CN109492120A - Model training method, search method, device, electronic equipment and storage medium - Google Patents
Model training method, search method, device, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN109492120A CN109492120A CN201811292397.0A CN201811292397A CN109492120A CN 109492120 A CN109492120 A CN 109492120A CN 201811292397 A CN201811292397 A CN 201811292397A CN 109492120 A CN109492120 A CN 109492120A
- Authority
- CN
- China
- Prior art keywords
- feature vector
- vector
- shape
- samples pictures
- class center
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Abstract
The embodiment of the present invention provides a kind of model training method, search method, device, electronic equipment and storage medium, it obtains samples pictures and the label for characterizing image category in the samples pictures, samples pictures includes: the projection image of two-dimentional cartographical sketching and multiple visual angles；By the 3D shape retrieval model of samples pictures input pre-training, first eigenvector and multiple second feature vectors are extracted；First eigenvector and each second feature vector are incorporated into the same high n-dimensional subspace n, obtain third feature vector and fourth feature vector；Based on third feature vector, multiple fourth feature vectors, label, multiple class center vectors and pre-set criteria, feature vector, updated parameter and class center vector with the 3D shape of institute's samples pictures are obtained, reduces the training time, improves retrieval precision.
Description
Technical field
The present invention relates to field of image processings, in particular to a kind of model training method, search method, device, electricity
Sub- equipment and storage medium.
Background technique
Last decade is the fast-developing phase of computer vision technique, especially with the gradually mature of depth learning technology,
The key problem (such as image recognition, target following, image segmentation and image labeling) in the field all realizes fast development.
For the mode of research object, computer vision technique can be mainly divided into two major classes 1) it is based on color, texture
Information (two dimensional image analysis)；2) space, shape information (3D shape analysis) are based on.However, two dimensional image is space object
Plane projection, the information such as a large amount of space and shape will be lost in projection process, and two dimensional image is easy by illumination
With the influence of attitudes vibration.On the contrary, the influence that 3D shape is illuminated by the light with attitudes vibration is unobvious, two dimensional image can be made up
Latent defect.Just because of this, 3D shape analysis is receive more and more attention, has all put into a large amount of manpower and material resources both at home and abroad
It is studied.
Due to the limitation of three-dimensional shape data acquisition modes, compared to conventional two-dimensional image database up to ten million quantity easily
The scale of grade, the scale in three-dimensional shape data library is generally smaller, such as maximum three-dimensional shape data library ShapeNet packet at present
About 3,000,000 shapes are contained, core subset ShapeNetCore only contains 55 classes totally 51300 3D shapes.It is smaller
Data set be that the training of complicated deep neural network brings obstacle.However, although mode is different, two dimensional image and three-dimensional shaped
The transportable property of shape knowledge between describing the correlation and mode for objective world with height.Therefore for being based on Freehandhand-drawing
The cross-module state 3D shape retrieval of sketch is an important research direction in the field.
However following problem of the existing technology: training process is extremely complex, needs to generate positive sample, negative sample carries out
Training, increases trained difficulty, training time, the twin generalization ability of network performance of the positive counter-example mode of generation is not strong, multi-angle of view mould
The center of the higher dimensional space of type characterization and cartographical sketching mode gap are larger, and the retrieval effectiveness finally obtained is bad.
Summary of the invention
In consideration of it, the embodiment of the present invention is designed to provide a kind of model training method, search method, device, electronics
Equipment and storage medium, to alleviate the above problem.
In a first aspect, the embodiment of the present invention provides a kind of model training method, which comprises obtain samples pictures and
For characterizing the label of image category in the samples pictures, wherein the samples pictures include: two-dimentional cartographical sketching and multiple
The projection image at visual angle；By the 3D shape retrieval model of samples pictures input pre-training, the two-dimentional Freehandhand-drawing grass is extracted
The second feature vector of every projection image in the first eigenvector of figure and multiple projection images；By the first eigenvector
Incorporate same high n-dimensional subspace n with each second feature vector, obtain third feature corresponding with the first eigenvector to
Amount and fourth feature vector corresponding with the second feature vector；Based on the third feature vector, the multiple described 4th
Feature vector, the label, multiple class center vectors and pre-set criteria, in the 3D shape retrieval model of the pre-training
Parameter and class center vector corresponding with the label be updated, obtain corresponding with samples pictures 3D shape
Feature vector and updated parameter and class center vector.
Second aspect, the embodiment of the present invention provide a kind of search method, are applied to above-mentioned 3D shape retrieval model, described
Method includes: to obtain Target Photo to be retrieved；Based on the Target Photo and the 3D shape retrieval model, described in acquisition
The corresponding target feature vector of Target Photo；Calculate separately all three-dimensionals in the target feature vector and three-dimensional shape data library
The Euclidean distance of the corresponding feature vector of shape, and the Euclidean distance is arranged according to retrieval ordering, acquisition is carried out from small to large
Sequence chained list.
