CN109902805A - The depth measure study of adaptive sample synthesis and device - Google Patents
The depth measure study of adaptive sample synthesis and device Download PDFInfo
- Publication number
- CN109902805A CN109902805A CN201910133351.2A CN201910133351A CN109902805A CN 109902805 A CN109902805 A CN 109902805A CN 201910133351 A CN201910133351 A CN 201910133351A CN 109902805 A CN109902805 A CN 109902805A
- Authority
- CN
- China
- Prior art keywords
- sample
- network
- difficulty
- depth
- adaptive
- 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
Landscapes
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the depth measure learning methods and device of a kind of adaptive sample synthesis, wherein method includes the following steps: extracting the depth characteristic of image from training set, and generates sample multi-component system；Sample difficulty is adaptively adjusted by linear interpolation；The sample for keeping difficulty is generated by generator；According to keeping the sample acquisition of difficulty to generate expression information of the difficult sample under metric space, while training whole network, information is indicated to obtain picture under metric space.This method uses generator, convert original sample characteristics to the feature for having information to the training of measurement, and the difficulty for controlling generation sample adaptively trains measurement network, and then the adaptive generation device module in most depth measure learning models can be integrated, realize the raising to depth measure learning method performance.
Description
Technical field
The present invention relates to computer vision and machine learning techniques field, in particular to a kind of depth of adaptive sample synthesis
Spend metric learning method and device.
Background technique
Metric learning has consequence in computer vision.The semantic information transmitted in image generally passes through feature
Vector is expressed, but the distance between feature vector generally can not be well reflected the true semantic distance between image, and big
Amount task such as image retrieval, face verification etc. all relies on the measurement to image similarity.Therefore, how to learn a standard out
Really, the distance metric of robust is an extremely crucial problem.
Metric learning in the related technology is broadly divided into two classes: linear and nonlinear metric study.Traditional linearity
Amount learning method mainly measures measurement between sample by learning a mahalanobis distance, such as LMNN (Large margin
Nearest neighbor, large-spacing nearest-neighbors), ITML (InformationTheoryMetricLearning, information reason
By study)；And nonlinear metric learning method is then carried out by kernel method or neural network method come the relationship to higher order
Modeling.From another point of view, metric learning method can be divided into non-supervisory and two classes of supervision again.Non-supervisory metric learning is mainly wished
Hope one mapping from higher-dimension to low-dimensional of study, and make retain as much as possible in higher-dimension in lower dimensional space between sample away from
From information.And the metric learning supervised then is based primarily upon a target, that is, maximizes between class distance and minimize inter- object distance.
Different supervision metric learning methods, is essentially all that the difference of the target is portrayed and modeled.As deep neural network exists
Huge success is achieved in image recognition, the current main approaches of metric learning also turn to depth network.Most of depths
Degree metric learning method is also regarded as depth insertion study, they are constructed empty from image space to insertion using deep neural network
Between mapping so that the Euclidean distance in the space is able to reflect the true semantic distance of image data.Depth measure
Practise substantially frame and key step it is as follows: 1) projected depth network model, by picture or video mention conversion be characterized it is embedding with it
Enter；2) multiple sample multi-component systems etc. are constructed from training set；3) by minimize the loss function defined in multi-component system etc. come
Training depth network；4) final distance metric is the Euclidean distance under depth internet startup disk space.As Hu et al. is based on
Sample measures the recognition of face being used under natural conditions with judgement index to having trained one.Song et al. is by calculating one
The distance between training sample matrix is criticized, a kind of optimization aim that can make full use of information in a collection of sample is devised, three
Good result is achieved on a image recognition library.Wang et al. is by limiting the triangle being made of in each training set triple
The angle of negative sample apex carrys out design optimization target in shape, and three rank geological informations between sample are utilized, further increase
The performance of metric learning method based on depth network.
