CN106504064A - Clothes classification based on depth convolutional neural networks recommends method and system with collocation - Google Patents
Clothes classification based on depth convolutional neural networks recommends method and system with collocation Download PDFInfo
- Publication number
- CN106504064A CN106504064A CN201610941721.1A CN201610941721A CN106504064A CN 106504064 A CN106504064 A CN 106504064A CN 201610941721 A CN201610941721 A CN 201610941721A CN 106504064 A CN106504064 A CN 106504064A
- Authority
- CN
- China
- Prior art keywords
- collocation
- garment image
- picture
- neural networks
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Finance (AREA)
- Accounting & Taxation (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Library & Information Science (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Image Analysis (AREA)
Abstract
The present invention proposes a kind of clothes classification based on depth convolutional neural networks and recommends method and system, method to include with collocation:By adding batch normalization, improving inception structures, add redundancy grader and improve original GoogleNet convolutional neural networks, feature extraction is carried out to garment image, the classification results of garment image are obtained;Carry out the picture augmentation of multi-method combination to storehouse training set of arranging in pairs or groups, and garment image is carried out distorting, is overturn, color notation conversion space process, train improved GoogleNet convolutional neural networks disaggregated model;The similar single product in suit storehouse and its collocation being found, identification information being generated to every garment image, similar pictures are obtained by contrasting identification information, garment coordination recommendation is comprehensively provided according to the corresponding sex of garment image, style, function information.The present invention can recommend corresponding garment coordination suggestion according to the garment image of input to consumer, high have the advantages that speed block, accuracy.
Description
Technical field
The present invention relates to Image Classfication Technology field, more particularly to a kind of clothes classification based on depth convolutional neural networks
Recommend method and system with collocation.
Background technology
Image recognition is that computer vision field is carried out obtaining, analyze, process, recognizing to image, to obtain automatically target
A kind of technology of perpetual object principal character in image, in many research fields such as object positioning, object classification, gesture recognitions all
There is highly important effect.Under many application backgrounds such as industrial quarters, recognition of face, commercial product recommending, target following, body-sensing detection
Also there are outstanding performance and high big commercial value.In fields such as artificial intelligence field, such as unmanned, full-automatic machine people
There is good development prospect.In the process of clothes identification, it is preferred that emphasis is the color of clothes, version pattern formula, striped flower pattern, style
The extraction of the features such as style.
Image recognition original adoption histograms of oriented gradients, local binary patterns, feature and Scale invariant features transform etc.
Method, refined crawl and tentatively are described by being input in grader and complete pattern classification.But this kind of method is to artificial
A kind of classification learning of feature crawl, is based entirely on mankind's recognition capability and systematic knowledge, and nicety of grading is almost completely dependent on
The manual features extraction level of system is set up, so application must be very fine, data scale is little, Generalization Ability is poor, it is difficult to
Many scenes, wide category development and application.
Deep learning to some extent solves problem above, and the neutral net of intensification has a large amount of intermediate hidden layers
The feature extraction of deep neural network and semantic conversion capability, it is ensured that feature extraction is more essential, more abstract, more meet the mankind recognizes
Know, so as to preferably be showed;Using hierarchical structure, under unsupervised mode of learning, " sequentially initializing " is realized, by
Level abstract characteristics, greatly reduce the difficulty of depth network abstraction training.This method has become and is substantially better than other solution party
The optimal selection of case, has surmounted the general human-subject test of the mankind at present completely.Simultaneously corresponding hardware platform is also obtained
High speed development, completing training using the graphics calculations ability of GPU can shorten more than half by the calculating time.
The growth of popularization and electric business along with internet, increasing people select shopping online, Chinese net in 2015
Network shopping marketplace transaction size 3.8 trillion, dress ornament class accounting surpass 20%, as starting earliest, largest Clothing Industry, net
, also more than three one-tenth, consumer is still in rising trend for the demand of purchase clothing online for purchase permeability.In this context, based on known
The Classification and Identification of clothes and collocation are recommended significant.On the one hand single product picture of consumer's offer can be directed to consumption
Person recommends which to arrange in pairs or groups, and helps consumer to be more readily available collocation strategy, saves the time and efforts of consumer;On the other hand,
The photo that shopping website is provided for consumer, can not only find similar single product, and the collocation that can more obtain consumer is inclined
Good, so as to provide recommendation, expand sales volume, reach doulbe-sides' victory.And do not have related art scheme at present and can realize object above, i.e.,
Provide the consumer with collocation to recommend.
Content of the invention
It is contemplated that at least solving one of above-mentioned technical problem.
For this purpose, it is an object of the present invention to proposing a kind of clothes classification based on depth convolutional neural networks and collocation
Recommendation method, the method can recommend corresponding garment coordination suggestion according to the garment image of input to consumer, with speed
The high advantage of block, accuracy.
Further object is that propose a kind of clothes classification based on depth convolutional neural networks pushing away with collocation
Recommend system.
To achieve these goals, the embodiment of first aspect present invention discloses a kind of based on depth convolutional neural networks
Clothes classification with collocation recommendation method, comprise the following steps:S1:By adding batch normalization, improving inception knots
Structure, adds redundancy grader and improves original GoogleNet convolutional neural networks, to carry out feature extraction to garment image, obtain
The classification results of the garment image;S2:The picture augmentation that multi-method combination is carried out to storehouse training set of arranging in pairs or groups, and to the clothes
Picture is carried out distorting, is overturn, color notation conversion space process, and trains improved GoogleNet convolutional neural networks disaggregated model;
S3:Similar single product and its collocation in suit storehouse is found according to similar pictures lookup algorithm, and mark is generated to every garment image
Information, obtains similar pictures by contrasting the identification information, and obtains clothes figure according to the classification results of the garment image
The corresponding sex of piece, style, function information, and be comprehensively given according to the corresponding sex of the garment image, style, function information
Garment coordination is recommended.
In addition, the clothes classification based on depth convolutional neural networks according to the above embodiment of the present invention and collocation recommendation side
Method can also have the technical characteristic for adding as follows:
In some instances, the S1, further includes:Improved inception construction design methods are based on, using office
The optimum sparsity structure in portion replaces original full connected mode, substitutes traditional convolutional layer increasing amount filtering using multilayer perceptron
Device, and training process is accelerated using dense calculating;Based on inception construction design methods, tied using inception
The compression unit of structure replaces original convolution pond pressure texture, to carry out parallel picture compression;Normalized based on batch
Specified dimension data are all deducted average by data preprocessing method, become the data set that average is 0.
In some instances, the S2, further includes:Based on the training set processing method of picture augmentation, using distortion,
Upset, the color notation conversion space process garment image, and different Enhancement Methods is used in color space, and adopt two angles
The local segment and horizontal reflection of point and central point cuts out mode, and image of clothing is adjusted.
In some instances, the utilization distortion, upset, the color notation conversion space process garment image, including:R leads to
Road is full, discount, and G passages are full, discount, and channel B is full, discount, and adaptive histogram equalization, Retinex contrasts increase
By force, contrast linearly strengthens, contrast linear reduction.
In some instances, the S3, further includes:Mark is generated to every garment image by perceiving hash algorithm
Information, and information contrast is identified to obtain similar pictures by Hamming distance, wherein, simplify via minification, color,
DCT is calculated, diminution DCT, calculating mean value, calculating hash are worth to the identification information;Corresponding according to the garment image
Sex, function and style, select the clothing that arranges in pairs or groups in the storehouse of sex corresponding with the garment image, function and style coupling
Clothes, and similar pictures are found using hash algorithm is perceived in suit storehouse to the garment image being input into, obtain taking for similar pictures
Match somebody with somebody, and the similar pictures for perceiving the single product of hash algorithm search collocation are reused in collocation storehouse, obtain final collocation and select.
Clothes classification based on depth convolutional neural networks according to embodiments of the present invention and collocation recommendation method, with such as
Under beneficial effect:
1st, more accurate clothes nicety of grading can be obtained according to the model of improved depth convolutional neural networks, and is had
There is the expansibility of other field;
2nd, the picture augmentation strategy of combination can be used to process training set, so as to simulate the Unified Field under different shooting conditions
Scape, increase picture training set scale, reach the effect for strengthening training;
3rd, sex, function and the style of input picture can be obtained according to the model of improved depth convolutional neural networks,
Foundation is provided for selecting collocation, high have the advantages that classification accuracy
4th, perception hash algorithm can be utilized to find similar pictures in collocation storehouse according to known picture, and according to collocation storehouse
In collocation be given collocation suggestion;
5th, the picture being input into can be classified, so as to obtain sex, function and the style information of clothes, also obtains which
To in requisition for collocation storehouse;On the other hand the picture to being input into finds similar pictures using hash algorithm is perceived in suit storehouse,
The collocation of similar pictures is obtained, and then is believed that the collocation is applied to the picture of input, on the other hand, more optional in order to obtain
The collocation list product that selects, reuse the similar pictures for perceiving the single product of hash algorithm search collocation, in collocation storehouse so as to obtain more
Plus abundant collocation is selected.
To achieve these goals, the embodiment of second aspect present invention discloses a kind of based on depth convolutional neural networks
Clothes classification with collocation commending system, including:Sort module, the sort module are used for normalizing, changing by adding batch
Enter inception structures, add redundancy grader and improve original GoogleNet convolutional neural networks, to carry out to garment image
Feature extraction, obtains the classification results of the garment image;Processing module, the processing module are used for entering storehouse training set of arranging in pairs or groups
The picture augmentation of row multi-method combination, and the garment image is carried out distorting, is overturn, color notation conversion space process, and train
Improved GoogleNet convolutional neural networks disaggregated model;Collocation recommending module, the collocation recommending module are used for according to similar
Picture searching algorithm finds the similar single product in suit storehouse and its collocation, generates identification information to every garment image, by right
Similar pictures are obtained than the identification information, and the corresponding property of garment image are obtained according to the classification results of the garment image
Not, style, function information, and comprehensively provide garment coordination and push away according to the corresponding sex of the garment image, style, function information
Recommend.
