CN107066586B - Shoes model index management method and system - Google Patents
Shoes model index management method and system Download PDFInfo
- Publication number
- CN107066586B CN107066586B CN201710249557.2A CN201710249557A CN107066586B CN 107066586 B CN107066586 B CN 107066586B CN 201710249557 A CN201710249557 A CN 201710249557A CN 107066586 B CN107066586 B CN 107066586B
- Authority
- CN
- China
- Prior art keywords
- normal pictures
- feature vector
- shoes model
- shoes
- reference feature
- 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
Links
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
- G06F16/51—Indexing; Data structures therefor; Storage structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/25—Integrating or interfacing systems involving database management systems
- G06F16/252—Integrating or interfacing systems involving database management systems between a Database Management System and a front-end application
-
- 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
- G06F16/53—Querying
- G06F16/532—Query formulation, e.g. graphical querying
-
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
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)
- Databases & Information Systems (AREA)
- Business, Economics & Management (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- Artificial Intelligence (AREA)
- Strategic Management (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Business, Economics & Management (AREA)
- Marketing (AREA)
- Economics (AREA)
- Development Economics (AREA)
- Image Analysis (AREA)
Abstract
Shoes model index management method and system first are pre-processed to obtain corresponding normal pictures to the shoes model picture that multiple contain shoeprints；Every normal pictures are taken to obtain multiple sampling segments by preset position and level；To each normal pictures, its multiple sampling segment and the unified scaling of the normal pictures are obtained into normalization segment；Each normalization segment is inputted into progress feature extraction in trained corresponding convolutional neural networks and obtains corresponding reference feature vector；Reference feature vector is associated with the unique identifying number of affiliated shoes model and is stored in feature database；Multiple reference feature vector constitutive characteristic matrixes of same normal pictures；It treats retrieving image execution abovementioned steps and obtains multiple feature vectors to be compared；Reference feature vector will be corresponded in each feature vector to be compared and eigenmatrix each in feature database verifying is compared and counts result, return and the most matched shoes model of sole in picture to be retrieved.
Description
Technical field
The present invention relates to the crossing domains of image procossing and machine learning, and in particular to is based on convolutional neural networks and depth
The shoes model index management method and system of study.
Background technique
Chang Cunfang on large-scale one factory and footwear B2B (Business-to-Business, business to business) transaction platform
There is the sole of the different models of enormous amount, sole is the pith of shoemaking processing, and carrying out retrieval to sole on demand is to accelerate
The important technical of production process.On footwear B2B transaction platform, previous way is by the sole model of each supplier
It sorts according to factors such as price, models, browses and select for buyer；And reality is then that buyer prefers to through given practical pin
The sole picture of sample is sold to go to retrieve the supplier of corresponding sole model as input, to can retrieve this sole and its
Supplier, in order to purchase a large amount of soles for manufacturing this shoes of given effective sale sample to the supplier.So,
Buyer can find quickly the sole needed for oneself based on picture retrieval, rather than go to identify with human eye in immense sole library
Required sole.As it can be seen that realizing that the shoes model index based on characteristics of image has critically important application value.Presently relevant skill
Art is business recommender system used in electric business platform, is scanned for particular by using picture, in the hope of that can search phase
Same or similar commodity.The picture retrieval method of this business recommender system uses characteristics of image abundant, includes almost
All external appearance characteristics (such as color of upper of a shoe, shoes, style) and dealing side describe attribute, and search result is to be more in line with user
Think the shoes type of purchase.And in one factory shoe mould searching, managing and footwear B2B transaction platform, upper of a shoe, color characteristic and buy
The factors such as seller's description belong to irrelevant factor, and only the appearance, structure of sole determine the classification of shoes model.
As it can be seen that the prior art can no longer meet current one factory and footwear B2B transaction platform for shoes model index and
The demand of management.
Summary of the invention
It is a primary object of the present invention to propose a kind of shoes model index pipe based on convolutional neural networks and deep learning
Reason method, this method can carry out feature extraction by the sole picture generated to the shoes model of input to describe shoes to be retrieved
Model, and verifying is compared with inventory's feature in the feature extraction result of the image of input in retrieval, it is retrieved with realizing
Matched target shoes model, thus solve the prior art can not be provided for one factory and footwear B2B transaction platform it is reliable, efficient
The problem of shoes model index manages.