An effective measurement is accurate, robustly measures two samples to learn for the target of depth measure learning method
Similarity between data, and the optimization aim of depth measure learning model entirety can be regarded as and be added using its sample
The loss function of power, this makes sampling policy become an important component part in its model.One about sampling policy
Main problem is a lack of the sample (commonly known as " hardly possible " sample) of information.Most of sample in training set may expire
The constraint that sufficient loss function applies, therefore effective supervision message cannot be transmitted.With trained progress, more and more samples
It will be classified as simple sample, thus lead to slower training speed.This allows for many depth measure learning methods to set
The efficient difficult sample Mining Strategy of meter.These strategy usually in training set sample carry out it is down-sampled, selection can provide compared with
The sample of big gradient, can both accelerate the convergence rate of model in this way, can also be improved the effect for finally learning measurement out.Such as
Schroff et al. selects the negative sample of " half is difficult " in a collection of sample, i.e., so that the distance of negative sample pair is smaller in triple, but
Still it is greater than distance between positive sample pair, the network based on triple loss function is trained with such triple.Wang etc.
People is using a kind of importance sampling method on line, the triple for selecting training to use using the similarity between sample.
The sample method for digging that Harwood et al. is searched for using approximate KNN, is adaptive selected the training sample of more challenge,
It can more efficiently training pattern.Yuan et al. proposes multiple to excavate by one group of Cascade with different complexity
The sample of grade of difficulty, and the model more complicated using the sample training being more difficult.
However, difficult sample Mining Strategy only selects sample in a subset of training set, and this set may be insufficient
Accurately to portray the whole geometry structure of metric space.That is, some data points are repeated as many times as required sampling, and other
Data point will not may be sampled always, and metric space over-fitting near the data point of over-sampling is caused, while owed again
Poor fitting near the data point of sampling.Since simple sample has generally taken up the major part in training set, such hardly possible sample is dug
Pick strategy, which is equivalent to, has ignored most data.The present invention is solved the above problems by the sample of synthesis self-adaptive difficulty.
The difficulty of synthesis sample should determine that the present invention controls synthesis sample according to the physical training condition of model according to the physical training condition of model
This difficulty, so that being trained to model of training when more abundant using more difficult synthesis sample.
In terms of the necessity of synthesis self-adaptive difficulty sample is following two.Firstly, at trained initial stage, metric space
And do not have accurate semantic structure, therefore current difficult sample may veritably not include significant information.This
In the case of, the difficult sample of synthesis may can be even some contradictory samples.In addition, difficult sample normally results in network parameter hair
Raw biggish variation, therefore be easy to destroy the structure of metric space using meaningless sample, cause one from the beginning
With the model of wrong way training.On the other hand, with trained progress, model can easily cope with the sample being more difficult,
Therefore the synthesis sample that is increasingly difficult to should be generated to maintain to keep high learning efficiency.
Summary of the invention
The present invention is directed to solve at least some of the technical problems in related technologies.
For this purpose, an object of the present invention is to provide a kind of depth measure learning method of adaptive sample synthesis, it should
Method can preferably train network, be integrally trained to network by using unified majorized function, finally learn
One accurate, robust measurement.
It is another object of the present invention to the depth measure learning devices for proposing a kind of adaptive sample synthesis.
In order to achieve the above objectives, one aspect of the present invention proposes the depth measure learning method of adaptive sample synthesis, packet
Include following steps: step S1 extracts the depth characteristic of image from training set, and generates sample multi-component system；Step S2, passes through
Linear interpolation adaptively adjusts sample difficulty；Step S3 generates the sample for keeping difficulty by generator；Step S4, according to institute
It states and the sample acquisition of difficulty is kept to generate expression information of the difficult sample under metric space, while training whole network, to obtain
Expression information of the picture under metric space.
The depth measure learning method of the adaptive sample synthesis of the embodiment of the present invention, by neural network by original sample
Eigen is converted into the sample of information, is supplemented existing difficult sample, to make full use of the information of whole samples；
Meanwhile the difficulty for generating sample is controlled by linear interpolation, and by the interaction with measurement network, generation is suitable for
The difficult sample of model state in the related technology, and integrate the adaptive generation device mould in depth measure learning model in the related technology
Block, the final raising realized to depth measure learning method performance.
In addition, the depth measure learning method of adaptive sample synthesis according to the above embodiment of the present invention can also have
Additional technical characteristic below:
Further, in one embodiment of the invention, the step S1 includes: by the training set partial image forward pass
Enter depth convolutional neural networks, and after the last full articulamentum by feature extraction network, obtains the depth of 1024 dimensions of image
Feature is spent, and triple or multi-component system are obtained according to the objective function that sample label and measurement use.