In addition, the clothes classification based on depth convolutional neural networks according to the above embodiment of the present invention recommends system with collocation
System can also have the technical characteristic for adding as follows:
In some instances, the sort module is used for:Improved inception construction design methods are based on, using office
The optimum sparsity structure in portion replaces original full connected mode, substitutes traditional convolutional layer increasing amount filtering using multilayer perceptron
Device, and training process is accelerated using dense calculating;Based on inception construction design methods, tied using inception
The compression unit of structure replaces original convolution pond pressure texture, to carry out parallel picture compression;Normalized based on batch
Specified dimension data are all deducted average by data preprocessing method, become the data set that average is 0.
In some instances, the processing module is used for training set processing method based on picture augmentation, using distorting, turn over
Turn, color notation conversion space processes the garment image, and different Enhancement Methods is used in color space, and adopt two angle points
Mode is cut out with the local segment of central point and horizontal reflection, image of clothing is adjusted.
In some instances, the processing module processes garment image using distortion, upset, color notation conversion space, including:
R passages are full, discount, and G passages are full, discount, and channel B is full, discount, adaptive histogram equalization, Retinex contrasts
Strengthen, contrast linearly strengthens, contrast linear reduction.
In some instances, the collocation recommending module is used for:Every garment image is generated by perceiving hash algorithm
Identification information, and information contrast is identified to obtain similar pictures by Hamming distance, wherein, via minification, color
Simplify, DCT is calculated, reduces DCT, calculate mean value, calculating hash is worth to the identification information;According to the garment image pair
Sex, function and the style that answers, selects collocation in the storehouse of sex corresponding with the garment image, function and style coupling
Clothes, and similar pictures are found using hash algorithm is perceived in suit storehouse to the garment image being input into, obtain similar pictures
Collocation, and the similar pictures for perceiving the single product of hash algorithm search collocation are reused in collocation storehouse, obtain final collocation choosing
Select.
Clothes classification based on depth convolutional neural networks according to embodiments of the present invention and collocation commending system, with such as
Under beneficial effect:
1st, more accurate clothes nicety of grading can be obtained according to the model of improved depth convolutional neural networks, and is had
There is the expansibility of other field;
2nd, the picture augmentation strategy of combination can be used to process training set, so as to simulate the Unified Field under different shooting conditions
Scape, increase picture training set scale, reach the effect for strengthening training;
3rd, sex, function and the style of input picture can be obtained according to the model of improved depth convolutional neural networks,
Foundation is provided for selecting collocation, high have the advantages that classification accuracy
4th, perception hash algorithm can be utilized to find similar pictures in collocation storehouse according to known picture, and according to collocation storehouse
In collocation be given collocation suggestion;
5th, the picture being input into can be classified, so as to obtain sex, function and the style information of clothes, also obtains which
To in requisition for collocation storehouse;On the other hand the picture to being input into finds similar pictures using hash algorithm is perceived in suit storehouse,
The collocation of similar pictures is obtained, and then is believed that the collocation is applied to the picture of input, on the other hand, more optional in order to obtain
The collocation list product that selects, reuse the similar pictures for perceiving the single product of hash algorithm search collocation, in collocation storehouse so as to obtain more
Plus abundant collocation is selected.
The additional aspect and advantage of the present invention will be set forth in part in the description, and partly will become from the following description
Obtain substantially, or recognized by the practice of the present invention.
Description of the drawings
The above-mentioned and/or additional aspect and advantage of the present invention will become from the description with reference to accompanying drawings below to embodiment
Substantially and easy to understand, wherein:
Fig. 1 is the clothes classification based on depth convolutional neural networks according to embodiments of the present invention and collocation recommendation method
Flow chart;
Fig. 2 is the original inception structural representations according to a specific embodiment of the invention;
Fig. 3 is the inception structural representations optimized according to the small-scale convolution of a specific embodiment of the invention;
Fig. 4 is the inception structural representations optimized according to the more small-scale convolution of a specific embodiment of the invention;
Fig. 5 is the schematic flow sheet of the convolutional neural networks picture compression method according to a specific embodiment of the invention;
Fig. 6 is the parallel image compression schematic diagram according to a specific embodiment of the invention;
Fig. 7 is the improved googlenet inception model schematics according to a specific embodiment of the invention;
Fig. 8 is the collocation storehouse list product taxonomic structure schematic diagram according to a specific embodiment of the invention;
Fig. 9 is to strengthen result schematic diagram according to the picture augmentation of a specific embodiment of the invention;
Figure 10 is the model recommended flowsheet schematic diagram based on classification according to a specific embodiment of the invention;
Figure 11 is clothes classification and the collocation result schematic diagram according to a specific embodiment of the invention;
Figure 12 is recommended based on clothes classification and the collocation of perception hash algorithm according to a specific embodiment of the invention
Schematic flow sheet;And
Figure 13 is the clothes classification based on depth convolutional neural networks according to embodiments of the present invention and collocation commending system
Structured flowchart.
Specific embodiment
Embodiments of the invention are described below in detail, the example of the embodiment is shown in the drawings, wherein from start to finish
Same or similar label represents same or similar element or the element with same or like function.Below with reference to attached
The embodiment of figure description is exemplary, is only used for explaining the present invention, and is not considered as limiting the invention.
In describing the invention, it is to be understood that term " " center ", " longitudinal direction ", " horizontal ", " on ", D score,
The orientation of instruction such as "front", "rear", "left", "right", " vertical ", " level ", " top ", " bottom ", " interior ", " outward " or position relationship are
Based on orientation shown in the drawings or position relationship, it is for only for ease of the description present invention and simplifies description, rather than indicate or dark
Show that the device or element of indication must be with specific orientation, with specific azimuth configuration and operation, therefore it is not intended that right
The restriction of the present invention.Additionally, term " first ", " second " are only used for describing purpose, and it is not intended that indicating or implying relative
Importance.
In describing the invention, it should be noted that unless otherwise clearly defined and limited, term " installation ", " phase
Even ", " connection " should be interpreted broadly, for example, it may be being fixedly connected, or being detachably connected, or be integrally connected;Can
Being to be mechanically connected, or electrically connect;Can be joined directly together, it is also possible to be indirectly connected to by intermediary, Ke Yishi
The connection of two element internals.For the ordinary skill in the art, above-mentioned term can be understood at this with concrete condition
Concrete meaning in invention.
Classify below in conjunction with the Description of Drawings clothes based on depth convolutional neural networks according to embodiments of the present invention and take
With recommendation method and system.
Fig. 1 is the clothes classification based on depth convolutional neural networks according to an embodiment of the invention and collocation recommendation side
The flow chart of method.As shown in figure 1, the method is comprised the following steps:
Step S1:By adding batch normalization, improving inception structures, interpolation redundancy grader improves original
GoogleNet convolutional neural networks, to carry out feature extraction to garment image, obtain the classification results of garment image.In other words,
The step builds the process of picture classification model based on improvement GoogleNet depth convolutional neural networks, based on improved depth
The clothes sorting algorithm of degree convolutional neural networks, in conjunction with the garment coordination and clothes category dataset of network crawl, for research
Object fast and effeciently sets up depth convolutional neural networks disaggregated model, so as to predict the category feature of input picture, specially:
By adding batch normalization, improving inception structures, add redundancy grader and improve original GoogleNet convolutional Neurals
Network, for extracting feature for picture, reaches classification results more accurately and quickly for garment image feature, so as to
To the classification information of clothes, the i.e. classification results of garment image.
In one embodiment of the invention, step S1, further includes:It is based on improved inception structure designs
Method, replaces original full connected mode using the sparsity structure of local optimum, substitutes traditional convolutional layer using multilayer perceptron
Increasing amount wave filter, is avoided redundancy to greatest extent, and training process is accelerated using dense calculating, compared to individual layer convolution
Model, multilayer perceptron can carry out more loaded down with trivial details calculating to each local sensing domain, so as to approach the feature for intentionally getting
As a result;Inception construction design methods are based on, and original convolution pond are replaced using the compression unit of inception structures
Pressure texture, to carry out parallel picture compression, so as to avoid with the intensification of transmission depth, the sign of nonlinear activation layer is special
Levy the Calculation bottleneck that quantity should be successively decreased therewith;The normalized data preprocessing method of batch is based on, will be whole for specified dimension data
Average is deducted, becomes the data set that average is 0, such that it is able to improve the validity in random assortment face, all may be used for ensureing per layer
With using higher learning rate, save computing resource, strengthen network replicability, evade the extreme situation of weight of local optimum with
And reduce picture distortion;In fabric increase redundancy grader, can make bottom can training parameter in gradient descent procedures
More fully trained.Redundancy grader can play a part of to help train, and its principle is to play to a certain extent
Normalized effect, therefore adds batch normalized structure before grader or Dropout processes structure and can play
Greater role.Redundancy grader really plays a role after network structure convergence of approximation simultaneously, is arranged using isolation at the training initial stage
Apply, computing resource can be saved.So embodiments of the invention add redundancy grader in fabric, normalize with batch
Structural interaction so that training is more abundant.