Provided technical solution is as follows for the above-mentioned purpose by the present invention:
A kind of shoes model index management method, comprising the following steps:
S1, the shoes model picture that multiple are contained with shoeprints pre-process, and obtain corresponding normal pictures, wherein
All normal pictures have consistent coloration, size and sole direction；
S2, to each normal pictures, taken by preset position and level, obtain multiple and different sampling segments；
S3, to each normal pictures, scale the multiple sampling segments formed in step S2 and the normal pictures are unified
To a predetermined size, multiple normalization segments are obtained；
S4, it each normalization segment is input in the convolutional neural networks of preparatory trained corresponding parameter carries out feature and mentions
It takes, to obtain the corresponding reference feature vector of each normalization segment；
S5, the corresponding reference feature vector of each normalization segment and the affiliated shoes of the normal pictures by each normal pictures
The unique identifying number of model is associated with and is stored in a feature database；Multiple reference feature vectors of same normal pictures constitute the mark
The eigenmatrix of quasi- picture；
S6, step S1 to S4 is executed to given picture to be retrieved, obtains multiple feature vectors to be compared；Will it is each it is described to
Feature vector is compared to be compared respectively with the reference feature vector of corresponding position and level in eigenmatrix each in feature database
To verify whether to belong to same shoes model；
S7, to the multiple feature vector to be compared and reference feature vector comparison corresponding in all eigenmatrixes
Result counted, with return with picture to be retrieved in the most matched shoes model of sole.
Above-mentioned shoes model index management method provided by the invention passes through step S1 to S5 for all shoes to be managed first
Model is stored in feature database after being described with eigenmatrix, then can be based on feature database to given shoes illustraton of model piece and retrieve,
Matched shoes model is found out, the unique identifying number of matched shoes model is obtained, corresponding shoes mould can be called by the identification number
Type picture, it can be seen that the information such as supplier.When this method is used for footwear B2B transaction platform, supplier and buying to shoe parts
All be for quotient it is quite convenient, find specific shoes model especially for the purchaser as one factory to purchase
When, it is very convenient, as long as inputting the picture comprising sole of the specific shoes model in B2B transaction platform, can retrieve
Matched shoes model, obtains its supplier information and model etc.；Even without most matched shoes model, can also be arranged according to similarity
More matched shoes model is set out in sequence.
Preferably, take specifically including by preset position and level in step S2: in the sole area of normal pictures
Multiple location points are marked out in domain；Then it is directed to each position point, multiple and different sizes are all taken centered on the location point
Segment.
Preferably, the number of the location point of mark is n, and each position point takes m different size of segments, so that
Every normal pictures obtain the different sampling segments of m × n, and every normal pictures are corresponding after step S3 obtains m × n+1 and return
One changes segment；M >=1, n >=1.
Preferably, further include the steps that being trained the convolutional neural networks for corresponding to parameter described in step S4: choosing w
(m × n+1) in total × w normalization segment of normal pictures, by the normalizing of the corresponding position of every normal pictures and level
Change segment and be divided into one group, while the corresponding normalization segment of each normal pictures itself forms m × n+1 group training number as one group
According to, with m × n+1 group training data go that the convolutional neural networks with same framework are respectively trained, with obtain m × n+1 set network
Parameter, w >=1；Every suit network parameter is correspondingly used for the normalization segment of corresponding position and level to all normal pictures
Carry out feature extraction.
Preferably, when carrying out the feature extraction, the framework of the convolutional neural networks by input terminal to output end successively
It include: multiple convolutional layers being sequentially connected in series；It is connected to the first full articulamentum of the last layer convolutional layer and is connected to second from the bottom
Second full articulamentum of layer convolutional layer；It is connected to the full articulamentum of third of the output end of the first and second full articulamentums simultaneously；Its
In, the output of the full articulamentum of third is the reference feature vector.
Preferably, step S6 is specifically included: each feature vector to be compared is corresponding with normal pictures each in feature database
The reference feature vector of position and level spliced respectively, is formed after multiple splicing vectors and is sequentially input to training in advance
SVM sentences in same classifier, to differentiate the feature vector to be compared for being constituted each splicing vector and corresponding reference feature vector
Whether belong to same shoes model and exports the differentiation result for representing yes/no.
Preferably, further include the steps that same classifier is sentenced to SVM to be trained: using fixed reference feature obtained in step S4
Two reference feature vectors of the corresponding position for belonging to same shoes model and level are carried out being spliced to form a positive example sample by vector
This vector, by two reference feature vectors for being not belonging to same shoes model of corresponding position and level be spliced to form one it is negative
Example sample vector；To obtain r positive example sample vector and r negative example sample vectors to each normal pictures；Using each
The corresponding position of normal pictures and the positive example sample vector of level and negative example sample vector go that the SVM is trained to sentence same classifier,
Same classifier is sentenced to obtain the r SVM with different parameters；Wherein r is the number of the normalization segment in a normal pictures
Amount, r >=2；Every SVM sentences same classifier and is based respectively on respective parameter, to the reference feature vector of corresponding position and level and
Feature vector to be compared is made whether to belong to the differentiation of a shoes model.