Further, in one embodiment of the invention, the step S2 includes: under metric space, to negative sample
The sample difficulty is adjusted to by changing distance, wherein the target sample difficulty is obtained according to measurement model performance；According to
Current loss function value measures the current state of measurement model, and controls the distance for generating sample.
Further, in one embodiment of the invention, the whole loss function of the generator are as follows:
Wherein, λ is balance factor, and y is sample characteristics, and y' is the composite character for not adjusting length,After adjustment length
Composite character, Y, Y' andFeature distribution respectively corresponding to them, l are the corresponding class label of sample, and c (y, y') is
Reconstruct loss, JsoftFor Softmax loss function.
Further, in one embodiment of the invention, the step S4 includes: by the difficult sample degree of passing through of the generation
It measures network and obtains the expression information under metric space；Feature extraction network, measurement network and generator are connected to the network,
With the whole network after being trained；Test specimens are obtained by the feature extraction network and the measurement network by samples pictures
This expression information under metric space.
In order to achieve the above objectives, another aspect of the present invention proposes a kind of depth measure study dress of adaptive sample synthesis
It sets, comprising: extraction module is used to extract the depth characteristic of image from training set, and generates sample multi-component system；Adjust module
For adaptively adjusting sample difficulty by linear interpolation；Generation module is used to generate the sample for keeping difficulty by generator；
Training module is used to generate expression information of the difficult sample under metric space according to the sample acquisition for keeping difficulty, instructs simultaneously
Practice whole network, indicates information to obtain picture under metric space.
The depth measure learning device of the adaptive sample synthesis of the embodiment of the present invention, by neural network by original sample
Eigen is converted into the sample of information, is supplemented existing difficult sample, to make full use of the information of whole samples；
Meanwhile the difficulty for generating sample is controlled by linear interpolation, and by the interaction with measurement network, generation is suitable for
The difficult sample of model state in the related technology, and integrate the adaptive generation device mould in depth measure learning model in the related technology
Block, the final raising realized to depth measure learning method performance.
In addition, the depth measure learning device of adaptive sample synthesis according to the above embodiment of the present invention can also have
Additional technical characteristic below:
Further, in one embodiment of the invention, the extraction module is further used for: by the training set point
It is passed to depth convolutional neural networks before image, and after the last full articulamentum by feature extraction network, obtains image
The depth characteristic of 1024 dimensions, and triple or multi-component system are obtained according to the objective function that sample label and measurement use.
Further, in one embodiment of the invention, the adjustment module is specifically used for: right under metric space
Negative sample adjusts the sample difficulty to by changing distance, wherein it is difficult to obtain the target sample according to measurement model performance
Degree；The current state of measurement model is measured according to current loss function value, and controls the distance for generating sample.
Further, in one embodiment of the invention, the whole loss function of the generator are as follows:
Wherein, λ is balance factor, and y is sample characteristics, and y' is the composite character for not adjusting length,After adjustment length
Composite character, Y, Y' andFeature distribution respectively corresponding to them, l are the corresponding class label of sample, and c (y, y') is
Reconstruct loss, JsoftFor Softmax loss function.
Further, in one embodiment of the invention, the training module is specifically used for: by the difficult sample of the generation
The expression information under metric space is obtained by measurement network；By feature extraction network, measurement network and generator net
Network connection, with the whole network after being trained；It is obtained by samples pictures by the feature extraction network and the measurement network
To the expression information under metric space of test sample.
The additional aspect of the present invention and advantage will be set forth in part in the description, and will partially become from the following description
Obviously, or practice through the invention is recognized.
Detailed description of the invention
Above-mentioned and/or additional aspect and advantage of the invention will become from the following description of the accompanying drawings of embodiments
Obviously and it is readily appreciated that, in which:
Fig. 1 is the depth measure learning method flow chart synthesized according to the adaptive sample of the embodiment of the present invention；
Fig. 2 is the specific of step S4 in the depth measure learning method synthesized according to the adaptive sample of the embodiment of the present invention
Network structure；
Fig. 3 is the depth measure learning device structural schematic diagram synthesized according to the adaptive sample of the embodiment of the present invention.
Specific embodiment
The embodiment of the present invention is described below in detail, examples of the embodiments are shown in the accompanying drawings, wherein from beginning to end
Same or similar label indicates same or similar element or element with the same or similar functions.Below with reference to attached
The embodiment of figure description is exemplary, it is intended to is used to explain the present invention, and is not considered as limiting the invention.
