for recognizing an object | |
CN110147836A (en) | Model training method, device, terminal and storage medium | |
EP3176563A1 (en) | Identification device and identification method | |
US20190087683A1 (en) | Method and apparatus for outputting information | |
CN109635676A (en) | A method of positioning source of sound from video | |
CN113361636B (en) | Image classification method, system, medium and electronic device | |
CN111160225A (en) | Human body analysis method and device based on deep learning | |
CN111444370A (en) | Image retrieval method, device, equipment and storage medium thereof | |
CN110751069A (en) | Face living body detection method and device | |
CN110175505A (en) | Determination method, apparatus, storage medium and the electronic device of micro- expression type | |
CN110909565B (en) | Image recognition and pedestrian re-recognition method and device, electronic and storage equipment | |
CN108614987A (en) | The method, apparatus and robot of data processing | |
CN108154183A (en) | A kind of objective classification method based on part and depth characteristic set | |
CN114140880A (en) | Gait recognition method and device | |
Pigou et al. | Sign classification in sign language corpora with deep neural networks | |
CN109255339A (en) | Classification method based on adaptive depth forest body gait energy diagram | |
KR102562377B1 (en) | Method and system for federated learning of artificial intelligence for providing diagnosis information of depression | |
CN112200862B (en) | Training method of target detection model, target detection method and device | |
CN113205045A (en) | Pedestrian re-identification method and device, electronic equipment and readable storage medium |
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: 20181102 |