Preferably, a SVM sentences same classifier one splicing vector of every differentiation and then exports a bit, and 1 indicates to sentence
Other result be it is yes, 0 indicate differentiate result be no；Then: all feature vectors to be compared of the picture to be retrieved and feature database
The reference feature vector of corresponding position and level is sentenced same classifier via corresponding SVM respectively and is differentiated in one eigenmatrix
Afterwards, the output that r all SVM sentences same classifier constitutes one r binary sequences；When all feature squares in feature database
After battle array is all compared, v r binary sequences are generated in total, wherein v is characterized the quantity of eigenmatrix in library, v >=1；
Step S7 is specifically included: being counted the number of binary number 1 respectively to each r of binary sequence, is then selected v
1 the largest number of k first, k >=1 in a binary sequence；K before being generated when verifying will be compared with the picture to be retrieved
The identification number of shoes model belonging to the k eigenmatrix difference of name binary sequence is returned as search result.
The present invention separately also proposes a kind of shoes model index management system, including model data memory, computer-readable deposits
Storage media and processor are stored with computer program on the computer readable storage medium, and the processor is for executing
The step of computer program is to realize claim 1 the method.
Preferably, eigenmatrix of the model data memory for being formed in storing step S5.
Above-mentioned shoes model index management system proposed by the present invention can be used for the progress such as one factory, footwear B2B transaction platform
The retrieval and management of shoes model, by the system, with feature vector after capable of handling the shoes model picture of user's input
Shoes model is described, then carries out that one by one comparison, differentiation are with the shoes model (describing with eigenmatrix) stored in system
It is no to belong to same shoes model or similarity with higher, and return and belong to same shoes model or the higher eigenmatrix of similarity
Associated shoes model identification number can call corresponding shoes model picture according to identification number, also would know that corresponding shoes model
The information such as supplier.
Detailed description of the invention
Fig. 1 is the illustraton of model taken to the segment of normal pictures progress different location and level；
Fig. 2 is a kind of exemplary architecture figure of convolutional neural networks；
Fig. 3 is the process schematic that SVM sentences that same classifier is differentiated.
Specific embodiment
The invention will be further described with specific embodiment with reference to the accompanying drawing.
A specific embodiment of the invention provides a kind of shoes model index pipe based on convolutional neural networks and deep learning
Reason method, by handling the shoes model picture to be retrieved containing shoeprints of input, by convolutional neural networks into
Row feature extraction describes shoes model, by SVM sentence same classifier by the feature of shoes model to be retrieved and it is pre-stored very
Verifying is compared in the feature of multiple shoes models respectively, to export matched shoes model.This method generally comprises following steps
S1~S7:
S1, the shoes model picture that multiple are contained with shoeprints pre-process, and obtain corresponding normal pictures, wherein
All normal pictures have consistent coloration, size and sole direction；
S2, to each normal pictures, taken by preset position and level, obtain multiple and different sampling segments；
S3, to each normal pictures, scale the multiple sampling segments formed in step S2 and the normal pictures are unified
To a predetermined size, multiple normalization segments are obtained；
S4, it each normalization segment is input in the convolutional neural networks of preparatory trained corresponding parameter carries out feature and mentions
It takes, to obtain the corresponding reference feature vector of each normalization segment；
S5, the corresponding reference feature vector of each normalization segment and the affiliated shoes of the normal pictures by each normal pictures
The unique identifying number of model is associated with and is stored in a feature database；Multiple reference feature vectors of same normal pictures constitute the mark
The eigenmatrix of quasi- picture；
S6, step S1 to S4 is executed to given picture to be retrieved, obtains multiple feature vectors to be compared；Will it is each it is described to
Feature vector is compared to be compared respectively with the reference feature vector of corresponding position and level in eigenmatrix each in feature database
To verify whether to belong to same shoes model；
S7, to the multiple feature vector to be compared and reference feature vector comparison corresponding in all eigenmatrixes
Result counted, with return with picture to be retrieved in the most matched shoes model of sole.
Each implementation steps of the above method are described in detail below by some specific embodiments.