The third aspect, the embodiment of the present invention provide a kind of model training apparatus, and described device includes: the first acquisition module,
Label for obtaining samples pictures and for characterizing image category in the samples pictures, wherein the samples pictures include:
The projection image of two-dimentional cartographical sketching and multiple visual angles；Input module, for the samples pictures to be inputted to the three-dimensional of pre-training
Shape-memory behavior model extracts in the first eigenvector and multiple projection images of the two-dimentional cartographical sketching every projection image
Second feature vector；Fusion Features module, it is same for incorporating the first eigenvector and each second feature vector
High n-dimensional subspace n obtains corresponding with first eigenvector third feature vector and corresponding with the second feature vector
Fourth feature vector；Update module, for being based on the third feature vector, multiple fourth feature vectors, the mark
Label, multiple class center vectors and pre-set criteria, in the 3D shape retrieval model of the pre-training parameter and with it is described
The corresponding class center vector of label is updated, and obtains the feature vector and more of corresponding with samples pictures 3D shape
Parameter and class center vector after new.
Fourth aspect, the embodiment of the present invention provide a kind of retrieval device, are applied to above-mentioned 3D shape retrieval model, described
Device includes: that Target Photo obtains module, for obtaining Target Photo to be retrieved；Target feature vector obtains module, is used for
Based on the Target Photo and the 3D shape retrieval model, the corresponding target feature vector of the Target Photo is obtained；Inspection
Rope module, for calculate separately target feature vector feature corresponding with 3D shapes all in three-dimensional shape data library to
The Euclidean distance of amount, and by the Euclidean distance according to retrieval ordering is carried out from small to large, obtain sequence chained list.
5th aspect, the embodiment of the present invention provide a kind of electronic equipment, the electronic equipment include processor and with it is described
The memory of processor connection, the memory is interior to store computer program, when the computer program is held by the processor
When row, so that the electronic equipment executes method described in first aspect and second aspect.
6th aspect, the embodiment of the present invention provide a kind of storage medium, are stored with computer program in the storage medium,
When the computer program is run on computers, so that the computer executes first aspect and closes side described in second aspect
Method.
Compared with prior art, model training method, search method, device, the electronics of various embodiments of the present invention proposition are set
Standby and storage medium beneficial effect is: samples pictures and the label for characterizing image category in the samples pictures are obtained,
Wherein, the samples pictures include: the projection image of two-dimentional cartographical sketching and multiple visual angles；The samples pictures are inputted into pre- instruction
Experienced 3D shape retrieval model extracts every throwing in the first eigenvector and multiple projection images of the two-dimentional cartographical sketching
The second feature vector of shadow picture；It is empty that the first eigenvector and each second feature vector are incorporated into same higher-dimension
Between, obtain and the corresponding third feature vector of the first eigenvector and the 4th spy corresponding with the second feature vector
Levy vector；Based on the third feature vector, multiple fourth feature vectors, the label, multiple class center vectors and
Pre-set criteria, in the 3D shape retrieval model of the pre-training parameter and class center vector corresponding with the label into
Row updates, obtain corresponding with samples pictures 3D shape feature vector and updated parameter and class center to
Amount.The classification that image is realized using class center vector, is trained without generating positive sample and negative sample, and by multiple visual angles
The corresponding feature vector of projection image feature vector corresponding with two-dimentional cartographical sketching incorporates the same high n-dimensional subspace n, overcomes
The center of the higher dimensional space of multi-angle of view model characterization and the larger drawback of cartographical sketching mode gap, reduce the training time, improve
Retrieval precision.
To enable the above objects, features and advantages of the present invention to be clearer and more comprehensible, preferred embodiment is cited below particularly, and cooperate
Appended attached drawing, is described in detail below.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached
Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair
The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this
A little attached drawings obtain other relevant attached drawings.
Fig. 1 is the structural schematic diagram of electronic equipment provided in an embodiment of the present invention；
Fig. 2 is a kind of flow chart of model training method provided in an embodiment of the present invention；
Fig. 3 is a kind of flow chart of search method provided in an embodiment of the present invention；
Fig. 4 is a kind of module diagram of model training apparatus provided in an embodiment of the present invention；
Fig. 5 is a kind of module diagram for retrieving device provided in an embodiment of the present invention.
Icon: 100- electronic equipment；110- memory；120- storage control；130- processor；140- Peripheral Interface；
150- input-output unit；170- display unit；200- model training apparatus；210- first obtains module；220- input module；
230- Fusion Features module；240- update module；300- retrieves device；310- Target Photo obtains module；320- target signature
Vector obtains module；330- retrieval module.