Step S2:The picture augmentation that multi-method combination is carried out to storehouse training set of arranging in pairs or groups, and garment image is carried out distorting, turned over
Turn, color notation conversion space is processed, and trains improved GoogleNet convolutional neural networks disaggregated model.In other words, the step is
Based on training set build marking model process, specially:The picture augmentation that multi-method combination is carried out to storehouse training set of arranging in pairs or groups, with
Improve the accuracy and speed of picture classification;Picture is processed using methods such as distortion, upset, color notation conversion spaces, to increase picture
Training set scale, while simulating the unified scene under different shooting conditions, reaches the effect for strengthening training.Number may finally be enriched
According to the accuracy and speed for collecting and improving classification.
In one embodiment of the invention, step S2, further includes:First, for example the python nets by increasing income
Network reptile downloads clothes list product and collocation suit from shopping website, and by single product sophisticated category;Training set based on picture augmentation
Processing method, using distortion, upset, the color notation conversion space process garment image, while increase picture training set scale,
The unified scene under different shooting conditions can also be simulated, the effect for strengthening training is reached;And in color space using different
Enhancement Method, and the local segment and horizontal reflection using two angle points and central point cuts out mode, and image of clothing is carried out
Adjustment, with the quality of training for promotion collection, improves accuracy of identification.Wherein, clothes are processed using distortion, upset, color notation conversion space
Picture, including:R passages are full, discount, and G passages are full, discount, and channel B is full, discount, adaptive histogram equalization,
Retinex contrasts strengthen, and contrast linearly strengthens, contrast linear reduction.
Step S3:Similar single product and its collocation in suit storehouse is found according to similar pictures lookup algorithm, to every clothes
Picture generates identification information, obtains similar pictures by contrasting identification information, and is taken according to the classification results of garment image
The corresponding sex of dress picture, style, function information, and be comprehensively given according to the corresponding sex of garment image, style, function information
Garment coordination is recommended.In other words, the step builds the process for recommending collocation model based on similar pictures and garment feature, specifically
For:Similar single product and its collocation in suit storehouse is found using similar pictures lookup algorithm, and " fingerprint " is generated (i.e. to every pictures
Identification information), similar pictures can be obtained by fingerprint (identification information) contrast, high-volume picture be processed using the method simple
Quickly, can according to the picture fast searching similar pictures of input, and obtained according to picture classification picture sex, style, work(
Energy informix provides garment coordination recommendation.
In one embodiment of the invention, step S3, further includes:By perceiving hash algorithm to every clothes figure
Piece generates identification information (such as " fingerprint "), and is identified information (fingerprint) contrast by Hamming distance to obtain similar diagram
Piece.Wherein, simplify via minification, color, DCT is calculated, reduces DCT, calculate mean value, calculating hash is worth to mark letter
Breath.Simple and quick using the method process high-volume picture, can be according to the picture fast searching similar pictures of input.It is based on and changes
The GoogleNet network trainings for entering go out clothes disaggregated model, due to training set classification fine, so for the picture of input can be with
Sex, function and style clear from the clothes.Further, according to the corresponding sex of garment image, function and style,
The clothes that arranges in pairs or groups is selected in the storehouse of sex corresponding with garment image, function and style coupling, and the garment image being input into is made
Similar pictures are found with hash algorithm is perceived in suit storehouse, obtain the collocation of similar pictures, then further it is believed that the collocation is fitted
For the picture being input into.Further, in order to obtain the single product of more selectable collocation, perception is reused in collocation storehouse and breathed out
The similar pictures of uncommon algorithm search collocation list product, the collocation so as to more be enriched select, that is, obtain final collocation and select,
And recommend consumer.
To sum up, the cardinal principle of the method for the embodiment of the present invention is summarized as:Based on training set and improved GoogleNet volume
Product neutral net obtains disaggregated model, based on the similar single product and its collocation that perceive in hash algorithm searching suit storehouse, and according to
Picture sex that classification is obtained, style, function information comprehensively provide garment coordination recommendation.The method can be based on known clothes
Classification and Identification and collocation are recommended consumer's collocation and advise and help businessman to promote sale.
In order to make it easy to understand, below in conjunction with accompanying drawing, being carried out in detail to the method for the above embodiment of the present invention with specific embodiment
Thin description.
1st, with regard to the description of improved GoogleNet convolutional neural networks structure.Specifically, based on improvement GoogleNet
The picture classification model of depth convolutional neural networks, by adding batch normalization, improving inception structures, adds redundancy
Grader improves original GoogleNet convolutional neural networks, for extracting feature for picture, reaches for garment image feature
Classification results more accurately and quickly, so that obtain the classification information of clothes.
1.1 with regard to improved inception structures description.Specifically, convolutional neural networks refer to process two-dimentional lattice
The multi-layer Network Framework built for the purpose of formula data, by the semanteme of convolutional calculation successively abstract picture different levels, leads to
The training fitting for crossing data set adjusts the inner parameter of each convolution kernel, so as to realize characteristic of division under the conditions of unsupervised
Crawl.Convolutional neural networks are made up of input layer, multilayer convolutional layer, multilayer sub-sampling layer and output layer.Convolutional layer passes through convolution
Verification 2-D data carries out convolutional calculation, and adds bias vector, so as to obtain the Feature Mapping of the feature extraction first time, and will be many
Input of the individual Feature Mapping weighted sum as sub-sampling layer (active coating), so that obtain the new spy through nonlinear characteristic crawl
Levy mapping.The work is repeated several times to obtain from bottom to high-rise semantic abstraction, eventually passes through the activation of multilayer convolution kernel, by feature
It is input in grader to complete to classify and recognize.
Googlenet greatly saves computing resource by the method for reducing parameter, while the new structure list of application
Unit, further expands network size in two dimensions of depth and width, it is proposed that inception structures, using Multilayer Perception
Device replaces the generalized linear structure in traditional convolutional neural networks.As realistic problem is almost non-linear, and Multilayer Perception
Utensil has more preferable non-linear classification, so multilayer perceptron suffers from preferably effect on classification accuracy and convergence rate
Really.
The core of Inception structures is to replace original full connected mode using the sparsity structure of local optimum, with many
Layer perceptron replaces the traditional approach of traditional convolutional layer increasing amount wave filter, avoids redundancy to greatest extent, and utilizes dense calculating
Realize the acceleration of training process.Compared to individual layer convolution model, multilayer perceptron can be carried out more to each local sensing domain
Loaded down with trivial details calculating, so that approach the characteristic results for intentionally getting.Per layer of feature extraction and calculation mode is as follows:
Initially inception versions adopt convolution piecemeal, such as shown in Fig. 2, using the convolution kernel and 3* of 1*1,3*3,5*5
3 maximum pondization operation, by different scales convolution kernel result with pond result together Federated filter, is exported.Due to four
Unit accepts preceding layer output respectively as input, and the amount of calculation that 5*5 convolution kernels bring is huge, so adopting network
The thought of in network, carries out merge operations to input feature value, is carried out again using its powerful nonlinear transformation ability
Processing with filtering, propose all large scale convolution are all carried out substituting with the convolutional layer of 3*3 obtain as shown in Figure 3
Inception structural models, subsequently attempt to more small-scale convolution kernel, obtain inception structures as shown in Figure 4
Model, so that obtain faster convergence rate and convergence precision.Further, using the inception structures shown in Fig. 4 and figure
Pressure texture shown in 5, composition inceptionv4 structures are classified to the data set of ILSVRC-2012, several so as to obtain
The nicety of grading contrast of different convolutional neural networks is as shown in table 1.According to nicety of grading contrast table (table 1), in the present embodiment
In, using two kinds of inception structures of Fig. 4.
Network structure | Top-1error | Top-5error |
Googlenet[5] | 29% | 9.2% |
BN-Inception[9] | 25.2% | 7.8% |
Inception-v3[7] | 21.2% | 5.6% |
Inception-ResNet-v1[8] | 21.3% | 5.5% |
Inception-v4[8] | 20.0% | 5.0% |
Table 1
1.2 with regard to picture compression method description.Specifically, the compression method in training process for picture size includes
Two methods as shown in Figure 5, left side are traditional convolutional neural networks compression method, and right side is by adjusting pondization and convolution order
Dimensional optimization is carried out, it is seen that wherein amount of calculation is huge.And the intensification with transmission depth, the sign of nonlinear activation layer
Feature quantity should be successively decreased to avoid Calculation bottleneck therewith, so the right figure of Fig. 5 is completed and this rule conflict by adjustment order, meeting
Cause the bottleneck of characteristic feature.Therefore, exist in the present embodiment, in conjunction with the inspiration of inception structures, final using such as Fig. 6 institutes
The version that shows.
1.3 with regard to the normalized description of Batch Normalization batches.Specifically, batch normalization is commonly used to
Enter line data set pretreatment, can help to a certain extent restrain.Natural image has very strong local correlations, such as same
Two neighbor pixels of image, exclude border only a few point, and its color character will not saltus step.Convolutional network initialization procedure
In, all train weight to generate using stochastical sampling, so as to cause iteration to lose as race deviates a space center.Therefore will
Certain one-dimensional data all deducts average, becomes the data set that average is 0, can improve the validity in random assortment face.In reality
In operation, Batch Normalization can make improvement in following many aspects and be lifted to convolutional network:
1) ensure per layer and can use higher learning rate.During unused Batch Normalization, each number of plies
Different according to category, corresponding learning rate floats.If same layer has different categories, the learning rate of minimum is selected as whole study
Rate is declined effectively, so as to cause learning rate low with ensureing gradient.Batch Normalization cause each layer adopt
Gradient decline is carried out with higher learning rate.