The depth measure study side of the adaptive sample proposed according to embodiments of the present invention synthesis is described with reference to the accompanying drawings
Method and device describe the depth measure study of the adaptive sample proposed according to embodiments of the present invention synthesis with reference to the accompanying drawings first
Method.
Fig. 1 is the depth measure learning method flow chart of the adaptive sample synthesis of one embodiment of the invention.
As shown in Figure 1, the depth measure learning method of adaptive sample synthesis the following steps are included:
In step sl, the depth characteristic of image is extracted from training set, and generates sample multi-component system.
Specifically, the embodiment of the present invention will be passed to depth convolutional neural networks before training set partial image, and pass through feature
After the last full articulamentum for extracting network, the depth characteristic of 1024 dimensions of image is obtained, and is made according to sample label and measurement
Objective function obtains triple or multi-component system.
In step s 2, sample difficulty is adaptively adjusted by linear interpolation.
Further, in one embodiment of the invention, step S2 includes: under metric space, to negative sample to logical
Cross change distance adjustment sample difficulty, wherein target sample difficulty is obtained according to measurement model performance；According to current loss letter
Numerical value measures the current state of measurement model, and controls the distance for generating sample.
Briefly, under metric space for negative sample pair, can by changing its distance come adjustment difficulty, reduce its away from
From being equivalent to raising difficulty.Therefore, it is possible to use linear interpolation controls the difficulty of negative sample pair.With trained progress, by
In the promotion of measurement model performance, the bigger sample of Ying Shengcheng difficulty.The embodiment of the present invention is come using current loss function value
The state of measurement model is measured, and controls the distance for generating sample with this.
In step s3, the sample for keeping difficulty is generated by generator.
It should be noted that generator is made of multilayer neural network, need to introduce loss function to train.Pass through life
It grows up to be a useful person and wishes the sample of the adaptive difficulty Jing Guo linear interpolation being mapped to feature space from metric space, so as to be used to
Training measurement network.The feature of generation should have the property that
1) difficulty is kept as much as possible
The sample of input generator is the sample of the adaptive difficulty generated by linear interpolation, and difficulty is by current degree
Amount model state is determined, it is thus possible to improve the training effectiveness of measurement network to the maximum extent.Thus, the sample characteristics of generation
Also such difficulty should be kept as much as possible.
2) with original sample distribution having the same
Final measurement network should be able to be fitted true data set, therefore the sample generated should be with script with identical
Distribution, thus keep its authenticity.
3) classification of original sample is not changed
For the metric learning of supervision, need to calculate loss function using the classification of sample.The embodiment of the present invention is direct
The classification of sample is generated using the classification of original sample, therefore the process generated allows for keeping the classification of sample, this
Sample just can guarantee that the adaptive sample of generation can train measurement network well.
Further, the whole loss function of generator are as follows:
Wherein, λ is balance factor, and y is sample characteristics, and y' is the composite character for not adjusting length,After adjustment length
Composite character, Y, Y' andFeature distribution respectively corresponding to them, l are the corresponding class label of sample, and c (y, y') is
Reconstruct loss, JsoftFor Softmax loss function.
In step s 4, according to keeping the sample acquisition of difficulty to generate expression information of the difficult sample under metric space, together
Shi Xunlian whole network indicates information to obtain picture under metric space.
Further, in one embodiment of the invention, step S4 includes: and will generate difficult sample to obtain by measuring network
Take the expression information under metric space；Feature extraction network, measurement network and generator are connected to the network, after being trained
Whole network；The table under metric space of test sample is obtained by feature extraction network and measurement network by samples pictures
Show information.
Specifically, as shown in Fig. 2, together by feature extraction network, measurement network and generator network connection.Measurement
Network is both using the sample optimization generated:
In formula, β is balance factor, JgenFor generator loss function, T is sample multi-component system,It is more for what is adaptively synthesized
Tuple, J are measurement loss function.
After training whole network, the expression under metric space of test sample passes through feature extraction by samples pictures
Network is obtained with measurement network, without being only used in the training process of network by generator.