Specific embodiment
Below in conjunction with attached drawing in the embodiment of the present invention, technical solution in the embodiment of the present invention carries out clear, complete
Ground description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Usually exist
The component of the embodiment of the present invention described and illustrated in attached drawing can be arranged and be designed with a variety of different configurations herein.Cause
This, is not intended to limit claimed invention to the detailed description of the embodiment of the present invention provided in the accompanying drawings below
Range, but it is merely representative of selected embodiment of the invention.Based on the embodiment of the present invention, those skilled in the art are not doing
Every other embodiment obtained under the premise of creative work out, shall fall within the protection scope of the present invention.
It should also be noted that similar label and letter indicate similar terms in following attached drawing, therefore, once a certain Xiang Yi
It is defined in a attached drawing, does not then need that it is further defined and explained in subsequent attached drawing.Meanwhile of the invention
In description, term " first ", " second " etc. are only used for distinguishing description, are not understood to indicate or imply relative importance.
The embodiment of the present invention provides the structural schematic diagram of a kind of electronic equipment 100, and the electronic equipment 100 can be individual
Computer (personal computer, PC), tablet computer, smart phone, personal digital assistant (personal digital
Assistant, PDA) etc..
As shown in Figure 1, the electronic equipment 100 may include: retrieval device 300, model training apparatus 200, memory
110, storage control 120, processor 130, Peripheral Interface 140, input-output unit 150, display unit 170.
The memory 110, storage control 120, processor 130, Peripheral Interface 140, input-output unit 150 and
Each element of display unit 170 is directly or indirectly electrically connected between each other, to realize the transmission or interaction of data.For example, this
A little elements can be realized by one or more communication bus or signal wire be electrically connected between each other.300 He of retrieval device
The model training apparatus 200 includes that at least one can be stored in the memory in the form of software or firmware (firmware)
In 110 or the software function module that is solidificated in the operating system (operating system, OS) of client device.The place
Reason device 130 is for executing the executable module stored in memory 110, such as the sequence.
Wherein, memory 110 may be, but not limited to, random access memory (Random Access Memory,
RAM), read-only memory (Read Only Memory, ROM), programmable read only memory (Programmable Read-Only
Memory, PROM), erasable read-only memory (Erasable Programmable Read-Only Memory, EPROM),
Electricallyerasable ROM (EEROM) (Electric Erasable Programmable Read-Only Memory, EEPROM) etc..
Wherein, memory 110 is for storing program, and the processor 130 executes described program after receiving and executing instruction, aforementioned
Method performed by the electronic equipment 100 for the flow definition that any embodiment of the embodiment of the present invention discloses can be applied to processor
In 130, or realized by processor 130.
Processor 130 may be a kind of IC chip, the processing capacity with signal.Above-mentioned processor 130 can
To be general processor, including central processing unit (Central Processing Unit, abbreviation CPU), network processing unit
(Network Processor, abbreviation NP) etc.；Can also be digital signal processor (DSP), specific integrated circuit (ASIC),
Field programmable gate array (FPGA) either other programmable logic device, discrete gate or transistor logic, discrete hard
Part component.It may be implemented or execute disclosed each method, step and the logic diagram in the embodiment of the present invention.General processor
It can be microprocessor or the processor be also possible to any conventional processor etc..
Various input/output devices are couple processor 130 and memory 110 by the Peripheral Interface 140.Some
In embodiment, Peripheral Interface 140, processor 130 and storage control 120 can be realized in one single chip.Other one
In a little examples, they can be realized by independent chip respectively.
Input-output unit 150 is used to be supplied to the interaction that user input data realizes user and electronic equipment 100.It is described
Input-output unit 150 may be, but not limited to, mouse and keyboard etc..
Display unit 170 provides an interactive interface (such as user interface) between electronic equipment 100 and user
Or it is referred to for display image data to user.In the present embodiment, the display unit 170 can be liquid crystal display or touching
Control display.It can be the touching of the capacitance type touch control screen or resistance-type of support single-point and multi-point touch operation if touch control display
Control screen etc..Single-point and multi-point touch operation is supported to refer to that touch control display can sense on the touch control display one or more
The touch control operation generated simultaneously at a position, and the touch control operation that this is sensed transfers to processor 130 to be calculated and handled.
Embodiment
Referring to figure 2., Fig. 2 is a kind of flow chart of model training method provided in an embodiment of the present invention.It below will be to Fig. 2
Shown in process be described in detail, the method be applied to Fig. 1 described in electronic equipment 100, which comprises
S100: samples pictures and the label for characterizing image category in the samples pictures are obtained, wherein the sample
Picture includes: the projection image of two-dimentional cartographical sketching and multiple visual angles.