2) original Local Response Normalization structures are banned.Local Response
Normalization structures do not possess astriction, can replace the structure using Batch Normalization, and save
Save computing resource.
3) Dropout structures are removed.Over-fitting is frequently experienced in the irregular obstacle body of data, and normalized causes data
Category becomes have mark follow, and therefore no longer needs to eliminate over-fitting by Dropout structures, strengthens the replicability of network.
4) L2 weight descent coefficients are cut down.Original L2 weights descent coefficient is in order to the extreme feelings of the weight of evading local optimum
Condition.Batch Normalization can solve this problem, and therefore descent coefficient can be reduced to the 1/5 of initial value.
5) reduce picture distortion to use.Distortion is not recycled to strengthen the Generalization Ability of network, in case original information destruction.
1.4 with regard to redundancy grader description.Specifically, increase redundancy grader in fabric, can make the bottom can
Training parameter is more fully trained in gradient descent procedures.Redundancy grader can play a part of to help train,
Its principle is to serve normalized effect to a certain extent, therefore before grader add batch normalized structure or
Person Dropout processes structure and can play greater role.Redundancy grader real performance after network structure convergence of approximation is made simultaneously
With adopting quarantine measures at the training initial stage, computing resource can be saved.So add redundancy classification herein in fabric
Device, with batch normalization structural interaction so that training is more abundant.
As shown in fig. 7, the googlenet inception network structures for finally improving are obtained, wherein with the figure on the right side of Fig. 6
Piece pressure texture substitutes original maxpool2 structures becomes maxpool2*;Substituted with the picture compression structure on the left of Fig. 6 original
Maxpool3, maxpool4 and avg pool structures become maxpool3*, maxpool4* and avg pool*;
Inception3a, 3b, 3c, 4a, 4b adopt original inception structures, and inception 4c*, 4d*, 4e* are using on the left of Fig. 4
Inception structures;Inception 5a*, 5b*, 5c* are using the inception structures on the right side of Fig. 4.In convolution each time
A batch normalization is executed after operation, and adds redundancy grader.
2nd, with regard to training set acquisition and the description of mark.Specifically, marking model is obtained based on training set, for arranging in pairs or groups
Storehouse training set carries out the picture augmentation of multi-method combination, to improve the accuracy and speed of picture classification.Using distortion, upset, face
The methods such as color space transformation process picture, can increase picture training set scale and can also simulate the unification under different shooting conditions
Scene, reaches the effect for strengthening training, may finally enrich data set and improve the accuracy and speed of classification.
2.1 with regard to obtain training set process description.Specifically, for example, using the python web crawlers that increases income from purchase
Clothes list product and collocation suit are downloaded in thing website, and single product are classified as shown in figure 8, wherein women's dress is divided into 9*9 81 classes, men's clothing totally
It is divided into 7*8 totally 56 classes, Men's Shoes are divided into 6 classes, and women's shoes is divided into 6*4 totally 24 classes, altogether 167 class, every class capturing pictures 1500 or so
Composition collocation storehouse.And in shopping website 1000 set groups of crawl suit into suit storehouse, suit includes clothes and shoes.Using collocation storehouse
Single product in neutralization suit storehouse train improved googlenet networks jointly, and clothes disaggregated model is obtained.
2.2 descriptions for strengthening training set with regard to picture augmentation.Specifically, using the side such as distortion, upset, color notation conversion space
Method processes picture, can increase the unified scene that picture training set scale can also be simulated under different shooting conditions, reach reinforcement
The effect of training.Principal component passage is carried out doubling to strengthen for example with PCA methods, enhancing ratio adopt average for 0 Gauss
Distribution, to simulate the image change under light intensity change.In the present embodiment, different Enhancement Methods are attempted in color space, and is adopted
Mode is cut out with the local segment and horizontal reflection of two angle points and central point, image is adjusted.
1) passage X is filled and is:min(X*2,255);Passage discount is X/2.
2) Retinex contrasts strengthen.
Wherein, S represents original image, and L represents light image, and R represents albedo image, and Retinex assumes S (x, y)=R
Image is gone to log-domain by (x, y) L (x, y), Retinex enhancing, and product switchs to and estimates illumination L from original image S, enter
And therefrom decomposite R.Formula is as follows:
Log (S)=log (RL);LogS=logR+logL;S=r+l,
L=f (s);R=s-f (s).
3) with regard to the description of adaptive histogram equalization.Specifically, if original pixel gray value is r (0≤r≤1), probability
Density is pr(r), after conversion pixel grey scale be s, probability density psS (), transforming function transformation function are T (r), j gray scale layer number of pixels is total
Pixel count is nj, the gray scale exponent number of total pixel number n, k for digital picture,
Then s=T (r), 0≤r≤1;ps(s) ds=pr(r) dr,
Order
Then:
Finally gray value is:Sk=(L-1) × sk.The result of the method and corresponding picture is for example shown in Fig. 9.
2.3 classification results are analyzed.Specifically, Caffe is the deep learning platform that increases income, and platform is developed by C++, built
One clear, efficiently, expression formula, modular framework, possess python, matlab and command line interface, and and
Turn round and look at the acceleration optimization and switching of GPU and CPU.Modular configuration settings cause define network structure more convenient, intuitive is more
By force;Adopt cuDNN simultaneously, optimize structure and calculating, greatly accelerate the process of learning training.So, in the present embodiment, adopt
Caffe Computational frames are carrying out the training that builds with model of structure.
Using improved googlenet deep learnings network to training set training pattern, 30 circuit training, study are set
From final clothes disaggregated model is obtained, the picture according to input can obtain the picture classification to rate.Classifying quality such as 2 institute of table
Show, it can be seen that based on the precision that improved googlenet models and picture augmentation can improve picture recognition, final classification essence
Degree can reach 74.81%.
Network structure | Accuracy | Top-5Accuracy |
Original googlenet | 70.08% | 96.98% |
Picture augmentation | 71.89% | 96.77% |
Improve inception | 74.22% | 97.58% |
Inception-bn | 74.81% | 97.76% |
Table 2
3. with regard to the description of the clothes recommendation process based on clothes classification and picture analyzing.Specifically, based on according to similar
Picture recommends collocation model with garment feature, finds the similar single product in suit storehouse using similar pictures lookup algorithm and its takes
Match somebody with somebody, using the method process high-volume picture simple and quick, can according to input picture fast searching similar pictures, and according to
Picture sex that picture classification is obtained, style, function information comprehensively provide garment coordination recommendation.
3.1 with regard to by perceiving the description that hash algorithm searches for the process of similar pictures.Specifically, calculated by perceiving Hash
Method generates " fingerprint " (i.e. identification information) to every pictures, can obtain similar pictures by fingerprint contrast, at the method
Reason high-volume picture is simple and quick, can be according to the picture fast searching similar pictures of input.Algorithm flow is as follows:
1) minification:Picture is narrowed down to 8*8, only retains the structure and light and shade information of picture, removed thin in picture
Section, reduces the difference that different proportion or size are brought.
2) color simplifies:The colour picture of 8*8 is converted into 64 grades of gray scales.
3) DCT is calculated:Dct transform is carried out to picture, the coefficient matrix of 32*32 is obtained.
4) DCT is reduced:The 8*8 matrixes in the coefficient matrix upper left corner of the 32*32 for obtaining, the part present minimum in picture
Frequency.
5) mean value is calculated:Calculate DCT averages.
6) hash values are calculated:According to 8*8DCT matrixes, " 1 " is set to more than or equal to DCT averages, less than " 0 " is set to, so as to obtain
To 64 hash values, and use hexadecimal representation.
7) Hamming distance obtains similar pictures:Calculate the Hamming distance of input picture fingerprint and fingerprint in fingerprint base, take away from
It is similar pictures from the picture less than or equal to 10.
3.2 based on the description for improving the process that the classification of Googlenet clothes obtains recommending.Specifically, according to mentioned above
Improved googlenet networks can train clothes disaggregated model, due to training set classification finely, so the figure for input
Piece can be clear from the sex of the clothes, function and style, such that it is able to select the clothes that arranges in pairs or groups with which in other storehouses.Such as
Input picture obtains picture sex for women, and style is Korea Spro's version, and function is cotta, then according to the picture of input recommendation sex can be
Women, style skirt and shoes for Korea Spro's version, select equally single product from corresponding storehouse at random, and handling process is as shown in Figure 10,
Arranging effect as shown in figure 11 can finally be obtained.
3.3 are classified based on clothes and perceive the description that hash algorithm obtains collocation strategic process.Specifically, due to basis point
Class chooses the single product of collocation at random from collocation storehouse certain randomness, it is contemplated that the arranging effect such as color, using shopping website under
Carry suit and constitute suit storehouse, the picture being input into is classified first, so as to obtain sex, function and the style information of clothes,
Also obtain its in requisition for collocation storehouse;On the other hand the picture to being input into finds phase using hash algorithm is perceived in suit storehouse
Like picture, the collocation of similar pictures is obtained, then further it is believed that the collocation is applied to the picture of input.More optional in order to obtain
The collocation list product that selects, reuse the similar pictures for perceiving the single product of hash algorithm search collocation, in collocation storehouse so as to obtain more
Plus abundant collocation is selected.Handling process is for example shown in Figure 12.