To sum up, the embodiment of the present invention is to avoid the difficult sample digging technology that depth measure learning method uses in the related technology
A large amount of easy a small amount of training samples that sample is ignored and it is used can not be portrayed with the distribution of sample in training set well
Problem, therefore propose the depth measure learning method of adaptive sample synthesis, by utilizing the generator of multilayer neural network composition,
The sample that will do not have helpful sample to be adaptively converted into information training originally.Firstly, to extracting from picture
Initial characteristics carry out the feature for the adaptive difficulty that linear interpolation is synthesized, and are entered into generator, are translated into
New feature, while keeping its difficulty and classification.Secondly, obtaining final number by measurement network with these new features
According to expression, and by simultaneously training by acting on generator and acting on the optimization aim dimerous of measurement network
Whole network.After the completion of network training, for the test sample of input, not by it by generator, and it is direct to measure network
Initial characteristics are acted on to obtain its expression in the case where learning metric space out.
For in another way, its initial feature is extracted using depth network for picture；Pass through linear interpolation later
It is adaptively adjusted the difficulty of feature；The generator using multilayer neural network is reused, converts initial characteristics to new spy
Sign guarantees not change the classification for generating feature, last measurement optimization aim is for life while keeping adaptive difficulty
The feature of generation of growing up to be a useful person also has good resolution performance.
The depth measure learning method of the adaptive sample synthesis proposed according to embodiments of the present invention, will by neural network
Original sample characteristics are converted into the sample of information, are supplemented existing difficult sample, to make full use of whole samples
This information；Meanwhile controlled by linear interpolation generate sample difficulty, and by with measurement network interaction,
The difficult sample for being suitable for model state in the related technology is generated, and is integrated adaptive in depth measure learning model in the related technology
Maker module is answered, the final raising realized to depth measure learning method performance.
Referring next to the depth measure study dress for the adaptive sample synthesis that attached drawing description proposes according to embodiments of the present invention
It sets.
Fig. 2 is the depth measure learning structure schematic diagram of the adaptive sample synthesis of one embodiment of the invention.
As shown in Fig. 2, the depth measure learning device 10 of the adaptive sample synthesis includes: extraction module 100, adjustment mould
Block 200, generation module 300 and training module 400.
Wherein, extraction module 100 is used to extract the depth characteristic of image from training set, and generates sample multi-component system.
Further, in one embodiment of the invention, extraction module 100 is further used for: by training set partial image
Preceding incoming depth convolutional neural networks, and after the last full articulamentum by feature extraction network, obtain 1024 dimensions of image
Depth characteristic, and triple or multi-component system are obtained according to sample label and the objective function that uses of measurement.
Adjustment module 200 is used to adaptively adjust sample difficulty by linear interpolation.
Further, in one embodiment of the invention, adjustment module 200 is specifically used for: under metric space, to negative
Sample is to by changing distance adjustment sample difficulty, wherein obtains target sample difficulty according to measurement model performance；According to current
Loss function value measure measurement model current state, and control generate sample distance.
Generation module 300 is used to generate the sample for keeping difficulty by generator.
Further, in one embodiment of the invention, the whole loss function of generator are as follows:
Wherein, λ is balance factor, and y is sample characteristics, and y' is the composite character for not adjusting length,After adjustment length
Composite character, Y, Y' andFeature distribution respectively corresponding to them, l are the corresponding class label of sample, and c (y, y') is
Reconstruct loss, JsoftFor Softmax loss function.
Training module 400 is used for according to the expression letter for keeping the difficult sample of sample acquisition generation of difficulty under metric space
Breath, while training whole network, indicate information to obtain picture under metric space.
Further, in one embodiment of the invention, training module 400 is specifically used for: will generate difficult sample and passes through
It measures network and obtains the expression information under metric space；Feature extraction network, measurement network and generator are connected to the network, with
Whole network after being trained；Measuring for test sample is obtained by feature extraction network and measurement network by samples pictures
Expression information under space.
It should be noted that the explanation of the aforementioned depth measure learning method embodiment to the synthesis of adaptive sample
Suitable for the device, details are not described herein again.
The depth measure learning device of the adaptive sample synthesis proposed according to embodiments of the present invention, will by neural network
Original sample characteristics are converted into the sample of information, are supplemented existing difficult sample, to make full use of whole samples
This information；Meanwhile controlled by linear interpolation generate sample difficulty, and by with measurement network interaction,
The difficult sample for being suitable for model state in the related technology is generated, and is integrated adaptive in depth measure learning model in the related technology
Maker module is answered, the final raising realized to depth measure learning method performance.