In the actual implementation process, samples pictures are the electricity that can input the electronic equipment 100 with image-capable
Sub-pictures, the content inside samples pictures can be cat, dog, apple, desk, people etc., wherein Felis belongs to one in a kind of, dog
Class, apple belong to one kind, and desk belongs to one kind, and Genus Homo by samples pictures and is used to characterize in the samples pictures and schemed in one kind
As the label input of classification has the electronic equipment 100 of image-capable.
S200: by the 3D shape retrieval model of samples pictures input pre-training, the two-dimentional cartographical sketching is extracted
First eigenvector and multiple projection images in every projection image second feature vector.
In the actual implementation process, by the way that the samples pictures are inputted the pre-training being stored in electronic equipment 100
3D shape retrieval model, wherein the 3D shape retrieval model of pre-training can for AlexNet, Vgg11, Vgg16,
Vgg19, ResNet18, ResNet34, ResNet50 etc. pass through the 3D shape retrieval model of pre-training, 100 energy of electronic equipment
Enough extract the second feature of every projection image in the first eigenvector and multiple projection images of the two-dimentional cartographical sketching to
Amount.
S300: incorporating the same high n-dimensional subspace n for the first eigenvector and each second feature vector, obtain with
The corresponding third feature vector of first eigenvector and fourth feature vector corresponding with the second feature vector.
In order to make two-dimentional cartographical sketching and the corresponding feature vector of the projection image at multiple visual angles be in same sub-spaces
It is interior, as an implementation, the present embodiment using matrix network by first eigenvector and each second feature to
Amount incorporates in same high n-dimensional subspace n, then obtain third feature vector corresponding with the first eigenvector and with institute
State the corresponding fourth feature vector of second feature vector, it is to be understood that the dimension of third feature vector sum fourth feature vector
Spend it is identical, then overcome multi-angle of view model characterization higher dimensional space center and the larger drawback of cartographical sketching mode gap.
Due to not being that every projection image is all useful in the projection image at multiple visual angles, to reduce computation complexity,
It needs to screen the corresponding fourth feature vector of the projection image at multiple visual angles, filters out useful fourth feature vector,
To reduce the computation complexity of model training, screening technique will be specifically described below:
Based on the first pre-set criteria, multiple fourth feature vectors are screened, the fourth feature after obtaining screening
Vector, wherein before the number of the fourth feature vector after screening is less than screening.
Assuming that the same three-dimensional samples model has n visual angle before screening to multiple fourth feature vector Ms
Fourth feature vector constitute dimension be (100 × n) matrix M, first pre-set criteria be M*=M × F (M), wherein F
(M) indicate that dimension is the weight matrix of n × m, by the way that M is substituted into M*It is special to obtain the higher-dimension that dimension is (100 × m) by=M × F (M)
Levy M*, wherein m is less than n, whereinWherein, WAttnAnd bAttnIt respectively indicates
The weight of Attention Model and biasing.Dimensionality reduction may be implemented by the step, then reduce computation complexity.
S400: based on the third feature vector, multiple fourth feature vectors, the label, multiple class centers to
Amount and pre-set criteria, in the 3D shape retrieval model of the pre-training parameter and class corresponding with label center
Vector is updated, obtain feature vector corresponding with the three-dimensional samples model and updated parameter and class center to
Amount.
As an implementation, it is based on the second pre-set criteria, wherein two pre-set criteria are as follows:K=1,2 ..., K. by by the third feature to
Measure (WTS+b k-th of class center vector C) and in multiple class center vectorskIt substitutes into second pre-set criteria, obtains described the
First distance in three feature vectors and multiple class center vectors between each class center vector.
As an implementation, it is based on third pre-set criteria, wherein institute's third pre-set criteria are as follows:a∈Rm×1, by by multiple fourth feature
Matrix (the W that vector is constitutedTM+b k-th of class center vector C) and in multiple class center vectorskSubstitute into the default standard of the third
Then, the second distance between multiple each class center vectors of fourth feature vector sum is obtained.
As an implementation, based on the default standard of multiple first distances, multiple second distances and third
Then, wherein the third pre-set criteria are as follows: L=CE (- dis, k)+λ * disk,diskIndicate the second feature vector, more
The distance between a fourth feature vector and k-th of class center vector, by multiple first distances and multiple described
Two distances substitute into the third pre-set criterias, in this example, it is assumed that there is K class, it is to be appreciated that have K class center to
C, the K first distances and the K second distances are measured, to the parameter in the 3D shape retrieval model of the pre-training
W, b class center vector C corresponding with for characterizing the label of kth classkIt is updated, obtains corresponding with the three-dimensional samples model
Feature vector and updated parameter and class center vector.