For the one factory or B2B transaction platform of the searching, managing for using this method to carry out shoes model, obtaining first will
The picture of the shoes model of owned enormous amount saves storage after carrying out image procossing.After the picture for getting shoes model,
It needs first to pre-process these pictures, for example carries out the picture that screening removal is free of sole, interaction is assisted to complete to sole
The outer rim positioning of image-region, corrective shoes bottom direction be unified direction (such as heel be unified on the left of, heelpiece be unified in the right side
The horizontal placement position of side), carry out colors countenance coloration to be uniformly converted to a fixed value, most to shoeprints region
After scale the images to fixed size (such as wide 400 pixel, high 140 pixel), form the normal pictures of many soles.
Preset position and level are for example shown in FIG. 1 in step S2, mark out in the sole region of normal pictures more
A location point, and take centered on the point of each position the segment of multiple and different sizes.Specifically, can disperse in sole region
Property terrestrial reference provide representational n location point (n >=1), such as n=7, in location point P6, P7 of heel portion, sole center
Location point P5, location point P1~P4 of heelpiece part, but this is merely exemplary, can be according to specific when this method is embodied
Situation carries out different position marks.After marking out location point, m different size of figures are taken centered on the point of each position
Block (m >=1), specific way for example may is that by taking location point P5 as an example, for this point of P5, can be divided centered on it
The segment of multiple and different level sizes out, m=3 level size L1, L2, the L3 of example as shown in figure 1.Each location point is carried out
Such distinguishing hierarchy and after taking, every normal pictures can obtain m × n=7 × 3=21 different size of samplings and scheme
Block.It is to be appreciated that the model in Fig. 1 is only exemplary, when the division of level size, each segment can not also be with correspondence
Location point centered on, as long as can also be with comprising the location point.
Continue to be illustrated by taking the model in Fig. 1 as an example, to each normal pictures, by obtain 21 sampling segments with
And itself zooms in and out uniform sizes, which is preferably identical as the smallest sampling size of segment or more
It is small.Thus obtaining 22 normalization segments, (wherein 21 correspond to the different local of normal pictures, and another 1 corresponds to standard drawing
The overall situation of piece).
In a preferred embodiment, for carry out feature extraction convolutional neural networks framework by input terminal extremely
Output end successively includes: multiple convolutional layers being sequentially connected in series；It is connected to the first full articulamentum and the connection of the last layer convolutional layer
In the second full articulamentum of layer convolutional layer second from the bottom；The third for the output end for being connected to the first and second full articulamentums simultaneously is complete
Articulamentum；Wherein, the output of the full articulamentum of the third is the reference feature vector.
In one embodiment, convolutional neural networks framework as shown in Figure 2 is provided, comprising:
First layer is convolutional layer conv1, and the input of convolution operation is the normalization segment of normal pictures, and convolution kernel size is
3*3, step-length 2, feature port number be 20, the feature sizes of output can be calculated according to formula, formula be (C, (H-k_h+1)/
Stride_h, (W-h_w+1)/stride_w), wherein C is the feature port number of output, and H and W are the height and width of input respectively,
K_h and k_w is the height and width of convolution kernel respectively, and stride_h and stride_w are the short transverse and width of convolution operation respectively
Direction step-length, such as step-length are 2 indicate to be highly 2 with width step-length；
The second layer is convolutional layer conv2, this layer input is 2*2 by core size for conv1, and step-length is 2 down-sampling layer
Characteristic layer afterwards, the calculating that the output size of down-sampling calculates convolution operation is similar, and the feature port number only exported is equal to defeated
Enter feature port number.The convolution kernel size and step-length of the convolutional layer are consistent with conv1, and output feature port number is 60；
Third layer is convolutional layer conv3, this layer input is 2*2 by core size for conv2, and step-length is 2 down-sampling layer
The convolution kernel size of characteristic layer afterwards, this convolutional layer is consistent with step-length with conv1, and the feature port number of output is 120；
4th layer is convolutional layer conv4, and it by core size is 2 that this layer input, which is conv3, after step-length is 2 down-sampling layer
Characteristic layer, the convolution kernel size of this convolutional layer is consistent with step-length with conv1, and the feature port number of output is 60；
Layer 5 is full articulamentum fc1_4, fc1_3, and fc1_4 is the full articulamentum that input is conv4, output feature
Size is the neuron number of conv4；Fc1_3 is the full articulamentum that input is conv3, and the size of output feature is conv3
Neuron number.
Layer 6 is full articulamentum fc2, and the input of this layer is two layers of fc1_3, fc1_4 splicing as a result, this layer output is
The feature vector of certain number neuron, the output of this layer are the result of feature extraction.
