CN110766065A - Hash learning method based on deep hyper-information - Google Patents

Hash learning method based on deep hyper-information Download PDF

Info

Publication number
CN110766065A
CN110766065A CN201910992620.0A CN201910992620A CN110766065A CN 110766065 A CN110766065 A CN 110766065A CN 201910992620 A CN201910992620 A CN 201910992620A CN 110766065 A CN110766065 A CN 110766065A
Authority
CN
China
Prior art keywords
hash
information
hyper
deep
learner
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
Application number
CN201910992620.0A
Other languages
Chinese (zh)
Inventor
于治楼
袭肖明
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shandong Inspur Artificial Intelligence Research Institute Co Ltd
Original Assignee
Shandong Inspur Artificial Intelligence Research Institute Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shandong Inspur Artificial Intelligence Research Institute Co Ltd filed Critical Shandong Inspur Artificial Intelligence Research Institute Co Ltd
Priority to CN201910992620.0A priority Critical patent/CN110766065A/en
Publication of CN110766065A publication Critical patent/CN110766065A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to the field of computer vision, and particularly provides a hash learning method based on deep hyper-information. Compared with the prior art, the deep hyper-information-based Hash learning method mainly comprises a training stage and a testing stage, wherein the training stage comprises the construction of a physical Hash learning device and the construction of a correlation multidimensional Hash code learning device, and the testing stage comprises an acquisition stage and a matching stage. The obtained hyper-information hash code is utilized to enable people to understand the physical significance of the hyper-information hash code more easily, the complex hash code obtaining process is avoided, the recognition precision and efficiency are improved, and the method has good popularization value.