To sum up, the clothes classification based on depth convolutional neural networks according to embodiments of the present invention and collocation recommendation method,
There is following beneficial effect:
1st, more accurate clothes nicety of grading can be obtained according to the model of improved depth convolutional neural networks, and is had
There is the expansibility of other field;
2nd, the picture augmentation strategy of combination can be used to process training set, so as to simulate the Unified Field under different shooting conditions
Scape, increase picture training set scale, reach the effect for strengthening training;
3rd, sex, function and the style of input picture can be obtained according to the model of improved depth convolutional neural networks,
Foundation is provided for selecting collocation, high have the advantages that classification accuracy
4th, perception hash algorithm can be utilized to find similar pictures in collocation storehouse according to known picture, and according to collocation storehouse
In collocation be given collocation suggestion;
5th, the picture being input into can be classified, so as to obtain sex, function and the style information of clothes, also obtains which
To in requisition for collocation storehouse;On the other hand the picture to being input into finds similar pictures using hash algorithm is perceived in suit storehouse,
The collocation of similar pictures is obtained, and then is believed that the collocation is applied to the picture of input, on the other hand, more optional in order to obtain
The collocation list product that selects, reuse the similar pictures for perceiving the single product of hash algorithm search collocation, in collocation storehouse so as to obtain more
Plus abundant collocation is selected.
Further embodiment of the present invention is proposed a kind of clothes classification based on depth convolutional neural networks and is pushed away with collocation
Recommend system.
Figure 13 is that the clothes classification based on depth convolutional neural networks according to an embodiment of the invention is recommended with collocation
The structured flowchart of system.As shown in figure 13, should be wrapped with collocation commending system 100 based on the classification of the clothes of depth convolutional neural networks
Include:Sort module 110, processing module 120 and collocation recommending module 130.
Wherein, sort module 110 is used for, by adding batch normalization, improving inception structures, adding redundancy classification
Device improves original GoogleNet convolutional neural networks, to carry out feature extraction to garment image, obtains the classification knot of garment image
Really.
Specifically, in one embodiment of the invention, sort module 110 is used for:It is based on improved inception structures
Method for designing, replaces original full connected mode using the sparsity structure of local optimum, substitutes conventional roll using multilayer perceptron
Lamination increasing amount wave filter, and training process is accelerated using dense calculating;Inception construction design methods are based on,
Original convolution pond pressure texture is replaced using the compression unit of inception structures, to carry out parallel picture compression;Base
In the normalized data preprocessing method of batch, specified dimension data are all deducted average, become the data set that average is 0.
Processing module 120 is used for the picture augmentation for carrying out multi-method combination to storehouse training set of arranging in pairs or groups, and garment image is entered
Row distortion, upset, color notation conversion space process, and train improved GoogleNet convolutional neural networks disaggregated model.
Specifically, in one embodiment of the invention, processing module 120 is used for the process of the training set based on picture augmentation
Method, processes garment image using distortion, upset, color notation conversion space, and uses different Enhancement Methods in color space, and
Mode is cut out using the local segment and horizontal reflection of two angle points and central point, image of clothing is adjusted.More have
Body ground, processing module 120 process garment image using distortion, upset, color notation conversion space, including:R passages are full, discount, G
Passage is full, discount, and channel B is full, discount, and adaptive histogram equalization, Retinex contrasts strengthen, and contrast is linear
Strengthen, contrast linear reduction.
Collocation recommending module 130 is used for finding the similar single product in suit storehouse according to similar pictures lookup algorithm and its takes
Match somebody with somebody, identification information is generated to every garment image, similar pictures, and dividing according to garment image is obtained by contrasting identification information
Class result obtains the corresponding sex of garment image, style, function information, and according to the corresponding sex of garment image, style, function
Informix provides garment coordination recommendation.
Specifically, in one embodiment of the invention, collocation recommending module 130 is used for:By perceiving hash algorithm pair
Every garment image generates identification information, and is identified information contrast to obtain similar pictures by Hamming distance, wherein, warp
Simplified by minification, color, DCT is calculated, reduces DCT, calculate mean value, calculating hash is worth to identification information;According to clothes
The corresponding sex of dress picture, function and style, in the storehouse of sex corresponding with garment image, function and style coupling select to take
The clothes that matches somebody with somebody, and similar pictures are found using hash algorithm is perceived in suit storehouse to the garment image being input into, obtain similar diagram
The collocation of piece, and the similar pictures for perceiving the single product of hash algorithm search collocation are reused in collocation storehouse, obtain final taking
With selection.
It should be noted that the clothes classification based on depth convolutional neural networks of the embodiment of the present invention recommends system with collocation
The clothes classification based on depth convolutional neural networks of the specific implementation of system and the embodiment of the present invention and collocation recommendation method
Specific implementation be similar to, specifically refer to the description of method part, in order to reduce redundancy, here is omitted.
To sum up, the clothes classification based on depth convolutional neural networks according to embodiments of the present invention and collocation commending system,
There is following beneficial effect:
1st, more accurate clothes nicety of grading can be obtained according to the model of improved depth convolutional neural networks, and is had
There is the expansibility of other field;
2nd, the picture augmentation strategy of combination can be used to process training set, so as to simulate the Unified Field under different shooting conditions
Scape, increase picture training set scale, reach the effect for strengthening training;
3rd, sex, function and the style of input picture can be obtained according to the model of improved depth convolutional neural networks,
Foundation is provided for selecting collocation, high have the advantages that classification accuracy
4th, perception hash algorithm can be utilized to find similar pictures in collocation storehouse according to known picture, and according to collocation storehouse
In collocation be given collocation suggestion;
5th, the picture being input into can be classified, so as to obtain sex, function and the style information of clothes, also obtains which
To in requisition for collocation storehouse;On the other hand the picture to being input into finds similar pictures using hash algorithm is perceived in suit storehouse,
The collocation of similar pictures is obtained, and then is believed that the collocation is applied to the picture of input, on the other hand, more optional in order to obtain
The collocation list product that selects, reuse the similar pictures for perceiving the single product of hash algorithm search collocation, in collocation storehouse so as to obtain more
Plus abundant collocation is selected.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or the spy described with reference to the embodiment or example
Point is contained at least one embodiment or example of the present invention.In this manual, to the schematic representation of above-mentioned term not
Identical embodiment or example are necessarily referred to.And, the specific features of description, structure, material or feature can be any
One or more embodiments or example in combine in an appropriate manner.
Although an embodiment of the present invention has been shown and described, it will be understood by those skilled in the art that:Not
These embodiments can be carried out with multiple changes, modification, replacement and modification in the case of the principle and objective that depart from the present invention, this
The scope of invention is limited by claim and its equivalent.
Claims (10)
1. a kind of clothes classification and recommendation method of arranging in pairs or groups based on depth convolutional neural networks, it is characterised in that including following step
Suddenly:
S1:By adding batch normalization, improving inception structures, interpolation redundancy grader improves original GoogleNet volume
Product neutral net, to carry out feature extraction to garment image, obtains the classification results of the garment image;
S2:Carry out the picture augmentation of multi-method combination to storehouse training set of arranging in pairs or groups, and the garment image is carried out distorting, is overturn,
Color notation conversion space process, and train improved GoogleNet convolutional neural networks disaggregated model;
S3:Similar single product and its collocation in suit storehouse is found according to similar pictures lookup algorithm, and every garment image is generated
Identification information, obtains similar pictures by contrasting the identification information, and is taken according to the classification results of the garment image
The corresponding sex of dress picture, style, function information, and according to the corresponding sex of the garment image, style, function information synthesis
Provide garment coordination recommendation.
2. the clothes classification and recommendation method of arranging in pairs or groups, its feature based on depth convolutional neural networks according to claim 1
It is that the S1 is further included:
Improved inception construction design methods are based on, and original full connection side are replaced using the sparsity structure of local optimum
Formula, is substituted traditional convolutional layer increasing amount wave filter using multilayer perceptron, and training process is accelerated using dense calculating;
Inception construction design methods are based on, and original convolution pondization pressure are replaced using the compression unit of inception structures
Shrinking structure, to carry out parallel picture compression;
Based on the normalized data preprocessing method of batch, specified dimension data are all deducted average, becomes the number that average is 0
According to collection.
3. the clothes classification and recommendation method of arranging in pairs or groups, its feature based on depth convolutional neural networks according to claim 1
It is that the S2 is further included:
Based on the training set processing method of picture augmentation, using distortion, upset, the color notation conversion space process garment image,
And use different Enhancement Methods in color space, and using two angle points and the local segment of central point and cutting for horizontal reflection
Sanction mode, is adjusted to image of clothing.
4. the clothes classification and recommendation method of arranging in pairs or groups, its feature based on depth convolutional neural networks according to claim 3
It is, the utilization distortion, upset, the color notation conversion space process garment image, including:R passages are full, discount, G passages
Full, discount, channel B is full, discount, and adaptive histogram equalization, Retinex contrasts strengthen, and contrast linearly strengthens,
Contrast linear reduction.
5. the clothes classification and recommendation method of arranging in pairs or groups, its feature based on depth convolutional neural networks according to claim 1
It is that the S3 is further included:
Identification information is generated to every garment image by perceiving hash algorithm, and information contrast is identified by Hamming distance
To obtain similar pictures, wherein, simplify via minification, color, DCT is calculated, diminution DCT, calculating mean value, calculating hash
It is worth to the identification information;
According to the corresponding sex of the garment image, function and style, in sex corresponding with the garment image, function and wind
The clothes that arranges in pairs or groups is selected in the storehouse of lattice coupling, and phase is found using hash algorithm is perceived in suit storehouse to the garment image being input into
Like picture, the collocation of similar pictures is obtained, and the similar of the single product of perception hash algorithm search collocation is reused in collocation storehouse
Picture, obtains final collocation and selects.