When carrying out the training of convolutional neural networks, the 7th layer of full articulamentum fc3 is also used, the input of this layer is fc2
The output of layer, the neuronal quantity of output when the training of convolutional neural networks module according to the existing sole classification number of acquisition by determining
Fixed, the result of output, which is sent to softmax layers, classifies, and exports existing class categories, mistake of this layer in training convolutional neural networks
It uses in journey, is not used in feature extraction.
Before carrying out feature extraction, when proceeding to step S3, convolutional neural networks can be trained to be closed
The suitable network parameter for feature extraction.Specific training process is as follows:
It needs to use the normalization segment of step S3 generation when training, then the example above, such as chooses 100 standards
22 × 100=2200 normalization the segment in total of picture, by the normalization figure of the corresponding position of every normal pictures and level
Block is divided into one group, and (such as level size is that the segment of L1 is classified as one group at the location point P1 of every normal pictures, at location point P1
Level size is that the segment of L2 is classified as one group again ... ..., then the local segment of all normal pictures has been divided into 21 groups), simultaneously
The corresponding normalization segment of each normal pictures itself forms 22 groups of training datas, goes to distinguish with 22 groups of training datas as one group
The training convolutional neural networks, i.e., each group is all trained the convolutional neural networks of same architecture, then and available 22 sets
Different network parameters；Every suit network parameter is correspondingly used for the normalization to the corresponding position and level of all normal pictures
Segment carries out feature extraction.
Then illustrate that step S4 carries out the process of feature extraction with example above-mentioned: for a certain normal pictures,
22 normalization segments, the convolutional neural networks that trained corresponding parameter is respectively adopted carry out feature extraction, to obtain 22
A feature vector executes each normal pictures as the reference feature vector of subsequent comparison in this way.As shown in Figure 2
Exemplary architecture in, the output of full articulamentum fc2 is 14 neurons, thus the obtained each fixed reference feature of feature extraction to
Amount is the one-dimensional vector of 14 elements.To each normal pictures, belonging to shoes model have unique identification number, by standard
The identification number of the reference feature vector of picture shoes model corresponding with normal pictures is saved in feature database after being associated.Simultaneously
22 1 × 14 reference feature vectors of each normal pictures may be constructed the eigenmatrix for describing the normal pictures, i.e.,
An eigenmatrix in feature database corresponds to a normal pictures.
At this point, it has been basically completed the process of the feature database to inventory's shoes model foundation searching, managing, the spy in feature database
For comparison when sign matrix can be used to retrieve.
When carrying out the retrieving of step S6, the comparison that same classifier carries out feature vector is preferably sentenced using SVM.Together
Sample, SVM sentences same classifier and is also required to be trained.The training process of same classifier is sentenced to SVM below based on example above-mentioned
It illustrates:
The reference feature vector generated after convolutional neural networks carry out feature extraction is selected, same shoes model will be belonged to
Corresponding position and two reference feature vectors of level carry out being spliced to form a positive example sample vector, by corresponding position and layer
The reference feature vector that secondary two are not belonging to same shoes model carries out being spliced to form a negative example sample vector.With reference to Fig. 1, example
Such as two reference feature vectors, all corresponding to level size location point P1 at, (segment herein can be with for the segment of L1
It is the same segment in same normal pictures；Be also possible to the segment of corresponding position and level in two normal pictures, and this two
A normal pictures belong to same shoes model), the two reference feature vectors are spliced into a positive example sample vector.To every
One normal pictures obtain r positive example sample vector and r negative example sample vectors；Using each normal pictures corresponding position and
The positive example sample vector of level and negative example sample vector go that the SVM is trained to sentence same classifier, to obtain r with different
The SVM of parameter sentences same classifier；Wherein r is the quantity of the normalization segment in a normal pictures, and r >=2 are (because of a normal pictures
Normalization segment include at least a local segment and a global segment)；Using example above-mentioned, then r=22 are obtained
SVM with different parameters sentences same classifier, is also considered as 22 different SVM and sentences same sub-classifier.Every SVM sentences with son
Classifier is respectively used to reference feature vector to corresponding position and level and feature vector to be compared is made whether to belong to one
The differentiation of shoes model, can refer to Fig. 3, S1, S2 ... to S22, i.e. representing 22 SVM sentences same sub-classifier.