Description

Hash learning method based on deep hyper-information
Technical Field
The invention relates to the field of computer vision, and particularly provides a hash learning method based on deep hyper-information.
Background
Hashing is the conversion of an input of arbitrary length, also called pre-mapping, into a fixed length output, the output being a hash value, by a hashing algorithm. This transformation is a compression mapping, i.e., the space of hash values is usually much smaller than the space of inputs, different inputs may hash to the same output, and so it is not possible to uniquely determine the input value from the hash value. In short, it is a function of compressing a message tooth of an arbitrary length to a message digest of a fixed length.
The hash has become an important data representation form in the big data era due to the advantages of the hash in storage and calculation. The traditional Hash learning method has no physical significance which is easy to be understood by people, and the interpretability is poor. And a cumbersome hash function needs to be designed to obtain each hash code. How to effectively solve the problems of poor hash understandability, complex acquisition process and the like in the prior art, and has important research significance and application value.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides the Hash learning method based on the depth super information, which has the advantages of reasonable design, safety, applicability and strong practicability.
The technical scheme adopted by the invention for solving the technical problems is as follows:
a Hash learning method based on deep hyper-information mainly comprises a training stage and a testing stage, and comprises the following steps:
(I) training phase
(1) Constructing a physical hash learning device: performing network architecture training based on the distinguishing characteristics of the identified objects, and constructing a physical hash learning device;
(2) constructing a correlation multidimensional hash code learner: constructing a correlation multidimensional hash code learner by utilizing a deep neural network;
(II) test phase
(1) An acquisition stage; and inputting the tested image into a trained learner, and acquiring the hyper-information Hash representation of the image.
(2) A matching stage: and comparing the hash representation of the hyper-information acquired in the acquisition stage with the registered hash template, measuring the similarity of the two hash characteristics, and judging whether the hyper-information belongs to the same class or not based on the similarity.
Further, in the acquisition stage, the test image is respectively input into a physical hash code learner and a correlation multidimensional hash code learner, and the hyper-information hash representation of the image is obtained.
Further, in the matching stage, the obtained super information hash representation is compared with the registered hash template, and the similarity of two hash characteristics measured by the sea name distance is utilized to judge whether the super information hash representation belongs to the same class.
Further, the construction of the physical hash learner comprises the following steps:
(1-1) identifying distinguishing characteristics of objects and designing a physical hash code;
and (1-2) training a physical hash code learner based on the obtained physical hash code.
Further, in the step (1-1), the distinguishing characteristics of the object are identified, and a plurality of physical hash codes are designed by using the distinguishing characteristics of the object, wherein each hash code is regarded as an important part of the composition target.
Further, in the step (1-2), according to each hash code designed in the step (1-1), all samples are divided into a positive class and a negative class, then training is carried out by using Resnet based on the collected samples and the physical hash codes, the Resnet obtained by training is a learner of the physical hash codes of the distinguishing features of the object, and finally, for the sample to be tested, the value of the corresponding physical hash code can be obtained only by inputting an image.
Further, due to the fact that the deep neural network can learn robust features, the relevance multidimensional hash code learner is constructed by the deep neural network.
Further, the network architecture of the correlation multidimensional hash code learner adopts a densenet, which is different from the traditional densenet in that the correlation multidimensional hash code learner introduces a correlation loss function at a loss layer:
Figure BDA0002238752590000021
in the above equation, t is the number of hash codes, Y(：，k)The value representing the kth hash code of all samples, X being the input image, an n X d matrix, wkIs the associated weight vector, α is the hyperparameter;
by minimizing the objective function, the correlation multidimensional Hash code learner can be obtained.
Further, the formula mainly comprises 2 terms, the first term is a fitting term, and the fitting term is mainly used for ensuring that the training error is minimum;
the second term is a regularization term, which is used to improve the generalization ability of the model.
Further, the regularization term is established by using an L2 paradigm, so that an objective function is smooth, the generalization capability of the model is enhanced, and α hyper-parameters are used for balancing the regularization term and the fitting term.
Compared with the prior art, the Hash learning method based on the depth super information has the following outstanding advantages that:
(1) the obtained super information hash code has good interpretability, and compared with the traditional hash learning method, the obtained super information hash code is easier for people to understand the physical significance of the super information hash code.
(2) The method integrates the correlation between the hashes, improves the resolving power of the hashes, avoids the complicated hash code obtaining process in the hash obtaining process, and improves the identification precision and efficiency.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
Fig. 1 is a flow diagram of a hash learning method based on deep hyper-information.
Detailed Description
The present invention will be described in further detail with reference to specific embodiments in order to better understand the technical solutions of the present invention. It is to be understood that the described embodiments are merely exemplary of the invention, and not restrictive of the full scope of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
A preferred embodiment is given below:
as shown in fig. 1, the hash learning method based on deep hyper-information in this embodiment is mainly divided into two stages, which are a training stage and a testing stage.
(I) training phase
(1) Constructing a physical hash learning device:
and (1-1) identifying the distinguishing characteristics of the object and designing a physical hash code. For example, to identify multiple images of cats, fish, birds, etc., several physical hash codes may be designed, each of which may be considered to be an important part of the composition of the object. In this embodiment, whether the object is aquatic, whether the object has wings, whether the object has limbs, and whether the object has tails is regarded as a distinctive feature of the object, and the physical hash code is designed based on the distinctive features.
And (1-2) training a physical hash code learner based on the obtained physical hash code. All samples are classified into a positive class and a negative class according to each hash code. For example, for water. Samples of objects such as dolphins, sharks, and tropical fishes may be regarded as positive classes, and samples of objects such as dogs, cats, and birds may be regarded as negative classes, and the hash value may be regarded as 0. Training is carried out by utilizing Resnet based on the collected samples and the physical hash codes, and the Resnet obtained by training is a learner for the physical hash codes of whether the water is generated or not. For the sample to be detected, the value of the corresponding physical hash code can be obtained only by inputting the image.
(2) Constructing a correlation multidimensional hash code learner: the deep neural network can learn the characteristics of robustness, so that the invention utilizes the deep neural network to construct a correlation multidimensional hash code learner. The network architecture adopts densenet, and different from the traditional densenet, the relevance multidimensional hash code learner introduces a relevance loss function in a loss layer:
Figure BDA0002238752590000041
the method mainly comprises 2 items, the first item is a fitting item, the fitting item is mainly used for ensuring the training error to be minimum, and the regularization item is used for improving the generalization capability of the model.
In the above equation, t is the number of hash codes, Y(：，k)The value representing the kth hash code of all samples, X being the input image, is an n X d matrix. w is akThe method comprises the steps of establishing a regularization term by using an L2 paradigm, smoothing an objective function and enhancing the generalization capability of a model, wherein α is a hyper-parameter and is used for balancing the regularization term and a fitting term, and a relevance multidimensional Hash code learning device can be obtained by minimizing the objective function.
(II) test phase
(1) An acquisition stage; and respectively inputting the test image into an attractive hash code learning device and a correlation multidimensional hash code learning device to obtain the hyper-information hash representation of the image.
(2) A matching stage: and comparing the obtained hyper-information hash representation with the registered hash template, and judging whether the obtained hyper-information hash representation belongs to the same class or not based on the similarity by measuring the similarity of two hash characteristics by using the sea name distance.
The above embodiments are only specific cases, and the scope of the present invention includes but is not limited to the above embodiments, and any suitable changes or substitutions that are consistent with the claims of the present deep super information based hash learning method and are made by those skilled in the art should fall within the scope of the present invention.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (10)

