CN108399211B - Large-scale image retrieval algorithm based on binary characteristics - Google Patents

Large-scale image retrieval algorithm based on binary characteristics Download PDF

Info

Publication number
CN108399211B
CN108399211B CN201810106624.XA CN201810106624A CN108399211B CN 108399211 B CN108399211 B CN 108399211B CN 201810106624 A CN201810106624 A CN 201810106624A CN 108399211 B CN108399211 B CN 108399211B
Authority
CN
China
Prior art keywords
real
similarity
loss function
difference
training
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810106624.XA
Other languages
Chinese (zh)
Other versions
CN108399211A (en
Inventor
鲁继文
周杰
陈志祥
袁鑫
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.)
Tsinghua University
Original Assignee
Tsinghua University
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 Tsinghua University filed Critical Tsinghua University
Priority to CN201810106624.XA priority Critical patent/CN108399211B/en
Publication of CN108399211A publication Critical patent/CN108399211A/en
Application granted granted Critical
Publication of CN108399211B publication Critical patent/CN108399211B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Library & Information Science (AREA)
  • Databases & Information Systems (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a large-scale image retrieval algorithm based on binary characteristics, which comprises the following steps: step S1: initializing neural network parameters, and initializing real-value output characteristics according to a training picture set; step S2: constructing a picture similarity matrix according to the training picture set, and constructing a Laplace matrix; step S3: constructing a loss function through weighting similarity measurement; step S4: the real-valued output characteristics are derived through a loss function, the real-valued output characteristics are updated by fixing the difference quantity, and meanwhile, the network parameters are updated; step S5: the difference is derived through a loss function, and the real-value output characteristic is fixed to update the difference; step S6: the high-order expansion weight is increased, and the real-valued output features and the network parameters are continuously updated according to steps S3 and S4 in combination with the loss function until the training is finished. The method can effectively compensate the problem caused by the imbalance of the positive training sample and the negative training sample in the input data pair, and effectively improve the retrieval precision.