6. a kind of clothes classification and commending system of arranging in pairs or groups based on depth convolutional neural networks, it is characterised in that include:
Sort module, the sort module are used for, by adding batch normalization, improving inception structures, adding redundancy point
Class device improves original GoogleNet convolutional neural networks, to carry out feature extraction to garment image, obtains the garment image
Classification results;
Processing module, the processing module are used for the picture augmentation for carrying out multi-method combination to storehouse training set of arranging in pairs or groups, and to described
Garment image is carried out distorting, is overturn, color notation conversion space process, and trains improved GoogleNet convolutional neural networks classification
Model;
Collocation recommending module, the collocation recommending module are used for finding the similar list in suit storehouse according to similar pictures lookup algorithm
Product and its collocation, generate identification information to every garment image, obtain similar pictures by contrasting the identification information, and according to
The classification results of the garment image obtain the corresponding sex of garment image, style, function information, and according to the garment image
Corresponding sex, style, function information comprehensively provide garment coordination recommendation.
7. the clothes classification and commending system of arranging in pairs or groups, its feature based on depth convolutional neural networks according to claim 6
It is, the sort module is used for:
Improved inception construction design methods are based on, and original full connection side are replaced using the sparsity structure of local optimum
Formula, is substituted traditional convolutional layer increasing amount wave filter using multilayer perceptron, and training process is accelerated using dense calculating;
Inception construction design methods are based on, and original convolution pondization pressure are replaced using the compression unit of inception structures
Shrinking structure, to carry out parallel picture compression;
Based on the normalized data preprocessing method of batch, specified dimension data are all deducted average, becomes the number that average is 0
According to collection.
8. the clothes classification and commending system of arranging in pairs or groups, its feature based on depth convolutional neural networks according to claim 6
It is, the processing module is used for the training set processing method based on picture augmentation, using distortion, upset, color notation conversion space
The garment image is processed, and different Enhancement Methods is used in color space, and using two angle points and the local of central point
Mode is cut out in fragment and horizontal reflection, and image of clothing is adjusted.
9. the clothes classification and commending system of arranging in pairs or groups, its feature based on depth convolutional neural networks according to claim 8
It is, the processing module processes garment image using distortion, upset, color notation conversion space, including:R passages are full, discount, G
Passage is full, discount, and channel B is full, discount, and adaptive histogram equalization, Retinex contrasts strengthen, and contrast is linear
Strengthen, contrast linear reduction.
10. the clothes classification and commending system of arranging in pairs or groups, its feature based on depth convolutional neural networks according to claim 6
It is, the collocation recommending module is used for:
Identification information is generated to every garment image by perceiving hash algorithm, and information contrast is identified by Hamming distance
To obtain similar pictures, wherein, simplify via minification, color, DCT is calculated, diminution DCT, calculating mean value, calculating hash
It is worth to the identification information;
According to the corresponding sex of the garment image, function and style, in sex corresponding with the garment image, function and wind
The clothes that arranges in pairs or groups is selected in the storehouse of lattice coupling, and phase is found using hash algorithm is perceived in suit storehouse to the garment image being input into
Like picture, the collocation of similar pictures is obtained, and the similar of the single product of perception hash algorithm search collocation is reused in collocation storehouse
Picture, obtains final collocation and selects.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610941721.1A CN106504064A (en) | 2016-10-25 | 2016-10-25 | Clothes classification based on depth convolutional neural networks recommends method and system with collocation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610941721.1A CN106504064A (en) | 2016-10-25 | 2016-10-25 | Clothes classification based on depth convolutional neural networks recommends method and system with collocation |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106504064A true CN106504064A (en) | 2017-03-15 |
Family
ID=58321792
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610941721.1A Pending CN106504064A (en) | 2016-10-25 | 2016-10-25 | Clothes classification based on depth convolutional neural networks recommends method and system with collocation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106504064A (en) |
Cited By (53)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106991439A (en) * | 2017-03-28 | 2017-07-28 | 南京天数信息科技有限公司 | Image-recognizing method based on deep learning and transfer learning |
CN107016415A (en) * | 2017-04-12 | 2017-08-04 | 合肥工业大学 | A kind of coloured image Color Semantic sorting technique based on full convolutional network |
CN107123033A (en) * | 2017-05-04 | 2017-09-01 | 北京科技大学 | A kind of garment coordination method based on depth convolutional neural networks |
CN107145518A (en) * | 2017-04-10 | 2017-09-08 | 同济大学 | Personalized recommendation system based on deep learning under a kind of social networks |
CN107292352A (en) * | 2017-08-07 | 2017-10-24 | 北京中星微电子有限公司 | Image classification method and device based on convolutional neural networks |
CN107688985A (en) * | 2017-08-23 | 2018-02-13 | 武汉恩祁瑞网络有限公司 | A kind of garment coordination commending system based on style element |
CN107730343A (en) * | 2017-09-15 | 2018-02-23 | 广州唯品会研究院有限公司 | A kind of user's merchandise news method for pushing and equipment based on picture attribute extraction |
CN107909076A (en) * | 2017-12-11 | 2018-04-13 | 奕响(大连)科技有限公司 | A kind of similar decision method of picture based on LBP |
CN107909580A (en) * | 2017-11-01 | 2018-04-13 | 深圳市深网视界科技有限公司 | A kind of pedestrian wears color identification method, electronic equipment and storage medium clothes |
CN108052969A (en) * | 2017-12-08 | 2018-05-18 | 奕响(大连)科技有限公司 | A kind of similar determination method of DCT pixel grey scales picture |
CN108171568A (en) * | 2017-12-11 | 2018-06-15 | 武汉纺织大学 | A kind of clothes in knowledge based library recommend method and system |
CN108197898A (en) * | 2018-01-26 | 2018-06-22 | 维沃移动通信有限公司 | Method and mobile terminal are recommended in a kind of dressing |
CN108230082A (en) * | 2017-06-16 | 2018-06-29 | 深圳市商汤科技有限公司 | The recommendation method and apparatus of collocation dress ornament, electronic equipment, storage medium |
CN108230297A (en) * | 2017-11-30 | 2018-06-29 | 复旦大学 | A kind of collocation of colour appraisal procedure replaced based on clothes |
CN108428166A (en) * | 2018-02-13 | 2018-08-21 | 东华大学 | The clothes commending system of figure and features feature recognition classification based on convolutional neural networks |
CN108494968A (en) * | 2018-04-04 | 2018-09-04 | Oppo广东移动通信有限公司 | Intelligent prompt method, device, storage medium and intelligent terminal |
CN108596094A (en) * | 2018-04-24 | 2018-09-28 | 杭州数为科技有限公司 | Personage's style detecting system, method, terminal and medium |
CN108615010A (en) * | 2018-04-24 | 2018-10-02 | 重庆邮电大学 | Facial expression recognizing method based on the fusion of parallel convolutional neural networks characteristic pattern |
CN108664922A (en) * | 2018-05-10 | 2018-10-16 | 东华大学 | A kind of infrared video Human bodys' response method based on personal safety |
CN108921030A (en) * | 2018-06-04 | 2018-11-30 | 浙江大学 | A kind of SAR automatic target recognition method of Fast Learning |
CN108986159A (en) * | 2018-04-25 | 2018-12-11 | 浙江森马服饰股份有限公司 | A kind of method and apparatus that three-dimensional (3 D) manikin is rebuild and measured |
CN109002863A (en) * | 2018-06-27 | 2018-12-14 | 佛山市顺德区中山大学研究院 | A kind of compact convolutional neural networks of resourceoriented constrained system |
CN109117779A (en) * | 2018-08-06 | 2019-01-01 | 百度在线网络技术(北京)有限公司 | One kind, which is worn, takes recommended method, device and electronic equipment |
CN109559761A (en) * | 2018-12-21 | 2019-04-02 | 广东工业大学 | A kind of risk of stroke prediction technique based on depth phonetic feature |
CN109597907A (en) * | 2017-12-07 | 2019-04-09 | 深圳市商汤科技有限公司 | Dress ornament management method and device, electronic equipment, storage medium |
CN109670591A (en) * | 2018-12-14 | 2019-04-23 | 深圳市商汤科技有限公司 | A kind of training method and image matching method, device of neural network |
CN109685756A (en) * | 2017-10-16 | 2019-04-26 | 乐达创意科技有限公司 | Image feature automatic identifier, system and method |