It, should so that the feature vector of mutually similar all two-dimentional sketches, three-dimensional samples model is all close by step S400
The corresponding class center vector of class, while the distance between inhomogeneity in higher dimensional space is remote enough, it can be real without positive negative sample
Now classify, reduce the training time and calculates cost.
Referring to figure 3., Fig. 3 is a kind of flow chart of search method provided in an embodiment of the present invention.It below will be to shown in Fig. 3
Process be described in detail, the method be applied to Fig. 1 in electronic equipment, which comprises
S600: Target Photo to be retrieved is obtained.
Wherein, Target Photo to be retrieved can be two-dimentional cartographical sketching, can also be the picture of video camera shooting, or
Picture to be retrieved input is had the electronic equipment 100 of image-capable, wherein the electronics by the projection image of threedimensional model
Equipment 100 can be realized by trained 3D shape retrieval model in advance to the corresponding three-dimensional shaped of Target Photo to be retrieved
Shape retrieval.
S700: it is based on the Target Photo and the 3D shape retrieval model, obtains the corresponding mesh of the Target Photo
Mark feature vector.
The Target Photo is input in preparatory trained 3D shape retrieval model by terminal device, obtains institute
State the corresponding target feature vector of Target Photo.
S800: target feature vector feature corresponding with 3D shapes all in three-dimensional shape data library is calculated separately
The Euclidean distance of vector, and by the Euclidean distance according to retrieval ordering is carried out from small to large, obtain sequence chained list.
Wherein, each three-dimensional samples shape and corresponding feature vector are stored in three-dimensional shape data library, by Europe
Formula distance-taxis obtains sequence chained list, and according to sequence chained list, chooses the corresponding three-dimensional shaped of the shortest feature vector of Euclidean distance
Shape.
Referring to figure 4., Fig. 4 is a kind of structural block diagram of model training apparatus 200 provided in an embodiment of the present invention.Below will
Structural block diagram shown in Fig. 4 is illustrated, shown device includes:
First obtains module 210, the mark for obtaining samples pictures and for characterizing image category in the samples pictures
Label, wherein the samples pictures include: the projection image of two-dimentional cartographical sketching and multiple visual angles.
Input module 220, for extracting described two for the 3D shape retrieval model of samples pictures input pre-training
Tie up the second feature vector of every projection image in the first eigenvector and multiple projection images of cartographical sketching.
Fusion Features module 230, for the first eigenvector and each second feature vector to be incorporated the same height
N-dimensional subspace n, obtains and the corresponding third feature vector of the first eigenvector and corresponding with the second feature vector
Fourth feature vector.
Update module 240, for based on the third feature vector, multiple fourth feature vectors, the label, more
A class center vector and pre-set criteria, in the 3D shape retrieval model of the pre-training parameter and with the label pair
The class center vector answered is updated, obtain feature vector corresponding with the three-dimensional samples model and updated parameter and
Class center vector.
As an implementation, described device further include:
Screening module screens multiple fourth feature vectors, obtains screening for being based on the first pre-set criteria
Fourth feature vector afterwards, wherein before the number of the fourth feature vector after screening is less than screening.
As an implementation, the update module 240, comprising:
First distance obtains module, for being based on the second pre-set criteria, obtains in the multiple classes of third feature vector sum
First distance in Heart vector between each class center vector.
Second distance obtains module, and for being based on third pre-set criteria, it is each to obtain multiple fourth feature vector sums
Second distance between class center vector.
Submodule is updated, for being based on multiple first distances, multiple second distances and third pre-set criteria,
To in the 3D shape retrieval model of the pre-training parameter and class center vector corresponding with the label be updated, obtain
Take the feature vector and updated parameter and class center vector of 3D shape corresponding with the samples pictures.
Referring to figure 5., Fig. 5 is a kind of structural block diagram of shape retrieval device 300 provided in an embodiment of the present invention.It below will be right
Structural block diagram shown in fig. 5 is illustrated, and shown device includes:
Target Photo obtains module 310, for obtaining Target Photo to be retrieved.
Target feature vector obtains module 320, for being based on the Target Photo and the 3D shape retrieval model, obtains
Take the corresponding target feature vector of the Target Photo.
Retrieval module 330, for calculating separately all three-dimensional shapeds in the target feature vector and three-dimensional shape data library
The Euclidean distance of the corresponding feature vector of shape, and the Euclidean distance is obtained into sequence according to retrieval ordering is carried out from small to large
Chained list.
In addition, it is stored with computer program in the storage medium the embodiment of the invention also provides a kind of storage medium,
When the computer program is run on computers, so that the computer executes any one of present invention embodiment and is provided
Training method and search method.
It is apparent to those skilled in the art that for convenience and simplicity of description, the model of foregoing description
The specific work process of training device 200 and retrieval device 300, can be with reference in foregoing model training method and search method
Corresponding process no longer excessively repeats herein.