In addition, term " first ", " second " are used for descriptive purposes only and cannot be understood as indicating or suggesting relative importance
Or implicitly indicate the quantity of indicated technical characteristic.Define " first " as a result, the feature of " second " can be expressed or
Implicitly include at least one this feature.In the description of the present invention, the meaning of " plurality " is at least two, such as two, three
It is a etc., unless otherwise specifically defined.
In the present invention unless specifically defined or limited otherwise, term " installation ", " connected ", " connection ", " fixation " etc.
Term shall be understood in a broad sense, for example, it may be being fixedly connected, may be a detachable connection, or integral；It can be mechanical connect
It connects, is also possible to be electrically connected；It can be directly connected, can also can be in two elements indirectly connected through an intermediary
The interaction relationship of the connection in portion or two elements, unless otherwise restricted clearly.For those of ordinary skill in the art
For, the specific meanings of the above terms in the present invention can be understood according to specific conditions.
In the present invention unless specifically defined or limited otherwise, fisrt feature in the second feature " on " or " down " can be with
It is that the first and second features directly contact or the first and second features pass through intermediary mediate contact.Moreover, fisrt feature exists
Second feature " on ", " top " and " above " but fisrt feature be directly above or diagonally above the second feature, or be merely representative of
First feature horizontal height is higher than second feature.Fisrt feature can be under the second feature " below ", " below " and " below "
One feature is directly under or diagonally below the second feature, or is merely representative of first feature horizontal height less than second feature.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example
Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not
It must be directed to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be in office
It can be combined in any suitable manner in one or more embodiment or examples.In addition, without conflicting with each other, the skill of this field
Art personnel can tie the feature of different embodiments or examples described in this specification and different embodiments or examples
It closes and combines.
Although the embodiments of the present invention has been shown and described above, it is to be understood that above-described embodiment is example
Property, it is not considered as limiting the invention, those skilled in the art within the scope of the invention can be to above-mentioned
Embodiment is changed, modifies, replacement and variant.
Claims (10)
1. a kind of depth measure learning method of adaptive sample synthesis, which comprises the following steps:
Step S1, extracts the depth characteristic of image from training set, and generates sample multi-component system；
Step S2 adaptively adjusts sample difficulty by linear interpolation；
Step S3 generates the sample for keeping difficulty by generator；And
Step S4 generates expression information of the difficult sample under metric space according to the sample acquisition for keeping difficulty, instructs simultaneously
Practice whole network, indicates information to obtain picture under metric space.
2. the depth measure learning method of adaptive sample synthesis according to claim 1, which is characterized in that the step
S1 includes:
Depth convolutional neural networks will be passed to before the training set partial image, and in the last full connection for passing through feature extraction network
After layer, the depth characteristic of 1024 dimensions of image is obtained, and ternary is obtained according to the objective function that sample label and measurement use
Group or multi-component system.
3. the depth measure learning method of adaptive sample synthesis according to claim 1, which is characterized in that the step
S2 includes:
Under metric space, the sample difficulty is adjusted to by changing distance to negative sample, wherein according to measurement model performance
Obtain the target sample difficulty；
The current state of measurement model is measured according to current loss function value, and controls the distance for generating sample.
4. the depth measure learning method of adaptive sample synthesis according to claim 1, which is characterized in that the generation
The whole loss function of device are as follows:
Wherein, λ is balance factor, and y is sample characteristics, and y' is the composite character for not adjusting length,For the synthesis after adjustment length
Feature, Y, Y' andFeature distribution respectively corresponding to them, l are the corresponding class label of sample, and c (y, y') is reconstruct damage
It loses, JsoftFor Softmax loss function.
5. the depth measure learning method of adaptive sample synthesis according to claim 1, which is characterized in that the step
S4 includes:
By the expression information for generating difficult sample by measurement network acquisition under metric space；
By feature extraction network, measurement network and generator network connection, with the whole network after being trained；
By samples pictures by the feature extraction network and it is described measurement network obtain test sample under metric space
Indicate information.
6. a kind of depth measure learning device of adaptive sample synthesis characterized by comprising
Extraction module for extracting the depth characteristic of image from training set, and generates sample multi-component system；
Module is adjusted, for adaptively adjusting sample difficulty by linear interpolation；
Generation module, for generating the sample for keeping difficulty by generator；And
Training module, for generating expression information of the difficult sample under metric space according to the sample acquisition for keeping difficulty,
Training whole network simultaneously, indicates information to obtain picture under metric space.