CN109784181A (en) * | 2018-12-14 | 2019-05-21 | 平安科技(深圳)有限公司 | Picture watermark recognition methods, device, equipment and computer readable storage medium |
WO2019120031A1 (en) * | 2017-12-20 | 2019-06-27 | Oppo广东移动通信有限公司 | Method, device, storage medium, and mobile terminal for making recommendation about clothing matching |
CN110020113A (en) * | 2017-09-28 | 2019-07-16 | 南京无界家居科技有限公司 | A kind of family product prediction technique and device based on characteristic matching |
CN110021061A (en) * | 2018-01-08 | 2019-07-16 | 广东欧珀移动通信有限公司 | Collocation model building method, dress ornament recommended method, device, medium and terminal |
CN110110209A (en) * | 2018-01-22 | 2019-08-09 | 青岛科技大学 | A kind of intersection recommended method and system based on local weighted linear regression model (LRM) |
CN110135032A (en) * | 2019-04-30 | 2019-08-16 | 厦门大学 | A kind of auxiliary clothes generation method and device generating network based on confrontation |
CN110232253A (en) * | 2019-06-20 | 2019-09-13 | 杭州时趣信息技术有限公司 | Generate computer installation, equipment, storage medium and the method for clothing matching scheme |
CN110232133A (en) * | 2019-05-16 | 2019-09-13 | 华中科技大学 | A kind of image of clothing search method and system classified based on Fusion Features and style |
CN110363213A (en) * | 2018-04-11 | 2019-10-22 | 国际商业机器公司 | The cognitive analysis of image of clothing and classification |
WO2019223302A1 (en) * | 2018-05-23 | 2019-11-28 | 山东大学 | Dress collocation method and system based on attention knowledge extraction, and storage medium |
WO2020037940A1 (en) * | 2018-08-23 | 2020-02-27 | 深圳码隆科技有限公司 | Method and device for screening image data of character clothing |
CN110874771A (en) * | 2018-08-29 | 2020-03-10 | 北京京东尚科信息技术有限公司 | Method and device for matching commodities |
CN110909746A (en) * | 2018-09-18 | 2020-03-24 | 深圳云天励飞技术有限公司 | Clothing recommendation method, related device and equipment |
CN111209475A (en) * | 2019-12-27 | 2020-05-29 | 武汉大学 | Interest point recommendation method and device based on space-time sequence and social embedded ranking |
CN111242154A (en) * | 2019-04-01 | 2020-06-05 | 泰州市康平医疗科技有限公司 | Storage management method for clothes cabinet |
CN111383081A (en) * | 2020-03-24 | 2020-07-07 | 东华大学 | Intelligent recommendation method for clothing matching |
CN111506761A (en) * | 2020-04-22 | 2020-08-07 | 上海极链网络科技有限公司 | Similar picture query method, device, system and storage medium |
CN111582979A (en) * | 2020-04-29 | 2020-08-25 | 上海风秩科技有限公司 | Clothing matching recommendation method and device and electronic equipment |
US10755228B1 (en) * | 2017-03-29 | 2020-08-25 | Blue Yonder Group, Inc. | Image processing system for deep fashion color recognition |
CN111680760A (en) * | 2020-06-16 | 2020-09-18 | 北京联合大学 | Clothing style identification method and device, electronic equipment and storage medium |
CN113330455A (en) * | 2019-01-16 | 2021-08-31 | 微软技术许可有限责任公司 | Finding complementary digital images using conditional generative countermeasure networks |
CN113869420A (en) * | 2021-09-28 | 2021-12-31 | 平安科技(深圳)有限公司 | Text recommendation method based on comparative learning and related equipment |
EP3905181A4 (en) * | 2019-01-24 | 2022-02-23 | Samsung Electronics Co., Ltd. | Electronic device and method for operating same |
US11538083B2 (en) | 2018-05-17 | 2022-12-27 | International Business Machines Corporation | Cognitive fashion product recommendation system, computer program product, and method |
CN111401456B (en) * | 2020-03-20 | 2023-08-22 | 杭州涂鸦信息技术有限公司 | Training method, system and device for face gesture recognition model |
US11790430B2 (en) | 2021-03-15 | 2023-10-17 | Tata Consultancy Services Limited | Method and system for determining unified user intention from digital environment for plurality of strategies |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103324650A (en) * | 2012-10-23 | 2013-09-25 | 深圳市宜搜科技发展有限公司 | Image retrieval method and system |
CN104391971A (en) * | 2014-12-05 | 2015-03-04 | 常州飞寻视讯信息科技有限公司 | Intelligent automatic dress collocation recommending method |
CN105761127A (en) * | 2016-02-05 | 2016-07-13 | 上海和鹰机电科技股份有限公司 | Clothing style intelligent matching method based on sign information and system thereof |
CN105844238A (en) * | 2016-03-23 | 2016-08-10 | 乐视云计算有限公司 | Method and system for discriminating videos |
-
2016
- 2016-10-25 CN CN201610941721.1A patent/CN106504064A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103324650A (en) * | 2012-10-23 | 2013-09-25 | 深圳市宜搜科技发展有限公司 | Image retrieval method and system |
CN104391971A (en) * | 2014-12-05 | 2015-03-04 | 常州飞寻视讯信息科技有限公司 | Intelligent automatic dress collocation recommending method |
CN105761127A (en) * | 2016-02-05 | 2016-07-13 | 上海和鹰机电科技股份有限公司 | Clothing style intelligent matching method based on sign information and system thereof |
CN105844238A (en) * | 2016-03-23 | 2016-08-10 | 乐视云计算有限公司 | Method and system for discriminating videos |
Cited By (77)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106991439A (en) * | 2017-03-28 | 2017-07-28 | 南京天数信息科技有限公司 | Image-recognizing method based on deep learning and transfer learning |
US11625677B2 (en) | 2017-03-29 | 2023-04-11 | Blue Yonder Group, Inc. | Image processing system for deep fashion color recognition |
US11995608B2 (en) | 2017-03-29 | 2024-05-28 | Blue Yonder Group, Inc. | Image processing system for deep fashion color recognition |
US10755228B1 (en) * | 2017-03-29 | 2020-08-25 | Blue Yonder Group, Inc. | Image processing system for deep fashion color recognition |
CN107145518B (en) * | 2017-04-10 | 2020-04-14 | 同济大学 | Personalized recommendation system based on deep learning in social network |
CN107145518A (en) * | 2017-04-10 | 2017-09-08 | 同济大学 | Personalized recommendation system based on deep learning under a kind of social networks |
CN107016415B (en) * | 2017-04-12 | 2019-07-19 | 合肥工业大学 | A kind of color image Color Semantic classification method based on full convolutional network |
CN107016415A (en) * | 2017-04-12 | 2017-08-04 | 合肥工业大学 | A kind of coloured image Color Semantic sorting technique based on full convolutional network |
CN107123033A (en) * | 2017-05-04 | 2017-09-01 | 北京科技大学 | A kind of garment coordination method based on depth convolutional neural networks |
US11232324B2 (en) | 2017-06-16 | 2022-01-25 | Shenzhen Sensetime Technology Co., Ltd. | Methods and apparatus for recommending collocating dress, electronic devices, and storage media |
WO2018228448A1 (en) * | 2017-06-16 | 2018-12-20 | 深圳市商汤科技有限公司 | Method and apparatus for recommending matching clothing, electronic device and storage medium |
CN108230082A (en) * | 2017-06-16 | 2018-06-29 | 深圳市商汤科技有限公司 | The recommendation method and apparatus of collocation dress ornament, electronic equipment, storage medium |
CN107292352A (en) * | 2017-08-07 | 2017-10-24 | 北京中星微电子有限公司 | Image classification method and device based on convolutional neural networks |
CN107292352B (en) * | 2017-08-07 | 2020-06-02 | 北京中星微人工智能芯片技术有限公司 | Image classification method and device based on convolutional neural network |
CN107688985A (en) * | 2017-08-23 | 2018-02-13 | 武汉恩祁瑞网络有限公司 | A kind of garment coordination commending system based on style element |
CN107730343A (en) * | 2017-09-15 | 2018-02-23 | 广州唯品会研究院有限公司 | A kind of user's merchandise news method for pushing and equipment based on picture attribute extraction |
CN110020113A (en) * | 2017-09-28 | 2019-07-16 | 南京无界家居科技有限公司 | A kind of family product prediction technique and device based on characteristic matching |
CN109685756A (en) * | 2017-10-16 | 2019-04-26 | 乐达创意科技有限公司 | Image feature automatic identifier, system and method |
CN107909580A (en) * | 2017-11-01 | 2018-04-13 | 深圳市深网视界科技有限公司 | A kind of pedestrian wears color identification method, electronic equipment and storage medium clothes |
CN108230297A (en) * | 2017-11-30 | 2018-06-29 | 复旦大学 | A kind of collocation of colour appraisal procedure replaced based on clothes |
CN108230297B (en) * | 2017-11-30 | 2020-05-12 | 复旦大学 | Color collocation assessment method based on garment replacement |
CN109597907A (en) * | 2017-12-07 | 2019-04-09 | 深圳市商汤科技有限公司 | Dress ornament management method and device, electronic equipment, storage medium |
CN108052969A (en) * | 2017-12-08 | 2018-05-18 | 奕响(大连)科技有限公司 | A kind of similar determination method of DCT pixel grey scales picture |
CN108171568A (en) * | 2017-12-11 | 2018-06-15 | 武汉纺织大学 | A kind of clothes in knowledge based library recommend method and system |
CN107909076A (en) * | 2017-12-11 | 2018-04-13 | 奕响(大连)科技有限公司 | A kind of similar decision method of picture based on LBP |
CN109949116A (en) * | 2017-12-20 | 2019-06-28 | 广东欧珀移动通信有限公司 | Clothing matching recommended method, device, storage medium and mobile terminal |
WO2019120031A1 (en) * | 2017-12-20 | 2019-06-27 | Oppo广东移动通信有限公司 | Method, device, storage medium, and mobile terminal for making recommendation about clothing matching |
CN110021061A (en) * | 2018-01-08 | 2019-07-16 | 广东欧珀移动通信有限公司 | Collocation model building method, dress ornament recommended method, device, medium and terminal |
CN110021061B (en) * | 2018-01-08 | 2021-10-29 | Oppo广东移动通信有限公司 | Collocation model construction method, clothing recommendation method, device, medium and terminal |
CN110110209A (en) * | 2018-01-22 | 2019-08-09 | 青岛科技大学 | A kind of intersection recommended method and system based on local weighted linear regression model (LRM) |