In conclusion various embodiments of the present invention propose model training method, search method, device, electronic equipment and deposit
Storage media: samples pictures and the label for characterizing image category in the samples pictures are obtained, wherein the samples pictures packet
It includes: the projection image of two-dimentional cartographical sketching and multiple visual angles；The 3D shape of samples pictures input pre-training is retrieved into mould
Type, extract the second feature of every projection image in the first eigenvector and multiple projection images of the two-dimentional cartographical sketching to
Amount；The first eigenvector and each second feature vector are incorporated into the same high n-dimensional subspace n, obtained and first spy
Levy the corresponding third feature vector of vector and fourth feature vector corresponding with the second feature vector；Based on the third
Feature vector, multiple fourth feature vectors, the label, multiple class center vectors and pre-set criteria, to the pre- instruction
Parameter and class center vector corresponding with the label in experienced 3D shape retrieval model are updated, and are obtained and described three
Tie up the corresponding feature vector of sample pattern and updated parameter and class center vector；Image is realized using class center vector
Classification, is trained without generating positive sample and negative sample, and by the corresponding feature vector of the projection image at multiple visual angles and two
It ties up the corresponding feature vector of cartographical sketching and incorporates the same high n-dimensional subspace n, overcome the higher dimensional space of multi-angle of view model characterization
Center and the larger drawback of cartographical sketching mode gap reduce the training time, improve retrieval precision.
Each functional module in each embodiment of the present invention can integrate one independent part of formation together, can also
To be modules individualism, an independent part can also be integrated to form with two or more modules.
The foregoing is only a preferred embodiment of the present invention, is not intended to restrict the invention, for the skill of this field
For art personnel, the invention may be variously modified and varied.All within the spirits and principles of the present invention, made any to repair
Change, equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.It should also be noted that similar label and letter exist
Similar terms are indicated in following attached drawing, therefore, once being defined in a certain Xiang Yi attached drawing, are then not required in subsequent attached drawing
It is further defined and explained.
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
Those familiar with the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all contain
Lid is within protection scope of the present invention.Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
It should be noted that, in this document, relational terms such as first and second and the like are used merely to a reality
Body or operation are distinguished with another entity or operation, are deposited without necessarily requiring or implying between these entities or operation
In any actual relationship or order or sequence.Moreover, the terms "include", "comprise" or its any other variant are intended to
Non-exclusive inclusion, so that the process, method, article or equipment including a series of elements is not only wanted including those
Element, but also including other elements that are not explicitly listed, or further include for this process, method, article or equipment
Intrinsic element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that
There is also other identical elements in process, method, article or equipment including the element.
Claims (10)
1. a kind of model training method, which is characterized in that the described method includes:
Obtain samples pictures and the label for characterizing image category in the samples pictures, wherein the samples pictures include:
The projection image of two-dimentional cartographical sketching and multiple visual angles；
By the 3D shape retrieval model of samples pictures input pre-training, the fisrt feature of the two-dimentional cartographical sketching is extracted
The second feature vector of every projection image in multiple projection images of vector sum；
The first eigenvector and each second feature vector are incorporated into the same high n-dimensional subspace n, obtained and first spy
Levy the corresponding third feature vector of vector and fourth feature vector corresponding with the second feature vector；
Based on the third feature vector, multiple fourth feature vectors, the label, multiple class center vectors and preset
Criterion, in the 3D shape retrieval model of the pre-training parameter and class center vector corresponding with the label carry out more
Newly, the feature vector of acquisition 3D shape corresponding with the samples pictures and updated parameter and class center vector.
2. the method according to claim 1, wherein being obtained corresponding with the label being based on the label
After class center vector, the method also includes:
Based on the first pre-set criteria, multiple fourth feature vectors are screened, the fourth feature vector after obtaining screening,
Wherein, before the number of the fourth feature vector after screening is less than screening.
3. according to the method described in claim 2, it is characterized in that, based on the third feature vector, multiple four spies
Vector, the label, multiple class center vectors and pre-set criteria are levied, in the 3D shape retrieval model of the pre-training
Parameter and class center vector corresponding with the label are updated, and obtain the spy of 3D shape corresponding with the samples pictures
Levy vector and updated parameter and class center vector, comprising:
Based on the second pre-set criteria, obtain in the multiple class center vectors of the third feature vector sum between each class center vector
First distance；
Based on third pre-set criteria, the second distance between multiple each class center vectors of fourth feature vector sum is obtained；
Based on multiple first distances, multiple second distances and third pre-set criteria, to the three-dimensional of the pre-training
Parameter and class center vector corresponding with the label in Shape-memory behavior model are updated, and are obtained and the samples pictures pair
The feature vector for the 3D shape answered and updated parameter and class center vector.