7. the depth measure learning device of adaptive sample synthesis according to claim 6, which is characterized in that the extraction
Module is further used for:
Depth convolutional neural networks will be passed to before the training set partial image, and in the last full connection for passing through feature extraction network
After layer, the depth characteristic of 1024 dimensions of image is obtained, and ternary is obtained according to the objective function that sample label and measurement use
Group or multi-component system.
8. the depth measure learning device of adaptive sample synthesis according to claim 6, which is characterized in that the adjustment
Module is specifically used for:
Under metric space, the sample difficulty is adjusted to by changing distance to negative sample, wherein according to measurement model performance
Obtain the target sample difficulty；
The current state of measurement model is measured according to current loss function value, and controls the distance for generating sample.
9. the depth measure learning device of adaptive sample synthesis according to claim 6, which is characterized in that the generation
The whole loss function of device are as follows:
Wherein, λ is balance factor, and y is sample characteristics, and y' is the composite character for not adjusting length,For the synthesis after adjustment length
Feature, Y, Y' andFeature distribution respectively corresponding to them, l are the corresponding class label of sample, and c (y, y') is reconstruct damage
It loses, JsoftFor Softmax loss function.
10. the depth measure learning device of adaptive sample synthesis according to claim 6, which is characterized in that the instruction
Practice module to be specifically used for:
By the expression information for generating difficult sample by measurement network acquisition under metric space；
By feature extraction network, measurement network and generator network connection, with the whole network after being trained；
By samples pictures by the feature extraction network and it is described measurement network obtain test sample under metric space
Indicate information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910133351.2A CN109902805A (en) | 2019-02-22 | 2019-02-22 | The depth measure study of adaptive sample synthesis and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910133351.2A CN109902805A (en) | 2019-02-22 | 2019-02-22 | The depth measure study of adaptive sample synthesis and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109902805A true CN109902805A (en) | 2019-06-18 |
Family
ID=66945186
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910133351.2A Pending CN109902805A (en) | 2019-02-22 | 2019-02-22 | The depth measure study of adaptive sample synthesis and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109902805A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110674692A (en) * | 2019-08-23 | 2020-01-10 | 北京大学 | Target accurate retrieval method and system based on difficult sample generation |
| CN110956255A (en) * | 2019-11-26 | 2020-04-03 | 中国医学科学院肿瘤医院 | Difficult sample mining method and device, electronic equipment and computer readable storage medium |
| CN111582365A (en) * | 2020-05-06 | 2020-08-25 | 吉林大学 | Junk mail classification method based on sample difficulty |
| CN111914109A (en) * | 2020-07-29 | 2020-11-10 | 厦门大学 | Depth measurement learning-based garment retrieval technology |
| CN112016539B (en) * | 2020-10-29 | 2021-03-26 | 上海特金信息科技有限公司 | Signal identification method and device, electronic equipment and storage medium |
| CN113450776A (en) * | 2020-03-24 | 2021-09-28 | 合肥君正科技有限公司 | Data enhancement method and system for improving crying detection model effect of baby |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105760889A (en) * | 2016-03-01 | 2016-07-13 | 中国科学技术大学 | Efficient imbalanced data set classification method |
| CN107392259A (en) * | 2017-08-16 | 2017-11-24 | 北京京东尚科信息技术有限公司 | The method and apparatus for building unbalanced sample classification model |
| CN107451954A (en) * | 2017-05-23 | 2017-12-08 | 南京邮电大学 | Iterated pixel interpolation method based on image low-rank property |
| CN108304864A (en) * | 2018-01-17 | 2018-07-20 | 清华大学 | Depth fights metric learning method and device |
| US20180276499A1 (en) * | 2017-03-24 | 2018-09-27 | Disney Enterprises, Inc. | One shot color calibrated metric learning for object re-identification |
-
2019
- 2019-02-22 CN CN201910133351.2A patent/CN109902805A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105760889A (en) * | 2016-03-01 | 2016-07-13 | 中国科学技术大学 | Efficient imbalanced data set classification method |