CN108197898A (en) * | 2018-01-26 | 2018-06-22 | 维沃移动通信有限公司 | Method and mobile terminal are recommended in a kind of dressing |
CN108197898B (en) * | 2018-01-26 | 2021-06-25 | 维沃移动通信有限公司 | Dressing recommendation method and mobile terminal |
CN108428166A (en) * | 2018-02-13 | 2018-08-21 | 东华大学 | The clothes commending system of figure and features feature recognition classification based on convolutional neural networks |
CN108494968A (en) * | 2018-04-04 | 2018-09-04 | Oppo广东移动通信有限公司 | Intelligent prompt method, device, storage medium and intelligent terminal |
CN110363213A (en) * | 2018-04-11 | 2019-10-22 | 国际商业机器公司 | The cognitive analysis of image of clothing and classification |
CN110363213B (en) * | 2018-04-11 | 2022-11-01 | 国际商业机器公司 | Method and system for cognitive analysis and classification of garment images |
CN108615010A (en) * | 2018-04-24 | 2018-10-02 | 重庆邮电大学 | Facial expression recognizing method based on the fusion of parallel convolutional neural networks characteristic pattern |
CN108596094B (en) * | 2018-04-24 | 2021-02-05 | 杭州数为科技有限公司 | Character style detection system, method, terminal and medium |
CN108615010B (en) * | 2018-04-24 | 2022-02-11 | 重庆邮电大学 | Facial expression recognition method based on parallel convolution neural network feature map fusion |
CN108596094A (en) * | 2018-04-24 | 2018-09-28 | 杭州数为科技有限公司 | Personage's style detecting system, method, terminal and medium |
CN108986159A (en) * | 2018-04-25 | 2018-12-11 | 浙江森马服饰股份有限公司 | A kind of method and apparatus that three-dimensional (3 D) manikin is rebuild and measured |
CN108664922A (en) * | 2018-05-10 | 2018-10-16 | 东华大学 | A kind of infrared video Human bodys' response method based on personal safety |
US11538083B2 (en) | 2018-05-17 | 2022-12-27 | International Business Machines Corporation | Cognitive fashion product recommendation system, computer program product, and method |
WO2019223302A1 (en) * | 2018-05-23 | 2019-11-28 | 山东大学 | Dress collocation method and system based on attention knowledge extraction, and storage medium |
CN108921030A (en) * | 2018-06-04 | 2018-11-30 | 浙江大学 | A kind of SAR automatic target recognition method of Fast Learning |
CN108921030B (en) * | 2018-06-04 | 2022-02-01 | 浙江大学 | SAR automatic target recognition method |
CN109002863A (en) * | 2018-06-27 | 2018-12-14 | 佛山市顺德区中山大学研究院 | A kind of compact convolutional neural networks of resourceoriented constrained system |
CN109002863B (en) * | 2018-06-27 | 2022-04-15 | 佛山市顺德区中山大学研究院 | Image processing method based on compact convolutional neural network |
CN109117779A (en) * | 2018-08-06 | 2019-01-01 | 百度在线网络技术(北京)有限公司 | One kind, which is worn, takes recommended method, device and electronic equipment |
WO2020037940A1 (en) * | 2018-08-23 | 2020-02-27 | 深圳码隆科技有限公司 | Method and device for screening image data of character clothing |
CN110874771A (en) * | 2018-08-29 | 2020-03-10 | 北京京东尚科信息技术有限公司 | Method and device for matching commodities |
CN110909746A (en) * | 2018-09-18 | 2020-03-24 | 深圳云天励飞技术有限公司 | Clothing recommendation method, related device and equipment |
WO2020119311A1 (en) * | 2018-12-14 | 2020-06-18 | 深圳市商汤科技有限公司 | Neural network training method and image matching method and device |
CN109784181A (en) * | 2018-12-14 | 2019-05-21 | 平安科技(深圳)有限公司 | Picture watermark recognition methods, device, equipment and computer readable storage medium |
CN109784181B (en) * | 2018-12-14 | 2024-03-22 | 平安科技(深圳)有限公司 | Picture watermark identification method, device, equipment and computer readable storage medium |
TWI760650B (en) * | 2018-12-14 | 2022-04-11 | 大陸商深圳市商湯科技有限公司 | Method and image matching method for neural network training and device thereof |
CN109670591A (en) * | 2018-12-14 | 2019-04-23 | 深圳市商汤科技有限公司 | A kind of training method and image matching method, device of neural network |
CN109559761A (en) * | 2018-12-21 | 2019-04-02 | 广东工业大学 | A kind of risk of stroke prediction technique based on depth phonetic feature |
CN113330455A (en) * | 2019-01-16 | 2021-08-31 | 微软技术许可有限责任公司 | Finding complementary digital images using conditional generative countermeasure networks |
EP3905181A4 (en) * | 2019-01-24 | 2022-02-23 | Samsung Electronics Co., Ltd. | Electronic device and method for operating same |
CN111242154A (en) * | 2019-04-01 | 2020-06-05 | 泰州市康平医疗科技有限公司 | Storage management method for clothes cabinet |
CN110135032A (en) * | 2019-04-30 | 2019-08-16 | 厦门大学 | A kind of auxiliary clothes generation method and device generating network based on confrontation |
CN110135032B (en) * | 2019-04-30 | 2020-12-18 | 厦门大学 | Auxiliary clothing generation method and device based on confrontation generation network |
CN110232133A (en) * | 2019-05-16 | 2019-09-13 | 华中科技大学 | A kind of image of clothing search method and system classified based on Fusion Features and style |
CN110232253B (en) * | 2019-06-20 | 2023-01-24 | 杭州卷瓜网络有限公司 | Computer device, equipment, storage medium and method for generating clothing matching scheme |
CN110232253A (en) * | 2019-06-20 | 2019-09-13 | 杭州时趣信息技术有限公司 | Generate computer installation, equipment, storage medium and the method for clothing matching scheme |
CN111209475B (en) * | 2019-12-27 | 2022-03-15 | 武汉大学 | Interest point recommendation method and device based on space-time sequence and social embedded ranking |
CN111209475A (en) * | 2019-12-27 | 2020-05-29 | 武汉大学 | Interest point recommendation method and device based on space-time sequence and social embedded ranking |
CN111401456B (en) * | 2020-03-20 | 2023-08-22 | 杭州涂鸦信息技术有限公司 | Training method, system and device for face gesture recognition model |
CN111383081A (en) * | 2020-03-24 | 2020-07-07 | 东华大学 | Intelligent recommendation method for clothing matching |
CN111506761B (en) * | 2020-04-22 | 2021-05-14 | 上海极链网络科技有限公司 | Similar picture query method, device, system and storage medium |
CN111506761A (en) * | 2020-04-22 | 2020-08-07 | 上海极链网络科技有限公司 | Similar picture query method, device, system and storage medium |
CN111582979A (en) * | 2020-04-29 | 2020-08-25 | 上海风秩科技有限公司 | Clothing matching recommendation method and device and electronic equipment |
CN111680760A (en) * | 2020-06-16 | 2020-09-18 | 北京联合大学 | Clothing style identification method and device, electronic equipment and storage medium |
US11790430B2 (en) | 2021-03-15 | 2023-10-17 | Tata Consultancy Services Limited | Method and system for determining unified user intention from digital environment for plurality of strategies |
CN113869420A (en) * | 2021-09-28 | 2021-12-31 | 平安科技(深圳)有限公司 | Text recommendation method based on comparative learning and related equipment |
CN113869420B (en) * | 2021-09-28 | 2024-06-21 | 平安科技(深圳)有限公司 | Text recommendation method and related equipment based on contrast learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106504064A (en) | Clothes classification based on depth convolutional neural networks recommends method and system with collocation | |
CN110378985B (en) | Animation drawing auxiliary creation method based on GAN | |
CN108717568B (en) | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network | |
CN110399821B (en) | Customer satisfaction acquisition method based on facial expression recognition | |
CN104809187B (en) | A kind of indoor scene semanteme marking method based on RGB D data | |
CN112950661B (en) | Attention-based generation method for generating network face cartoon | |
CN108961675A (en) | Fall detection method based on convolutional neural networks | |
Gando et al. | Fine-tuning deep convolutional neural networks for distinguishing illustrations from photographs | |
CN107748858A (en) | A kind of multi-pose eye locating method based on concatenated convolutional neutral net | |
CN109409435A (en) | A kind of depth perception conspicuousness detection method based on convolutional neural networks | |
CN108898145A (en) | A kind of image well-marked target detection method of combination deep learning | |
CN109325443A (en) | A kind of face character recognition methods based on the study of more example multi-tag depth migrations | |
CN107808132A (en) | A kind of scene image classification method for merging topic model | |
CN107220657A (en) | A kind of method of high-resolution remote sensing image scene classification towards small data set | |
CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
CN106326899A (en) | Tobacco leaf grading method based on hyperspectral image and deep learning algorithm | |
CN108345850A (en) | The scene text detection method of the territorial classification of stroke feature transformation and deep learning based on super-pixel | |
CN107292250A (en) | A kind of gait recognition method based on deep neural network | |
CN105844295A (en) | Video smog fine classification method based on color model and motion characteristics | |
CN107844795A (en) | Convolutional neural network feature extraction method based on principal component analysis | |
CN108009637B (en) | Station caption segmentation method of pixel-level station caption identification network based on cross-layer feature extraction | |
CN111242841A (en) | Image background style migration method based on semantic segmentation and deep learning | |
CN111553837A (en) | Artistic text image generation method based on neural style migration | |
CN102750385A (en) | Correlation-quality sequencing image retrieval method based on tag retrieval | |
CN104298974A (en) | Human body behavior recognition method based on depth video sequence |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170315 |