1. A Hash learning method based on deep hyper-information is characterized by mainly comprising a training stage and a testing stage, and comprises the following steps:
(I) training phase
(1) Constructing a physical hash learning device: performing network architecture training based on the distinguishing characteristics of the identified objects, and constructing a physical hash learning device;
(2) constructing a correlation multidimensional hash code learner: constructing a correlation multidimensional hash code learner by utilizing a deep neural network;
(II) test phase
(1) An acquisition stage; and inputting the tested image into a trained learner, and acquiring the hyper-information Hash representation of the image.
(2) A matching stage: and comparing the hash representation of the hyper-information acquired in the acquisition stage with the registered hash template, measuring the similarity of the two hash characteristics, and judging whether the hyper-information belongs to the same class or not based on the similarity.
2. The hash learning method based on the deep hyper-information as claimed in claim 1, wherein in the obtaining stage, the test image is input to the physical hash code learner and the correlation multidimensional hash code learner respectively, so as to obtain the hyper-information hash representation of the image.
3. The deep hyper-information based hash learning method according to claim 1 or 2, wherein in the matching stage, the obtained hyper-information hash representation is compared with a registered hash template, and the similarity between two hash features is measured by using sea name distance to judge whether the two hash features belong to the same class.
4. The hash learning method based on the deep hyper-information as claimed in claim 1, wherein the physical hash learning is constructed by the following steps:
(1-1) identifying distinguishing characteristics of objects and designing a physical hash code;
and (1-2) training a physical hash code learner based on the obtained physical hash code.
5. The deep super information-based hash learning method according to claim 4, wherein the distinguishing characteristics of the object are identified in step (1-1), and a plurality of physical hash codes are designed according to the distinguishing characteristics of the object, wherein each hash code is regarded as an important part of the composition target.
6. The hash learning method based on the deep meta-information as claimed in claim 5, wherein in step (1-2), all samples are classified into positive and negative classes according to each hash code designed in step (1-1), then training is performed by using Resnet based on the collected samples and physical hash codes, the Resnet obtained by training is a learner of the physical hash code of the object distinguishing features, and finally, for the sample to be measured, only an image needs to be input, and the value of the corresponding physical hash code can be obtained.
7. The hash learning method based on deep hyper-information as claimed in claim 6, wherein the correlation multidimensional hash code learner is constructed by using a deep neural network.
8. The hash learning method based on deep hyper-information as claimed in claim 7, wherein the network architecture of the correlation multidimensional hash code learner employs a densenet, which is different from the conventional densenet in that the correlation multidimensional hash code learner introduces a correlation loss function at a loss layer:
Figure FDA0002238752580000021
in the above equation, t is the number of hash codes, Y(：，k)The value representing the kth hash code of all samples, X being the input image, an n X d matrix, wkIs the associated weight vector, α is the hyperparameter;
by minimizing the objective function, the correlation multidimensional Hash code learner can be obtained.
9. The hash learning method based on the deep hyper-information as claimed in claim 8, wherein the formula mainly comprises 2 terms, the first term is a fitting term, and the fitting term is mainly used for ensuring that the training error is minimum;
the second term is a regularization term, which is used to improve the generalization ability of the model.
10. The hash learning method based on the deep super information as claimed in claim 9, wherein the regularization term is established using an L2 paradigm to smooth the objective function and enhance the generalization ability of the model, and α super parameters are used to balance the regularization term and the fitting term.
CN201910992620.0A 2019-10-18 2019-10-18 Hash learning method based on deep hyper-information Pending CN110766065A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910992620.0A CN110766065A (en) 2019-10-18 2019-10-18 Hash learning method based on deep hyper-information