It continues with based on example above-mentioned and illustrates to sentence using trained SVM the mistake that same classifier carries out assisted retrieval
Journey, as shown in figure 3, the picture to be retrieved for one, is obtaining 22 features to be compared after the processing of step S1~S4
Vector, the feature that this 22 feature vectors equally also can be configurable for describing the corresponding normal pictures of picture to be retrieved describe square
Battle array, all eigenmatrixes stored in feature Description Matrix and feature database to be compared are all compared respectively, are compared
When specifically between matrix two corresponding feature vectors be compared two-by-two respectively.In a preferred embodiment, feature vector
Comparison be then exactly that the progress of same classifier is sentenced by SVM.Due to each feature vector Y1 to Y22 to be compared correspond in
The segment of some level size at some location point, while it is also respectively correspondingly by some position that each SVM, which sentences same sub-classifier,
The feature vector training of the segment of some level size is formed at setting, therefore, 22 different parameters that aforementioned training is formed
SVM to sentence same sub-classifier S1 to S22 be also the feature vector Y1 to Y22 to be compared with 22 there are one-to-one relationship, than
This sentences same sub-classifier using which SVM when to which feature vector, this is fixed and known.
With continued reference to Fig. 3, for example, the spy of a certain normal pictures of representative in 22 feature vectors and feature database to be compared
The process that sign matrix is compared: by a feature vector Y1 to be compared and corresponding reference feature vector in this feature matrix
It carries out being spliced to form a splicing vector, the SVM which inputs corresponding parameter sentences in same sub-classifier S1, is made whether
Belong to the differentiation of same shoes model, if so, output binary number 1；If otherwise exporting binary number 0.Then this 22 it is to be compared
Feature vector be compared with an eigenmatrix after, 22 SVM sentence same sub-classifier and export 22 bits in total,
Constitute a binary sequence.After all compare with all eigenmatrixes in feature database, multiple binary systems are formed
Sequence.Finally to these binary sequences, 1 number in each binary sequence is counted respectively, can also further ask 1
The shared ratio of number, the highest several former binary sequence of ratio can know that it is compared with which eigenmatrix,
Therefore can obtain the eigenmatrix in the feature database corresponding to the several former binary sequence, and these eigenmatrixes be all with
The identification number of affiliated shoes model is associated, can also export the identification number of shoes model the most matched.Alternatively, search result
Return is to be arranged with the descending of matching degree, rather than only return most matched one or several.
In further embodiments, can also yet further by the identification number returned call corresponding shoes model picture and
The information such as supplier are supplied to user.Alternatively, user can voluntarily go in their database to call according to the identification number of return.?
This is no longer limited.
Another embodiment of the present invention additionally provides a kind of shoes model index management system, including model data is deposited
Reservoir, computer readable storage medium and processor are stored with computer program on the computer readable storage medium, institute
Processor is stated for executing the step of computer program is to realize aforementioned shoes model index management method.Wherein, pattern number
For the eigenmatrix of comparison when according to memory for memory scan.
The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be said that
Specific implementation of the invention is only limited to these instructions.For those skilled in the art to which the present invention belongs, it is not taking off
Under the premise of from present inventive concept, several equivalent substitute or obvious modifications can also be made, and performance or use is identical, all answered
When being considered as belonging to protection scope of the present invention.
Claims (9)
1. a kind of shoes model index management method, comprising the following steps:
S1, the shoes model picture that multiple are contained with shoeprints pre-process, and obtain corresponding normal pictures, wherein all
Normal pictures have consistent coloration, size and sole direction；
S2, to each normal pictures, taken by preset position and level, obtain multiple and different sampling segments；It is described
Take by preset position and level includes: that multiple location points are marked out in the sole region of normal pictures；Then needle
To each position point, the segment of multiple and different sizes is all taken centered on the location point；
S3, to each normal pictures, the multiple sampling segments formed in step S2 and the normal pictures are uniformly zoomed to one
Predetermined size obtains multiple normalization segments；
S4, it each normalization segment is input in the convolutional neural networks of correspondence parameter trained in advance carries out feature extraction,
To obtain the corresponding reference feature vector of each normalization segment；
S5, the corresponding reference feature vector of each normalization segment and the affiliated shoes model of the normal pictures by each normal pictures
Unique identifying number be associated with and be stored in a feature database；Multiple reference feature vectors of same normal pictures constitute the standard drawing
The eigenmatrix of piece；
S6, step S1 to S4 is executed to given picture to be retrieved, obtains multiple feature vectors to be compared；It will be each described to be compared
Feature vector is compared respectively with the reference feature vector of corresponding position and level in eigenmatrix each in feature database to test
Whether card belongs to same shoes model；
S7, to the knot of the multiple feature vector to be compared and reference feature vector comparison corresponding in all eigenmatrixes
Fruit is counted, to return and the most matched shoes model of sole in picture to be retrieved.