| US20180276499A1 (en) * | 2017-03-24 | 2018-09-27 | Disney Enterprises, Inc. | One shot color calibrated metric learning for object re-identification |
| CN107451954A (en) * | 2017-05-23 | 2017-12-08 | 南京邮电大学 | Iterated pixel interpolation method based on image low-rank property |
| CN107392259A (en) * | 2017-08-16 | 2017-11-24 | 北京京东尚科信息技术有限公司 | The method and apparatus for building unbalanced sample classification model |
| CN108304864A (en) * | 2018-01-17 | 2018-07-20 | 清华大学 | Depth fights metric learning method and device |
Non-Patent Citations (2)
| Title |
|---|
| SEAN BELL 等: "Learning visual similarity for product design with convolutional netural networks", 《ACM TRANSACTION ON GRAPHICS》 * |
| 宫志晨: "模型空间中的时间序列分类算法及其在不平衡数据上的应用", 《中国优秀硕士学位论文全文数据库基础科学辑》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110674692A (en) * | 2019-08-23 | 2020-01-10 | 北京大学 | Target accurate retrieval method and system based on difficult sample generation |
| CN110956255A (en) * | 2019-11-26 | 2020-04-03 | 中国医学科学院肿瘤医院 | Difficult sample mining method and device, electronic equipment and computer readable storage medium |
| CN110956255B (en) * | 2019-11-26 | 2023-04-07 | 中国医学科学院肿瘤医院 | Difficult sample mining method and device, electronic equipment and computer readable storage medium |
| CN113450776A (en) * | 2020-03-24 | 2021-09-28 | 合肥君正科技有限公司 | Data enhancement method and system for improving crying detection model effect of baby |
| CN111582365A (en) * | 2020-05-06 | 2020-08-25 | 吉林大学 | Junk mail classification method based on sample difficulty |
| CN111914109A (en) * | 2020-07-29 | 2020-11-10 | 厦门大学 | Depth measurement learning-based garment retrieval technology |
| CN112016539B (en) * | 2020-10-29 | 2021-03-26 | 上海特金信息科技有限公司 | Signal identification method and device, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109902805A (en) | The depth measure study of adaptive sample synthesis and device | |
| CN105787458B (en) | The infrared behavior recognition methods adaptively merged based on artificial design features and deep learning feature | |
| CN110287849A (en) | A kind of lightweight depth network image object detection method suitable for raspberry pie | |
| CN109299216A (en) | A kind of cross-module state Hash search method and system merging supervision message | |
| CN107423700A (en) | The method and device of testimony verification | |
| WO2021208601A1 (en) | Artificial-intelligence-based image processing method and apparatus, and device and storage medium | |
| CN108491874A (en) | A kind of image list sorting technique for fighting network based on production | |
| CN108304864A (en) | Depth fights metric learning method and device | |
| CN109344759A (en) | A kind of relatives' recognition methods based on angle loss neural network | |
| CN107123415A (en) | A kind of automatic music method and system | |
| CN108596203A (en) | Optimization method of the pond layer in parallel to pantograph carbon slide surface abrasion detection model | |
| CN113837366A (en) | Multi-style font generation method | |
| CN110427989A (en) | Chinese character skeleton automatic synthesis method and extensive Chinese word library automatic generation method | |
| CN109063649A (en) | Pedestrian's recognition methods again of residual error network is aligned based on twin pedestrian | |
| CN110457515A (en) | The method for searching three-dimension model of the multi-angle of view neural network of polymerization is captured based on global characteristics | |
| CN110287806A (en) | A kind of traffic sign recognition method based on improvement SSD network | |
| CN109271546A (en) | The foundation of image retrieval Feature Selection Model, Database and search method | |
| CN108491863A (en) | Color image processing method based on Non-negative Matrix Factorization and convolutional neural networks | |
| CN114663962A (en) | Lip-shaped synchronous face forgery generation method and system based on image completion | |
| CN114283083B (en) | Aesthetic enhancement method of scene generation model based on decoupling representation | |
| CN115564801A (en) | Attention-based single target tracking method | |
| CN111506760A (en) | Depth integration measurement image retrieval method based on difficult perception | |
| CN114357221A (en) | Self-supervision active learning method based on image classification | |
| CN113240594B (en) | High-quality point cloud completion method based on multiple hierarchies | |
| CN104778272A (en) | Method for estimating positions of images based on area mining and space encoding |
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: 20190618 |