4. a kind of search method, which is characterized in that retrieve mould applied to 3D shape described in any claim in claim 1-3
Type, which comprises
Obtain Target Photo to be retrieved；
Based on the Target Photo and the 3D shape retrieval model, obtain the corresponding target signature of the Target Photo to
Amount；
Calculate separately the Europe of target feature vector feature vector corresponding with 3D shapes all in three-dimensional shape data library
Formula distance, and by the Euclidean distance according to retrieval ordering is carried out from small to large, obtain sequence chained list.
5. a kind of training device, which is characterized in that described device includes:
First obtains module, the label for obtaining samples pictures and for characterizing image category in the samples pictures, wherein
The samples pictures include: the two-dimentional cartographical sketching of three-dimensional samples model and the projection image at multiple visual angles；
Input module, for extracting the two-dimentional Freehandhand-drawing for the 3D shape retrieval model of samples pictures input pre-training
The second feature vector of every projection image in the first eigenvector of sketch and multiple projection images；
Fusion Features module, it is empty for the first eigenvector and each second feature vector to be incorporated same higher-dimension
Between, obtain and the corresponding third feature vector of the first eigenvector and the 4th spy corresponding with the second feature vector
Levy vector；
Update module, for based in the third feature vector, multiple fourth feature vectors, the label, multiple classes
Heart vector and pre-set criteria, to the parameter and class corresponding with the label in the 3D shape retrieval model of the pre-training
Center vector is updated, obtain corresponding with samples pictures 3D shape feature vector and updated parameter and
Class center vector.
6. device according to claim 5, which is characterized in that further include:
Screening module screens multiple fourth feature vectors, after obtaining screening for being based on the first pre-set criteria
Fourth feature vector, wherein before the number of the fourth feature vector after screening is less than screening.
7. device according to claim 6, which is characterized in that the update module, comprising:
First distance obtain module, for be based on the second pre-set criteria, obtain the multiple class centers of the third feature vector sum to
First distance in amount between each class center vector；
Second distance obtains module, for being based on third pre-set criteria, obtains in multiple each classes of fourth feature vector sum
Second distance between Heart vector；
Submodule is updated, for being based on multiple first distances, multiple second distances and third pre-set criteria, to institute
State the parameter in the 3D shape retrieval model of pre-training and class center vector corresponding with the label be updated, obtain with
The corresponding corresponding feature vector of 3D shape of samples pictures and updated parameter and class center vector.
8. a kind of retrieval device, which is characterized in that retrieve mould applied to 3D shape described in any claim in claim 5-7
Type, described device include:
Target Photo obtains module, for obtaining Target Photo to be retrieved；
Target feature vector obtains module, for being based on the Target Photo and the 3D shape retrieval model, described in acquisition
The corresponding target feature vector of Target Photo；
Retrieval module, it is corresponding with all 3D shapes in three-dimensional shape data library for calculating separately the target feature vector
The Euclidean distance of feature vector, and by the Euclidean distance according to retrieval ordering is carried out from small to large, obtain sequence chained list.
9. a kind of electronic equipment, which is characterized in that including processor and the memory being connected to the processor, the storage
Computer program is stored in device, when the computer program is executed by the processor, so that the electronic equipment right of execution
Benefit requires method described in any one of 1-4.