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910992620.0A CN110766065A (en) 2019-10-18 2019-10-18 Hash learning method based on deep hyper-information

Publications (1)

Publication Number Publication Date
CN110766065A true CN110766065A (en) 2020-02-07

Family

ID=69332848

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910992620.0A Pending CN110766065A (en) 2019-10-18 2019-10-18 Hash learning method based on deep hyper-information

Country Status (1)

Country Link
CN (1) CN110766065A (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106503106A (en) * 2016-10-17 2017-03-15 北京工业大学 A kind of image hash index construction method based on deep learning
CN108549915A (en) * 2018-04-27 2018-09-18 成都考拉悠然科技有限公司 Image hash code training pattern algorithm based on two-value weight and classification learning method
CN108805157A (en) * 2018-04-11 2018-11-13 南京理工大学 Classifying Method in Remote Sensing Image based on the random supervision discrete type Hash in part

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106503106A (en) * 2016-10-17 2017-03-15 北京工业大学 A kind of image hash index construction method based on deep learning
CN108805157A (en) * 2018-04-11 2018-11-13 南京理工大学 Classifying Method in Remote Sensing Image based on the random supervision discrete type Hash in part
CN108549915A (en) * 2018-04-27 2018-09-18 成都考拉悠然科技有限公司 Image hash code training pattern algorithm based on two-value weight and classification learning method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
高宪军: "半监督哈希算法研究", 《中国优秀博硕士学位论文全文数据库（硕士）信息科技辑》 *

Similar Documents

Publication Publication Date Title
CN109101938B (en) Multi-label age estimation method based on convolutional neural network
CN115018021B (en) Machine room abnormity detection method and device based on graph structure and abnormity attention mechanism
CN111581405A (en) Cross-modal generalization zero sample retrieval method for generating confrontation network based on dual learning
CN109241317A (en) Based on the pedestrian's Hash search method for measuring loss in deep learning network
CN111222434A (en) Method for obtaining evidence of synthesized face image based on local binary pattern and deep learning
CN105184260B (en) A kind of image characteristic extracting method and pedestrian detection method and device
Ensafi et al. A bag of words based approach for classification of HEp-2 cell images
CN108520215B (en) Single-sample face recognition method based on multi-scale joint feature encoder
CN110866134B (en) Image retrieval-oriented distribution consistency keeping metric learning method
Zhong et al. A comparative study of image classification algorithms for Foraminifera identification
CN112633382A (en) Mutual-neighbor-based few-sample image classification method and system
CN104615634A (en) Direction feature based palm vein guiding quick retrieval method
Li et al. Dating ancient paintings of Mogao Grottoes using deeply learnt visual codes
CN112784921A (en) Task attention guided small sample image complementary learning classification algorithm
CN114897085A (en) Clustering method based on closed subgraph link prediction and computer equipment
CN112163106B (en) Second-order similar-perceived image hash code extraction model establishment method and application thereof
CN113269274A (en) Zero sample identification method and system based on cycle consistency
CN117153268A (en) Cell category determining method and system
Bi et al. Critical direction projection networks for few-shot learning
He et al. Wood species identification based on an ensemble of deep convolution neural networks
Karlinsky et al. RepMet: Representative-based metric learning for classification and one-shot object detection
CN110766065A (en) Hash learning method based on deep hyper-information
Bai et al. A unified deep learning model for protein structure prediction
CN110533636B (en) Image analysis device
CN111898579A (en) Extreme gradient lifting-based unbiased semi-supervised classification model for high-resolution remote sensing images

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: 20200207