Description

Large-scale image retrieval algorithm based on binary characteristics
Technical Field
The invention relates to the technical field of computer vision and multimedia, in particular to a large-scale image retrieval algorithm based on binary characteristics.
Background
Image retrieval is a data search method used to find pictures. The system searches some pictures similar to the input pictures from the database according to the retrieval information such as keywords, pictures and the like input by the user and feeds back the pictures to the user. The measure of similarity may be based on picture auxiliary information (e.g. keywords) or picture content characteristics such as picture texture, color, shape, etc.
Content-based image retrieval is an application of computational vision in the field of image retrieval. Such algorithms aim to avoid retrieving images using textual information, but rather by the characteristics of the texture, color, shape, etc. of the picture itself. Such algorithms require calculation of euclidean distances in feature space for the search image and the database image. In large-scale datasets, both the storage overhead using real-valued features and the time overhead of calculating euclidean distances at the time of retrieval are unacceptable.
The hash-based image retrieval can solve the problem of excessive time and storage overhead. Hash-based image retrieval algorithms store and retrieve images using binary features rather than real-valued features. The distance calculation between the binary features can be quickly realized by utilizing exclusive-or operation, and meanwhile, because each bit in the binary features only needs 1-bit storage space, the storage cost of the database picture features can be obviously reduced. The binary feature is called a hash feature, a function mapped from an original space to a hamming space is called a hash function, and a process of learning to obtain the hash feature is called hash learning.
The biggest difficulty in hash learning is that the optimization problem of solving the optimal hash characteristics is an NP problem. This is determined by the property that the value of the hash feature can only take 0,1 or ± 1. The integer optimization problem cannot solve the optimal solution through a traditional numerical optimization method, and therefore constraint conditions need to be relaxed. There are three main types of relaxation methods: and directly discarding the binary constraint, introducing a quantization error optimization term, and relaxing the step function into a sigmoid function. The first method is directly neglected by the constraint condition, so that the learned hash function has huge quantization error. The second method needs to introduce real-valued hidden layer characteristics and other auxiliary variables, decompose the original integer optimization problem into a plurality of solvable subproblems, and seek a local optimal solution through a step-by-step iterative optimization method. Sometimes, the sub-optimization problem of the hash feature is still an NP-hard problem without a closed-form solution, and a local optimal solution of the sub-problem needs to be converged to by a coordinated descent method. In the third method, the convergence speed of the training model is obviously reduced due to the introduction of a nonlinear function. In any of the above methods, there is always a difference between the trained hash function Φ and the hash function Ψ ═ sgn (Φ) actually used. This can result in a reduction in the effectiveness of the search on data outside the sample set.
Furthermore, when the training data is given in the form of pairs of pairwise data, previous approaches have defined the supervision flags of similar pairs of pictures as 1 and the supervision flags of dissimilar pairs of pictures as 0 or-1. The imbalance between positive and negative training samples is caused because the number of similarities in the pair of pairwise data constructed from most training sets is always much smaller than the number of dissimilarities.
Disclosure of Invention
The present invention is directed to solving, at least to some extent, one of the technical problems in the related art.
Therefore, the invention aims to provide a large-scale image retrieval algorithm based on binary characteristics, and the method can effectively improve the retrieval precision.
In order to achieve the above object, an embodiment of the present invention provides a large-scale image retrieval algorithm based on binary features, including the following steps: step S1: initializing neural network parameters, and initializing real-value output characteristics according to a training picture set; step S2: constructing a picture similarity matrix according to the training picture set, and constructing a Laplace matrix; step S3: constructing a loss function through weighting similarity measurement; step S4: the real-valued output characteristics are derived through the loss function, the fixed difference quantity updates the real-valued output characteristics, and meanwhile, network parameters are updated; step S5: the difference is derived through the loss function, and the difference is updated by fixing the real-value output characteristic; step S6: the higher-order unfolding weights are increased, and the real-valued output features and the network parameters are continuously updated according to the steps S3 and S4 in combination with the loss function until the training is finished.