2. shoes model index management method as described in claim 1, it is characterised in that: the number of the location point of mark is
N, each position point take m different size of segments, so that every normal pictures obtain m × n different sampling segments, warp
Every normal pictures correspondence obtains m × n+1 normalization segment after step S3；M >=1, n >=1.
3. shoes model index management method as claimed in claim 2, it is characterised in that: further include to corresponding described in step S4
The step of convolutional neural networks of parameter are trained: all normalization segments of multiple normal pictures are chosen, by every standard
The corresponding position of picture and the normalization segment of level are divided into one group, while the corresponding normalization segment of each normal pictures itself is made
It is one group, forms m × n+1 group training data, removes to be respectively trained the mind of the convolution with same framework with m × n+1 group training data
Through network, m × n+1 set network parameter is obtained；
Every suit network parameter is correspondingly used for carrying out the corresponding position of all normal pictures and the normalization segment of level special
Sign is extracted.
4. shoes model index management method as claimed in claim 3, it is characterised in that: described when carrying out the feature extraction
The framework of convolutional neural networks successively includes: by input terminal to output end
Multiple convolutional layers being sequentially connected in series；
It is connected to the first full articulamentum of the last layer convolutional layer and is connected to the second full articulamentum of layer convolutional layer second from the bottom；
It is connected to the full articulamentum of third of the output end of the first and second full articulamentums simultaneously；
Wherein, the output of the full articulamentum of the third is the reference feature vector.
5. shoes model index management method as described in claim 1, it is characterised in that: step S6 is specifically included: will respectively to than
Feature vector and the corresponding position of normal pictures each in feature database and the reference feature vector of level are spliced respectively, shape
Sentence in same classifier at sequentially inputting after multiple splicing vectors to the SVM trained in advance, with differentiate constituted it is each it is described splice to
Whether the feature vector to be compared and corresponding reference feature vector of amount, which belong to same shoes model and export, represents sentencing for yes/no
Other result.
6. shoes model index management method as claimed in claim 5, it is characterised in that: further include SVM is sentenced same classifier into
The step of row training: using reference feature vector obtained in step S4, the corresponding position and level of same shoes model will be belonged to
Two reference feature vectors carry out being spliced to form a positive example sample vector, two of corresponding position and level are not belonging to same
The reference feature vector of one shoes model carries out being spliced to form a negative example sample vector；To obtain r to each normal pictures
A positive example sample vector and r negative example sample vectors；Using each normal pictures corresponding position and level positive example sample to
Amount and negative example sample vector go that the SVM is trained to sentence same classifier, so that obtaining the r SVM with different parameters sentences same point
Class device；Wherein r is the quantity of the normalization segment in a normal pictures, r >=2；
Every SVM sentences same classifier and is based respectively on respective parameter, reference feature vector to corresponding position and level and to than
Feature vector is made whether to belong to the differentiation of a shoes model.
7. such as shoes model index management method described in claim 5 or 6, it is characterised in that: a SVM sentences same classifier and often sentences
Other splicing vector then exports a bit, 1 indicate to differentiate result be it is yes, 0 indicates that differentiation result is no；Then: described
The fixed reference feature of corresponding position and level in all feature vectors to be compared of picture to be retrieved and an eigenmatrix of feature database
Vector is sentenced after same classifier differentiated via corresponding SVM respectively, and the output that r all SVM sentences same classifier constitutes one
A r of binary sequence；After all eigenmatrixes in feature database are all compared, v r binary system sequences are generated in total
Column, wherein v is characterized the quantity of eigenmatrix in library, v >=1；
Step S7 is specifically included: being counted the number of binary number 1 respectively to each r of binary sequence, is then selected v two
1 the largest number of k first, k >=1 in system sequence；K name two before being generated when verifying will be compared with the picture to be retrieved
The identification number of shoes model belonging to the k eigenmatrix difference of system sequence is returned as search result.
8. a kind of shoes model index management system, it is characterised in that: including model data memory, computer readable storage medium
And processor, be stored with computer program on the computer readable storage medium, the processor by execute it is described based on
The step of calculation machine program is to realize claim 1 the method.