10. a kind of storage medium, which is characterized in that computer program is stored in the storage medium, when the computer journey
When sequence is run on computers, so that the computer executes the method as described in any one of claim 1-4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811292397.0A CN109492120B (en) | 2018-10-31 | 2018-10-31 | Model training method, retrieval method, device, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811292397.0A CN109492120B (en) | 2018-10-31 | 2018-10-31 | Model training method, retrieval method, device, electronic equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109492120A true CN109492120A (en) | 2019-03-19 |
| CN109492120B CN109492120B (en) | 2020-07-03 |
Family
ID=65692107
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811292397.0A Active CN109492120B (en) | 2018-10-31 | 2018-10-31 | Model training method, retrieval method, device, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109492120B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796180A (en) * | 2019-10-12 | 2020-02-14 | 吉林大学 | Model training system and method based on artificial intelligence |
| CN113191400A (en) * | 2021-04-14 | 2021-07-30 | 中国海洋大学 | Method and device for retrieving corresponding three-dimensional model based on two-dimensional image |
| CN113240716A (en) * | 2021-05-31 | 2021-08-10 | 西安电子科技大学 | Twin network target tracking method and system with multi-feature fusion |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104361347A (en) * | 2014-10-21 | 2015-02-18 | 浙江大学 | Numerically-controlled machine tool design module three-dimension model retrieval method based on single image |
| US20170032214A1 (en) * | 2014-02-04 | 2017-02-02 | Fraunhofer-Gesellschaft Zur Foerderung Der Angewandten Forschung E.V. | 2D Image Analyzer |
| CN107122396A (en) * | 2017-03-13 | 2017-09-01 | 西北大学 | Three-dimensional model searching algorithm based on depth convolutional neural networks |
| CN107220277A (en) * | 2017-04-14 | 2017-09-29 | 西北大学 | Image retrieval algorithm based on cartographical sketching |
-
2018
- 2018-10-31 CN CN201811292397.0A patent/CN109492120B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170032214A1 (en) * | 2014-02-04 | 2017-02-02 | Fraunhofer-Gesellschaft Zur Foerderung Der Angewandten Forschung E.V. | 2D Image Analyzer |
| CN104361347A (en) * | 2014-10-21 | 2015-02-18 | 浙江大学 | Numerically-controlled machine tool design module three-dimension model retrieval method based on single image |
| CN107122396A (en) * | 2017-03-13 | 2017-09-01 | 西北大学 | Three-dimensional model searching algorithm based on depth convolutional neural networks |
| CN107220277A (en) * | 2017-04-14 | 2017-09-29 | 西北大学 | Image retrieval algorithm based on cartographical sketching |
Non-Patent Citations (1)
| Title |
|---|
| 沙腾: "《基于子空间学习的人脸特征分析及应用》", 《中国优秀硕士论文全文数据库（电子期刊）》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110796180A (en) * | 2019-10-12 | 2020-02-14 | 吉林大学 | Model training system and method based on artificial intelligence |
| CN110796180B (en) * | 2019-10-12 | 2022-06-07 | 吉林大学 | Model training system and method based on artificial intelligence |
| CN113191400A (en) * | 2021-04-14 | 2021-07-30 | 中国海洋大学 | Method and device for retrieving corresponding three-dimensional model based on two-dimensional image |
| CN113191400B (en) * | 2021-04-14 | 2022-04-19 | 中国海洋大学 | Method and device for retrieving corresponding three-dimensional model based on two-dimensional image |
| CN113240716A (en) * | 2021-05-31 | 2021-08-10 | 西安电子科技大学 | Twin network target tracking method and system with multi-feature fusion |
| CN113240716B (en) * | 2021-05-31 | 2023-04-18 | 西安电子科技大学 | Twin network target tracking method and system with multi-feature fusion |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109492120B (en) | 2020-07-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109948425B (en) | Pedestrian searching method and device for structure-aware self-attention and online instance aggregation matching | |
| Jiao et al. | SAR images retrieval based on semantic classification and region-based similarity measure for earth observation | |
| US7860317B2 (en) | Generating search results based on duplicate image detection | |
| CN110059807A (en) | Image processing method, device and storage medium | |
| CN107003977A (en) | System, method and apparatus for organizing the photo of storage on a mobile computing device | |
| CN101479728A (en) | Visual and multi-dimensional search | |
| Ip et al. | Saliency-assisted navigation of very large landscape images | |
| CN109492120A (en) | Model training method, search method, device, electronic equipment and storage medium | |
| CN105205135B (en) | A kind of 3D model retrieval methods and its retrieval device based on topic model | |
| CN110472082A (en) | Data processing method, device, storage medium and electronic equipment | |
| Purificato et al. | Multimedia and geographic data integration for cultural heritage information retrieval | |
| Bayraktar et al. | A hybrid image dataset toward bridging the gap between real and simulation environments for robotics: Annotated desktop objects real and synthetic images dataset: ADORESet | |
| Pham et al. | SHREC’18: Rgb-d object-to-cad retrieval | |
| CN115222896B (en) | Three-dimensional reconstruction method, three-dimensional reconstruction device, electronic equipment and computer readable storage medium | |
| CN108509567B (en) | Method and device for building digital culture content library | |
| Hong et al. | Art painting detection and identification based on deep learning and image local features | |
| CN110046669A (en) | Half Coupling Metric based on sketch image identifies the pedestrian retrieval method of dictionary learning | |
| CN109583371A (en) | Landmark information based on deep learning extracts and matching process | |
| Zhang et al. | Re-abstraction and perturbing support pair network for few-shot fine-grained image classification | |
| Manzo et al. | FastGCN+ ARSRGemb: a novel framework for object recognition | |
| CN107909091A (en) | A kind of iconic memory degree Forecasting Methodology based on sparse low-rank regression model | |
| Polsley et al. | SketchSeeker: finding similar sketches | |
| Sommer et al. | Deep learning based person search in aerial imagery | |
| Gupta et al. | Image feature detection using an improved implementation of maximally stable extremal regions for augmented reality applications | |
| Silva et al. | Web-based supervised thematic mapping |
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 |