According to the large-scale image retrieval algorithm based on the binary characteristics, an original binary optimization problem in Hash learning is converted into an optimization problem which can be guided to a Hash function, a binary constraint and similarity keeping target in Hash learning is decoupled, so that the converted problem can be optimized and solved through a simple interactive iteration frame, the problem caused by unbalance of positive and negative training samples in an input data pair is effectively compensated through a weighting similarity measurement method, the problem that the Hash function obtained through training is inconsistent with the Hash function used actually is effectively solved, and retrieval precision is improved.
Further, in an embodiment of the present invention, the step S1 further includes: step S101: initializing neural network parameters; step S102: and acquiring real-value output characteristics of pictures in a training picture set by using the initialized neural network, taking the real-value output characteristics as the initial real-value output characteristics H and the discrete output characteristics B as sgn (H), and taking the discrete output characteristics B as picture hash codes.
Further, the method can be used for preparing a novel materialIn one embodiment of the present invention, the step S2 includes the following steps: step S201: calculating the similarity of any two pictures in the training picture set through a binary similarity function, and recording the similarity of the ith picture and the jth picture as sijThe picture similarity matrix formed by the method is marked as S; step S202: obtaining a Laplace matrix of the picture similarity matrix S, and recording the Laplace matrix as Lsym。
Further, in an embodiment of the present invention, the step S202 specifically includes:
computing
Figure BDA0001567934510000031
D＝diag(d1,d2,......,dN)，L＝D-S；
The Laplace matrix of the picture similarity matrix S
Figure BDA0001567934510000032
Further, in an embodiment of the present invention, the step S3 includes the following steps:
step S301: compensating imbalance of positive and negative training samples by using a weighted similarity measurement method, calculating weighted similarity of any two pictures in the training picture set through the similarity, and recording the weighted similarity of the ith picture and the jth picture as
Figure BDA0001567934510000033
The weighted similarity matrix of the two sets is recorded as
Figure BDA0001567934510000034
Step S302: for any pictures i and j in the training picture set, according to the discrete output characteristic biAnd bjAnd constructing a loss function:
Figure BDA0001567934510000035
step S303: summing all sample pairs in the training picture set to construct a loss function:
Figure BDA0001567934510000036
the loss function matrix form is:
Figure BDA0001567934510000037
wherein the content of the first and second substances,
Figure BDA0001567934510000038
d is a diagonal matrix and diagonal elements
Figure BDA0001567934510000039
Step S304: defining a difference Δ ═ B-H, according to a taylor series, the loss function develops at the real-valued output characteristic H as follows:
Figure BDA00015679345100000310
wherein, if the real-valued output characteristic H and the difference quantity delta are connected with each element in a row manner, then
Figure BDA00015679345100000311
Is the ith column vector of real-valued output features H,
Figure BDA00015679345100000312
is the ith column vector of the difference Δ;
step S305: according to the expanded form of step S304, the loss function in step S303 is:
Figure BDA00015679345100000313
step S306: combining the step S303 and the step S305, constructing the loss function with respect to the real-valued output feature H and the difference Δ:
Figure BDA0001567934510000041
wherein, (H + delta) epsilon { -1,1}n×lWherein λ is1And λ2Weights are expanded for higher orders.
Further, in an embodiment of the present invention, the step S301 specifically includes:
for any pictures i and j in the training picture set, according to the similarity sijCalculating the weighted similarity:
Figure BDA0001567934510000042
in order to ensure that the similarity of the similar pictures is a positive value, and the similarity of the non-similar pictures is a negative value in the similarity measurement, let β <1 > be 0, and if β is 0.5, the original similarity is obtained by 0.5-fold scaling, and a weighted similarity measurement method is not used.
Further, in an embodiment of the present invention, the step S4 includes the following steps:
step S401: fixing a difference quantity, constructing the loss function with respect to the real-valued output features:
Figure BDA0001567934510000043
step S402: fixing the difference, calculating the derivative of the loss function on the real-valued output characteristic:
Figure BDA0001567934510000044
step S403: and updating the real-valued output characteristics and the network parameters according to a random gradient descent method.
Further, in an embodiment of the present invention, the step S5 includes the following steps:
step S501: fixing the real-valued output characteristics, constructing a loss function with respect to the difference quantity:
Figure BDA0001567934510000045
satisfies (H + delta) epsilon { -1,1}n×l；
Step S502: and (3) fixing real-value output characteristics, calculating the derivation of the loss function on the difference quantity:
Figure BDA0001567934510000046
step S503: the difference amount Δ is updated.
Further, in an embodiment of the present invention, the step S503 specifically includes:
according to
Figure BDA0001567934510000047
And iteratively updating the difference quantity.
Further, in an embodiment of the present invention, the step S6 includes the following steps: step S601: increasing the higher order expansion weight λ1And λ2(ii) a Step S602: according to the steps S4 and S5, the real-valued output features and the network parameters are continuously updated in combination with the loss function until the training is finished.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
The foregoing and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a flow chart of a large scale image retrieval algorithm based on binary features according to an embodiment of the present invention;
FIG. 2 is a flowchart of a large-scale image retrieval algorithm based on binary features according to an embodiment of the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are illustrative and intended to be illustrative of the invention and are not to be construed as limiting the invention.
A proposed binary-feature-based large-scale image retrieval algorithm according to an embodiment of the present invention is described below with reference to the accompanying drawings.
FIG. 1 is a flowchart of a large-scale image retrieval algorithm based on binary features according to an embodiment of the present invention.
As shown in fig. 1 and fig. 2, the binary-feature-based large-scale image retrieval algorithm includes the following steps:
step S1: initializing neural network parameters, and initializing real-value output characteristics according to the training picture set.
Further, in an embodiment of the present invention, the step S1 further includes: step S101: initializing neural network parameters; step S102: and acquiring real-value output characteristics of the pictures in the training picture set by using the initialized neural network, taking the real-value output characteristics as initial real-value output characteristics H and discrete output characteristics B as sgn (H), and taking the discrete output characteristics B as picture hash codes.
Step S2: and constructing a picture similarity matrix according to the training picture set, and constructing a Laplace matrix.
Further, in an embodiment of the present invention, the step S2 includes the following steps: step S201: calculating the similarity of any two pictures in the training picture set through a binary similarity function, and recording the similarity of the ith picture and the jth picture as sijThe picture similarity matrix formed by the method is marked as S; step S202: obtain the Laplace matrix of the picture similarity matrix S, and record as Lsym。
Further, in an embodiment of the present invention, step S202 specifically includes:
computing
Figure BDA0001567934510000061
D＝diag(d1,d2,......,dN)，L＝D-S；
Laplace matrix of picture similarity matrix S
Figure BDA0001567934510000062
Step S3: the loss function is constructed by weighting the similarity measure.
Further, in an embodiment of the present invention, the step S3 includes the following steps:
step S301: using a weighted similarity measurement method to compensate the imbalance of positive and negative training samples, calculating the weighted similarity of any two pictures in a training picture set through the similarity, and recording the weighted similarity of the ith picture and the jth picture as
Figure BDA0001567934510000063
The weighted similarity matrix of the two sets is recorded as
Figure BDA0001567934510000064
Step S302: for any pictures i and j in the training picture set, according to the discrete output characteristic biAnd bjAnd constructing a loss function:
Figure BDA0001567934510000065
step S303: summing all sample pairs in the training picture set to construct a loss function:
Figure BDA0001567934510000066
the loss function matrix form is:
Figure BDA0001567934510000067
wherein the content of the first and second substances,
Figure BDA0001567934510000068
d is a diagonal matrix and diagonal elements
Figure BDA0001567934510000069
Step S304: defining the difference Δ ═ B-H, the loss function develops at the real-valued output characteristic H according to a taylor series as follows:
Figure BDA00015679345100000610
wherein, if the real-valued output characteristic H and the difference quantity delta are connected with each element in a row manner, then
Figure BDA00015679345100000611
Is the ith column vector of real-valued output features H,
Figure BDA00015679345100000612
is the ith column vector of the difference Δ;
step S305: according to the expanded form of step S304, the loss function in step S303 is:
Figure BDA00015679345100000613
step S306: combining step S303 and step S305, a loss function is constructed for the real-valued output feature H and the difference Δ:
Figure BDA0001567934510000071
wherein, (H + delta) epsilon { -1,1}n×lWherein λ is1And λ2Weights are expanded for higher orders.
Further, in an embodiment of the present invention, step S301 specifically includes:
for any pictures i and j in the training picture set, according to the similarity sijCalculating the weighted similarity:
Figure BDA0001567934510000072
in order to ensure that the similarity measure has a positive similarity for similar pictures and a negative similarity for non-similar pictures, let β <1 > be 0, and if β is 0.5, obtain the original similarity by 0.5-fold scaling, a weighted similarity measure method is not used.
Step S4: and (4) the real-valued output characteristics are derived through a loss function, the real-valued output characteristics are updated by fixing the difference quantity, and meanwhile, the network parameters are updated.
Further, in an embodiment of the present invention, the step S4 includes the following steps:
step S401: fixing the difference quantity, constructing a loss function with respect to the real-valued output features:
Figure BDA0001567934510000073
step S402: fixing the difference, calculating the derivation of the loss function on the real-valued output characteristics:
Figure BDA0001567934510000074
step S403: and updating the real-valued output characteristics and the network parameters according to a random gradient descent method.
Step S5: and (4) carrying out derivation on the difference quantity through a loss function, and fixing the real-value output characteristic to update the difference quantity.
Further, in an embodiment of the present invention, the step S5 includes the following steps:
step S501: and (3) fixing real-value output characteristics, constructing a loss function about the difference:
Figure BDA0001567934510000075
satisfies (H + delta) epsilon { -1,1}n×l；
Step S502: fixing real value output characteristics, calculating the derivation of the loss function on the difference quantity:
Figure BDA0001567934510000076
step S503: the difference amount Δ is updated.
Further, in an embodiment of the present invention, step S503 specifically includes:
according to
Figure BDA0001567934510000081
And iteratively updating the difference quantity.
Step S6: the high-order expansion weight is increased, and the real-valued output features and the network parameters are continuously updated according to steps S3 and S4 in combination with the loss function until the training is finished.
Further, in an embodiment of the present invention, the step S6 includes the following steps: step S601: increasing the higher order expansion weight λ1And λ2(ii) a Step S602: according to steps S4 and S5, the real-valued output features and the network parameters are continuously updated by combining the loss function until the training is finished.
According to the large-scale image retrieval algorithm based on the binary characteristics, the original binary optimization problem in the Hash learning is converted into the optimization problem which can be guided to the Hash function, the binary constraint and the similarity keeping target in the Hash learning are decoupled, the converted problem can be optimized and solved through a simple interactive iteration framework, the problem caused by unbalance of positive and negative training samples in an input data pair is effectively compensated through a weighted similarity measurement method, the problem that the Hash function obtained through training is inconsistent with the Hash function used actually is effectively solved, and the retrieval precision is improved.
In the description of the present invention, it is to be understood that the terms "central," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," "circumferential," and the like are used in the orientations and positional relationships indicated in the drawings for convenience in describing the invention and to simplify the description, and are not intended to indicate or imply that the referenced devices or elements must have a particular orientation, be constructed and operated in a particular orientation, and are therefore not to be considered limiting of the invention.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.
In the present invention, unless otherwise expressly stated or limited, the terms "mounted," "connected," "secured," and the like are to be construed broadly and can, for example, be fixedly connected, detachably connected, or integrally formed; can be mechanically or electrically connected; they may be directly connected or indirectly connected through intervening media, or they may be connected internally or in any other suitable relationship, unless expressly stated otherwise. The specific meanings of the above terms in the present invention can be understood by those skilled in the art according to specific situations.
In the present invention, unless otherwise expressly stated or limited, the first feature "on" or "under" the second feature may be directly contacting the first and second features or indirectly contacting the first and second features through an intermediate. Also, a first feature "on," "over," and "above" a second feature may be directly or diagonally above the second feature, or may simply indicate that the first feature is at a higher level than the second feature. A first feature being "under," "below," and "beneath" a second feature may be directly under or obliquely under the first feature, or may simply mean that the first feature is at a lesser elevation than the second feature.
In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims (7)