9. shoes model index management system as claimed in claim 8, it is characterised in that: the model data memory is for depositing
The eigenmatrix formed in storage step S5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710249557.2A CN107066586B (en) | 2017-04-17 | 2017-04-17 | Shoes model index management method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710249557.2A CN107066586B (en) | 2017-04-17 | 2017-04-17 | Shoes model index management method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107066586A CN107066586A (en) | 2017-08-18 |
| CN107066586B true CN107066586B (en) | 2019-07-05 |
Family
ID=59600103
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710249557.2A Active CN107066586B (en) | 2017-04-17 | 2017-04-17 | Shoes model index management method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107066586B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107590690B (en) * | 2017-09-05 | 2021-02-23 | 创新先进技术有限公司 | Data processing method and device and server |
| CN107808126A (en) * | 2017-10-11 | 2018-03-16 | 苏州科达科技股份有限公司 | Vehicle retrieval method and device |
| US10825071B2 (en) * | 2018-04-09 | 2020-11-03 | International Business Machines Corporation | Adaptive multi-perceptual similarity detection and resolution |
| CN113505204B (en) * | 2021-09-09 | 2021-12-07 | 腾讯科技（深圳）有限公司 | Recall model training method, search recall device and computer equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104537393A (en) * | 2015-01-04 | 2015-04-22 | 大连理工大学 | Traffic sign recognizing method based on multi-resolution convolution neural networks |
| CN104572965A (en) * | 2014-12-31 | 2015-04-29 | 南京理工大学 | Search-by-image system based on convolutional neural network |
| CN104992191A (en) * | 2015-07-23 | 2015-10-21 | 厦门大学 | Image classification method based on deep learning feature and maximum confidence path |
| WO2016075095A2 (en) * | 2014-11-10 | 2016-05-19 | Ventana Medical Systems, Inc. | Using contextual information to classify nuclei in histology images |
-
2017
- 2017-04-17 CN CN201710249557.2A patent/CN107066586B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016075095A2 (en) * | 2014-11-10 | 2016-05-19 | Ventana Medical Systems, Inc. | Using contextual information to classify nuclei in histology images |
| CN104572965A (en) * | 2014-12-31 | 2015-04-29 | 南京理工大学 | Search-by-image system based on convolutional neural network |
| CN104537393A (en) * | 2015-01-04 | 2015-04-22 | 大连理工大学 | Traffic sign recognizing method based on multi-resolution convolution neural networks |
| CN104992191A (en) * | 2015-07-23 | 2015-10-21 | 厦门大学 | Image classification method based on deep learning feature and maximum confidence path |
Non-Patent Citations (1)
| Title |
|---|
| 基于卷积神经网络的多标签图像自动标注;黎健成 等;《计算机科学》;20160715;第43卷(第7期);第41-45页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107066586A (en) | 2017-08-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110825899B (en) | Clothing image retrieval method integrating color features and residual network depth features | |
| CN107066586B (en) | Shoes model index management method and system | |
| CN107679250A (en) | A kind of multitask layered image search method based on depth own coding convolutional neural networks | |
| CN107403424A (en) | A kind of car damage identification method based on image, device and electronic equipment | |
| CN107358596A (en) | A kind of car damage identification method based on image, device, electronic equipment and system | |
| CN104281572B (en) | A kind of target matching method and its system based on mutual information | |
| CN110287873A (en) | Noncooperative target pose measuring method, system and terminal device based on deep neural network | |
| Zhu et al. | Semi-supervised center-based discriminative adversarial learning for cross-domain scene-level land-cover classification of aerial images | |
| CN103186538A (en) | Image classification method, image classification device, image retrieval method and image retrieval device | |
| WO2023173599A1 (en) | Method and apparatus for classifying fine-granularity images based on image block scoring | |
| CN108415937A (en) | A kind of method and apparatus of image retrieval | |
| Shen et al. | A large benchmark for fabric image retrieval | |
| CN111967527B (en) | Peony variety identification method and system based on artificial intelligence | |
| EP4070207A1 (en) | Systems and methods for product identification using image analysis from image mask and trained neural network | |
| CN109993187A (en) | A kind of modeling method, robot and the storage device of object category for identification | |
| CN108280469A (en) | A kind of supermarket's commodity image recognition methods based on rarefaction representation | |
| CN110147780A (en) | The landform recognition methods of real-time field robot and system based on level landform | |
| CN114511452A (en) | Remote sensing image retrieval method integrating multi-scale cavity convolution and triple attention | |
| CN113379504A (en) | Commodity information search recommendation method, system and computer storage medium | |
| CN107533547A (en) | Product index editing method and its system | |
| CN108229505A (en) | Image classification method based on FISHER multistage dictionary learnings | |
| Hu et al. | Hierarchical attention vision transformer for fine-grained visual classification | |
| CN112699261A (en) | Automatic clothing image generation system and method | |
| CN111914772A (en) | Method for identifying age, and training method and device of age identification model | |
| Laaksonen et al. | Picsom-a framework for content-based image database retrieval using self-organizing maps |
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 |