1. A large-scale image retrieval method based on binary characteristics is characterized by comprising the following steps:
step S1: initializing neural network parameters, and initializing real-value output characteristics according to a training picture set;
step S2: constructing a picture similarity matrix according to the training picture set, and constructing a Laplace matrix;
step S3: constructing a loss function through weighting similarity measurement;
step S4: the real-valued output characteristics are derived through the loss function, the fixed difference quantity updates the real-valued output characteristics, and meanwhile, network parameters are updated;
step S5: the difference is derived through the loss function, and the difference is updated by fixing the real-value output characteristic; and
step S6: increasing the high-order expansion weight, and continuously updating the real-valued output features and the network parameters according to the step S4 and the step S5 in combination with the loss function until the training is finished;
wherein the step S1 further includes:
step S101: initializing neural network parameters;
step S102: acquiring real-value output characteristics of pictures in a training picture set by using the initialized neural network, taking the real-value output characteristics as initial real-value output characteristics H and discrete output characteristics B as sgn (H), and taking the discrete output characteristics B as picture hash codes;
the step S2 includes the steps of:
step S201: calculating the similarity of any two pictures in the training picture set through a binary similarity function, and recording the similarity of the ith picture and the jth picture as sijThe picture similarity matrix formed by the method is marked as S;
step S202: obtaining a Laplace matrix of the picture similarity matrix S, and recording the Laplace matrix as Lsym；
The step S3 includes the steps of:
step S301: compensating imbalance of positive and negative training samples by using a weighted similarity measurement method, calculating weighted similarity of any two pictures in the training picture set through the similarity, and recording the weighted similarity of the ith picture and the jth picture as
Figure FDA0002570865110000011
The weighted similarity matrix of the two sets is recorded as
Figure FDA0002570865110000012
Step S302: for any pictures i and j in the training picture set, according to the discrete output characteristic biAnd bjAnd constructing a loss function:
Figure FDA0002570865110000013
step S303: summing all sample pairs in the training picture set to construct a loss function:
Figure FDA0002570865110000014
the loss function matrix form is:
Figure FDA0002570865110000021
wherein the content of the first and second substances,
Figure FDA0002570865110000022
d is a diagonal matrix and diagonal elements
Figure FDA0002570865110000023
Step S304: defining a difference Δ ═ B-H, according to a taylor series, the loss function develops at the real-valued output characteristic H as follows:
Figure FDA0002570865110000024
wherein, if the real-valued output characteristic H and the difference quantity delta are connected with each element in a row manner, then
Figure FDA0002570865110000025
Is the ith column vector of real-valued output features H,
Figure FDA0002570865110000026
is the ith column vector of the difference Δ;
step S305: according to the expanded form of step S304, the loss function in step S303 is:
Figure FDA0002570865110000027
step S306: combining the step S303 and the step S305, constructing the loss function with respect to the real-valued output feature H and the difference Δ:
Figure FDA0002570865110000028
wherein, (H + delta) epsilon { -1,1}n×lWherein λ is1And λ2Weights are expanded for higher orders.
2. The method for retrieving a large-scale image based on binary features according to claim 1, wherein the step S202 specifically comprises:
computing
Figure FDA0002570865110000029
D＝diag(d1,d2,......,dN)，L＝D-S；
The Laplace matrix of the picture similarity matrix S
Figure FDA00025708651100000210
3. The method for retrieving a large-scale image based on binary features according to claim 1, wherein the step S301 specifically comprises:
for any pictures i and j in the training picture set, according to the similarity sijCalculating the weighted similarity:
Figure FDA00025708651100000211
in order to ensure that the similarity of the similar pictures is a positive value, and the similarity of the non-similar pictures is a negative value in the similarity measurement, let β <1 > be 0, and if β is 0.5, the original similarity is obtained by 0.5-fold scaling, and a weighted similarity measurement method is not used.
4. The binary-feature-based large-scale image retrieval method according to claim 1 or 3, wherein the step S4 includes the steps of:
step S401: fixing a difference quantity, constructing the loss function with respect to the real-valued output features:
Figure FDA00025708651100000212
step S402: fixing the difference, calculating the derivative of the loss function on the real-valued output characteristic:
Figure FDA0002570865110000031
step S403: and updating the real-valued output characteristics and the network parameters according to a random gradient descent method.
5. The binary-feature-based large-scale image retrieval method according to claim 4, wherein the step S5 comprises the steps of:
step S501: fixing the real-valued output characteristics, constructing a loss function with respect to the difference quantity:
Figure FDA0002570865110000032
satisfies (H + delta) epsilon { -1,1}n×l；
Step S502: and (3) fixing real-value output characteristics, calculating the derivation of the loss function on the difference quantity:
Figure FDA0002570865110000033
step S503: the difference amount Δ is updated.
6. The large-scale image retrieval method based on binary features as claimed in claim 5, wherein said step S503 specifically is:
according to
Figure FDA0002570865110000034
And iteratively updating the difference quantity.
7. The binary-feature-based large-scale image retrieval method according to claim 5, wherein the step S6 comprises the steps of:
step S601: increasing the higher order expansion weight λ1And λ2；
Step S602: according to the steps S4 and S5, the real-valued output features and the network parameters are continuously updated in combination with the loss function until the training is finished.
CN201810106624.XA 2018-02-02 2018-02-02 Large-scale image retrieval algorithm based on binary characteristics Active CN108399211B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810106624.XA CN108399211B (en) 2018-02-02 2018-02-02 Large-scale image retrieval algorithm based on binary characteristics

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810106624.XA CN108399211B (en) 2018-02-02 2018-02-02 Large-scale image retrieval algorithm based on binary characteristics

Publications (2)

Publication Number Publication Date
CN108399211A CN108399211A (en) 2018-08-14
CN108399211B true CN108399211B (en) 2020-11-24

Family

ID=63096218

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810106624.XA Active CN108399211B (en) 2018-02-02 2018-02-02 Large-scale image retrieval algorithm based on binary characteristics

Country Status (1)

Country Link
CN (1) CN108399211B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109918532B (en) * 2019-03-08 2023-08-18 苏州大学 Image retrieval method, device, equipment and computer readable storage medium
CN109977194B (en) * 2019-03-20 2021-08-10 华南理工大学 Text similarity calculation method, system, device and medium based on unsupervised learning
CN113157739B (en) * 2021-04-23 2024-01-09 平安科技（深圳）有限公司 Cross-modal retrieval method and device, electronic equipment and storage medium
CN113705589A (en) * 2021-10-29 2021-11-26 腾讯科技（深圳）有限公司 Data processing method, device and equipment

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103336801A (en) * 2013-06-20 2013-10-02 河海大学 Multi-feature locality sensitive hashing (LSH) indexing combination-based remote sensing image retrieval method
CN107085585A (en) * 2016-02-12 2017-08-22 奥多比公司 Accurate label dependency prediction for picture search

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105069173B (en) * 2015-09-10 2019-04-19 天津中科智能识别产业技术研究院有限公司 The fast image retrieval method of Hash is kept based on the topology for having supervision
CN105469096B (en) * 2015-11-18 2018-09-25 南京大学 A kind of characteristic bag image search method based on Hash binary-coding
CN105512289B (en) * 2015-12-07 2018-08-14 郑州金惠计算机系统工程有限公司 Image search method based on deep learning and Hash
CN106021364B (en) * 2016-05-10 2017-12-12 百度在线网络技术（北京）有限公司 Foundation, image searching method and the device of picture searching dependency prediction model

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103336801A (en) * 2013-06-20 2013-10-02 河海大学 Multi-feature locality sensitive hashing (LSH) indexing combination-based remote sensing image retrieval method
CN107085585A (en) * 2016-02-12 2017-08-22 奥多比公司 Accurate label dependency prediction for picture search

Also Published As

Publication number Publication date
CN108399211A (en) 2018-08-14

Similar Documents

Publication Publication Date Title
CN108399211B (en) Large-scale image retrieval algorithm based on binary characteristics
CN110059198B (en) Discrete hash retrieval method of cross-modal data based on similarity maintenance
US11651037B2 (en) Efficient cross-modal retrieval via deep binary hashing and quantization
US8438011B2 (en) Suggesting spelling corrections for personal names
CN111160564B (en) Chinese knowledge graph representation learning method based on feature tensor
CN109284411B (en) Discretization image binary coding method based on supervised hypergraph
CN103473307B (en) Across media sparse hash indexing means
US8386490B2 (en) Adaptive multimedia semantic concept classifier
CN108399268B (en) Incremental heterogeneous graph clustering method based on game theory
KR101623860B1 (en) Method for calculating similarity between document elements
CN111274424B (en) Semantic enhanced hash method for zero sample image retrieval
CN103488713A (en) Cross-modal search method capable of directly measuring similarity of different modal data
CN112395438A (en) Hash code generation method and system for multi-label image
Cintia Ganesha Putri et al. Design of an unsupervised machine learning-based movie recommender system
CN111753190A (en) Meta learning-based unsupervised cross-modal Hash retrieval method
CN112800344B (en) Deep neural network-based movie recommendation method
Zhang et al. Modeling the Homophily Effect between Links and Communities for Overlapping Community Detection.
JP7131616B2 (en) Time-series data processor
CN116821519A (en) Intelligent recommendation method for system filtering and noise reduction based on graph structure
CN116383437A (en) Cross-modal material recommendation method based on convolutional neural network
CN116958613A (en) Depth multi-view clustering method and device, electronic equipment and readable storage medium
WO2023279685A1 (en) Method for mining core users and core items in large-scale commodity sales
CN110659375A (en) Hash model training method, similar object retrieval method and device
Wang et al. An overview of t-SNE optimization algorithms
Song et al. Hybrid Recommendation Based on Matrix Factorization and Deep Learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant