CN109145947A - A kind of Fashionable women dress image fine grit classification method based on component detection and visual signature - Google Patents
A kind of Fashionable women dress image fine grit classification method based on component detection and visual signature Download PDFInfo
- Publication number
- CN109145947A CN109145947A CN201810784023.4A CN201810784023A CN109145947A CN 109145947 A CN109145947 A CN 109145947A CN 201810784023 A CN201810784023 A CN 201810784023A CN 109145947 A CN109145947 A CN 109145947A
- Authority
- CN
- China
- Prior art keywords
- image
- feature
- model
- fashionable women
- dress
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000000007 visual effect Effects 0.000 title claims abstract description 44
- 238000000034 method Methods 0.000 title claims abstract description 32
- 238000001514 detection method Methods 0.000 title claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 39
- 238000000605 extraction Methods 0.000 claims abstract description 23
- 230000004044 response Effects 0.000 claims description 33
- 239000013598 vector Substances 0.000 claims description 16
- 238000012360 testing method Methods 0.000 claims description 12
- 238000007637 random forest analysis Methods 0.000 claims description 11
- 239000000284 extract Substances 0.000 claims description 8
- 230000036544 posture Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000006073 displacement reaction Methods 0.000 claims description 3
- 238000011049 filling Methods 0.000 claims description 3
- 210000004705 lumbosacral region Anatomy 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000005259 measurement Methods 0.000 claims description 3
- 239000003086 colorant Substances 0.000 claims description 2
- 238000007689 inspection Methods 0.000 claims 1
- 230000009452 underexpressoin Effects 0.000 claims 1
- 238000013527 convolutional neural network Methods 0.000 description 4
- 229910002056 binary alloy Inorganic materials 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 241000282412 Homo Species 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 239000011248 coating agent Substances 0.000 description 1
- 238000000576 coating method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 230000037303 wrinkles Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of Fashionable women dress image fine grit classification methods based on component detection and visual signature, belong to computer vision and image application field.The component that the present invention carries out physical feeling to image in the Fashionable women dress image to be classified and training set of input first detects；Secondly, the Fashionable women dress image after detection is extracted respectively, and 4 kinds of HOG, LBP, color histogram and boundary operator low-level image features of training Fashionable women dress image, the image after obtaining feature extraction；Then, the visual signature descriptor of definition is matched with the 4 kinds of low-level image features extracted, using multiclass SVM supervised learning training fine grit classification device model；Finally, realizing fine grit classification by the fine grit classification device after training to the Fashionable women dress image of feature extraction, exporting the classification results of Fashionable women dress image.Detection and the classification method accuracy rate with higher that the present invention uses.
Description
Technical field
The present invention relates to a kind of Fashionable women dress image fine grit classification methods based on component detection and visual signature, belong to
Computer vision and image application field.
Background technique
Shopping at network receives the very big welcome of people, shows the development trend of universalness, globalization, mobile, makes
The topic that fashion clothing classification becomes more and more popular is obtained, fashion clothing is sorted in the fields such as e-commerce and is used widely.
Therefore, also there are many improved methods in fashion clothing classification, including word packet model the most classical, is based on depth
The fashion clothing classification method of habit and based on random forest, SVM (Support Vector Machine, support vector machines, letter
Claim SVM), the methods of CNN (Convolutional Neural Network, convolutional neural networks, abbreviation CNN).Known method
Classify mostly both for the coarseness of fashion clothing image, lacks the analysis between similar style classification, cannot achieve more smart
Thin division and multilayer subseries.Due to a Fashionable women dress great variety of models, different from the classification task of coarseness, Fashionable women dress particulate
The classification precision for spending image is more careful, and difference is subtleer between style classification, often can only be by means of small local difference ability
Distinguish different styles.In addition, the signal-to-noise ratio very little of fine granularity image, the information containing enough discriminations are present in very tiny
Regional area in.Therefore, how to find and efficiently use useful local region information, more finely, accurately and efficiently realize
Fashionable women dress image fine grit classification has important theory significance and practical value.In existing known method, as Berg (<
POOF:Part-Based One-vs.-One Features for Fine-Grained Categorization,Face
Verification, and Attribute Estimation >, 2013:955-962.) propose the one-to-one spy based on position
The image focusing study that the POOFs method of sign can be marked automatically from one group of specific area and with specific position and classification is big
Measure the different mid-level features with high distinction.Each feature can be according to the appearance features of object specific position come area
It is divided to two different classes.Bossard (<Apparel classification with style>, 2012,321-335.) is directed to
How to be identified in natural scene and fashion clothing of classifying propose a complete method, key be using it is multiple based on
The study of machine forest, and use the learner of powerful recognition capability as decision node, while being also extended to random forest and possessing
The migration forest that different field can be converted.Cui(<Fine-Grained Categorization and Dataset
Bootstrapping Using Deep Metric Learning with Humans in the Loop>,2015:1153-
1162) propose that the general iterative frame learnt based on depth measure for fine grit classification, is embedded into each classification with study
The low-dimensional feature of upper anchor point.Zhang(<Weakly Supervised Fine-Grained Categorization With
Part-Based Image Representation >, 2016,25 (4): 1713-1725.) propose one be easy to dispose it is thin
Granularity image categorizing system is annotated in training or in test phase without using any object or part, and only uses trained figure
The class label of picture.
In conclusion although the realization means of the classification method of fashion clothing image have very much, since clothes fashion is more
Sample, texture and accessories are changeable and clothes flexibility is easily-deformable, so that the pattern of clothes itself changes, these factors are to classification
Identification brings very big difficulty.Certain defect and limitation are still had in known method, and due to photographed scene and people
Body posture is numerous, so that how to detect human body different zones is particularly important.In terms of feature extraction and classification, known method
The low-level image features such as color, texture are mostly based on to realize feature extraction, local message cannot be utilized very well, for fashion clothing it
Between between subtle style class and in class the feature extraction of difference have some limitations, can only realize the coarseness of fashion clothing
Classification.
Summary of the invention
The present invention relates to a kind of Fashionable women dress image fine grit classification methods based on component detection and visual signature, with suitable
It answers the physical feeling of different gestures and view transformation to detect, meets the fine granularity point of e-commerce Fashionable women dress image on the way
Class.
The technical scheme is that a kind of Fashionable women dress image fine grit classification based on component detection and visual signature
Method includes the following steps: that Step1 to the training Fashionable women dress image T and Fashionable women dress image I to be sorted of input, is used
Improved DPM model carries out component detection to the human body under different postures and visual angle；Firstly, to training Fashionable women dress image
T and Fashionable women dress image I to be sorted extracts HOG ((Histogram of Oriented Gradient, direction gradient histogram
Figure, abbreviation HOG) and obtain after being normalized DPM (Deformable Part Model, deformable component model, referred to as
DPM) feature；Secondly, according to human body attitude, visual angle adjust DPM human testing model, by human testing model be divided into root model and
Partial model；Then, the response score of root model and partial model is calculated separately according to DPM feature, carries out response transform calculating
Goal hypothesis score obtains optimal location to calculate the comprehensive response score of each position of target, finally obtains detection knot
Fruit.
Improved DPM model is made of a root model and several partial models, and the object model of n component is expressed as one
A (n+2) tuple (F0,P1,...Pi,...Pn, b), wherein F0It is root filter, PiIt is the model of i-th of component, b is one inclined
From loss coefficient, in l0Scale layer, with (x0,y0) be anchor point response score are as follows:
Wherein,For the response score of root model, viIt is a bivector, for specifying i-th of filter
Coordinate of the anchor point position (normal place when i.e. deformation occurs) relative to root position,For
The response score of n partial model, λ are the number of levels of the Feature Mapping calculated in feature pyramid with twice of resolution ratio；
After calculating response score, the response of converting member filter simultaneously considers spatial location laws, and response transform calculates public
Formula is as follows:
Wherein, (x, y) is i-th of partial model in the ideal position of scale layer, and l is the number of levels of feature pyramid H,
(dx, dy) is the offset of opposite (x, y), Ri,l(x+dx, y+dy) is matching score of the partial model at (x+dx, y+dy),
di.φd(dx, dy) is the score for deviating (dx, dy) and being lost, φd(dx, dy)=(dx, dy, dx2,dy2) it is DPM feature, di
To deviate loss coefficient, when the parameter model for needing to learn when being model training initializes, di=(0,0,1,1) deviates damage
Lose the Euclidean distance for offset relative ideal position；
Each goal hypothesis specifies position of each filter in feature pyramid H in model: z=(p0,...,
pn), pi=(xi,yi,li) it is layer and position coordinates where i-th of filter, the score of goal hypothesis calculates as follows:
Wherein Fi'.φ(H,pi) be i-th of filter score, φ (H, pi) be feature pyramid H feature vector,
Fi' it is to connect weight vectors in i-th of filter and the vector that obtains, (dxi,dyi)=(xi,yi)-(2(x0,y0)+vi) give
Displacement of i-th of filter position relative to its anchor point position is gone out, optimal location, root is obtained by goal hypothesis score
According to the comprehensive response score of each position of best position calculation:
Multiple examples that target is detected by the comprehensive response score of each position, obtain testing result.
Step2 extract respectively detection after training Fashionable women dress image T' and Fashionable women dress image I' to be sorted HOG,
4 kinds of LBP (Local Binary Pattern, local binary patterns), color histogram and boundary operator low-level image features, obtain spy
Training Fashionable women dress image T " and Fashionable women dress image I " to be sorted after sign extraction.
Step3 matches the visual signature descriptor of definition with the 4 kinds of low-level image features extracted, using multiclass SVM
Supervised learning trains fine grit classification device model；Fashionable women dress is divided into upper body women's dress and lower part of the body women's dress first, wherein upper body clothes
Dress is divided into 14 kinds of styles, and lower body garment is divided into 6 kinds of styles, and whole body clothes fashion is divided into 3 kinds of styles, according to different attribute (such as clothing
Neck, sleeve, sleeve type, color, type pattern etc.) carry out attribute labeling；Secondly, by defining visual signature descriptor to fashion female
The style and attribute for filling image are described, 4 kinds of low-level image features for then extracting visual signature descriptor and step2 into
Row characteristic matching, wherein visual signature descriptor is divided into upper body visual signature descriptor, lower part of the body visual signature descriptor and the overall situation
Feature descriptor；Finally by random forest and Multiclass SVM method supervised learning to the training Fashionable women dress figure after feature extraction
Picture T " is trained, and obtains the fine grit classification device of style and attribute.
Step4 realizes fine granularity by the fine grit classification device after training, to the Fashionable women dress image I " of feature extraction
Classification exports the classification results of Fashionable women dress image.
The beneficial effects of the present invention are:
1, the detection method of well known fashion clothing image is carried out primarily directed to the fashion clothing image under ideal scenario
Detection, but due to photographed scene and take pictures posture and illumination, the multifactor interference such as block, there is certain limitation.
The present invention can preferably adapt to different scenes, difference using the component detection based on human body position of improved DPM model
The detection of the human body of posture and view transformation.
2, well known feature extracting method is mostly based on color characteristic and global characteristics, and characteristic attribute is relatively simple, can not
Obtain fine granularity important local feature and attribute.The present invention passes through the perceptual property descriptor of definition, which is divided into
Body visual signature descriptor, lower part of the body visual signature descriptor and global characteristics descriptor.By visual signature descriptor with extract
Training Fashionable women dress image 4 kinds of low-level image features carry out characteristic matching, improve the accuracy rate of Visual Feature Retrieval Process and expression.
3, well known fashion clothing classification method is mostly based on SVM classifier, however traditional SVM classifier is in fine granularity
It has some limitations in image classification, exercises supervision respectively in the present invention to the different fashion clothing attributes of defined mistake
It practises, establishes the fine grit classification device model of Fashionable women dress image, and combine to feature extraction by using random forest and SVM
Fashionable women dress image realizes fine grit classification, exports the classification results of Fashionable women dress image, classification accuracy with higher.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is example flow diagram figure in the present invention；
Fig. 3 is that Fashionable women dress low-level image feature extracts exemplary diagram in the present invention；
Fig. 4 is Fashionable women dress attributed graph in the present invention；
Fig. 5 is Fashionable women dress classifying quality figure in the present invention；
Specific embodiment
Present invention will be further explained below with reference to the attached drawings and specific embodiments.
Embodiment 1: as shown in Figs. 1-2, a kind of Fashionable women dress image fine granularity point based on component detection and visual signature
Class method carries out body to the Fashionable women dress image in the Fashionable women dress image and Fashionable women dress training set to be sorted of input first
The component at position detects；Secondly, distinguishing the Fashionable women dress image after extracting parts detection and training Fashionable women dress image
4 kinds of HOG, LBP, color histogram and boundary operator low-level image features, the image after obtaining feature extraction；Then, by the view of definition
Feel that feature descriptor is matched with the 4 kinds of low-level image features extracted, it is thin using random forest and the training of multiclass SVM supervised learning
Grain-size classification device model；Finally, realizing particulate to the Fashionable women dress image of feature extraction by the fine grit classification device after training
Degree classification, exports the classification results of Fashionable women dress image.
Specific step is as follows:
Step1 is to the training Fashionable women dress image T and Fashionable women dress image I to be sorted of input, using improved DPM mould
Type carries out component detection to the human body under different postures and visual angle；Firstly, to training Fashionable women dress image T and to be sorted
Fashionable women dress image I extracts HOG and obtains DPM feature after being normalized；Secondly, adjusting DPM people according to human body attitude, visual angle
Human testing model is divided into root model and partial model by body detection model；Then, root model is calculated separately according to DPM feature
It with the response score of partial model, carries out response transform and calculates goal hypothesis score, it is every to calculate target to obtain optimal location
The comprehensive response score of a root position, finally obtains testing result；
Step2 extract respectively detection after training Fashionable women dress image T' and Fashionable women dress image I' to be sorted HOG,
4 kinds of LBP, color histogram and boundary operator low-level image features, training Fashionable women dress image T " after obtaining feature extraction and to point
The Fashionable women dress image I " of class；
Step3 matches the visual signature descriptor of definition with the 4 kinds of low-level image features extracted, using multiclass SVM
Supervised learning trains fine grit classification device model；Fashionable women dress is divided into upper body women's dress and lower part of the body women's dress first, wherein upper body clothes
Dress is divided into 14 kinds of styles, and lower body garment is divided into 6 kinds of styles, and whole body clothes fashion is divided into 3 kinds of styles, is belonged to according to different attribute
Property mark；Secondly, the style and attribute of Fashionable women dress image are described by defining visual signature descriptor, then will
4 kinds of low-level image features that visual signature descriptor and step2 are extracted carry out characteristic matching；Finally by random forest and multiclass
SVM method supervised learning is trained the training Fashionable women dress image T " after feature extraction, obtains the particulate of style and attribute
Spend classifier；
Step4 realizes fine granularity by the fine grit classification device after training, to the Fashionable women dress image I " of feature extraction
Classification exports the classification results of Fashionable women dress image.
Embodiment 2: wherein improved DPM model is made of a root model and several partial models, the object of n component
Model is expressed as (n+2) tuple (F0,P1,...Pi,...Pn, b), wherein F0It is root filter, PiIt is the mould of i-th of component
Type, b is a deviation loss coefficient, in l0Scale layer, with (x0,y0) be anchor point response score are as follows:
Wherein,For the response score of root model, viIt is a bivector, for specifying i-th of filter
Coordinate of the anchor point position (normal place when i.e. deformation occurs) relative to root position,For
The response score of n partial model, λ are the number of levels of the Feature Mapping calculated in feature pyramid with twice of resolution ratio；
After calculating response score, the response of converting member filter simultaneously considers spatial location laws, and response transform calculates public
Formula is as follows:
Wherein, (x, y) is i-th of partial model in the ideal position of scale layer, and l is the number of levels of feature pyramid H,
(dx, dy) is the offset of opposite (x, y), Ri,l(x+dx, y+dy) is matching score of the partial model at (x+dx, y+dy),
di.φd(dx, dy) is the score for deviating (dx, dy) and being lost, φd(dx, dy) and=(dx, dy, dx2,dy2) it is DPM feature, di
To deviate loss coefficient, when the parameter model for needing to learn when being model training initializes, di=(0,0,1,1) deviates damage
Lose the Euclidean distance for offset relative ideal position；
Each goal hypothesis specifies position of each filter in feature pyramid H in model: z=(p0,...,
pn), pi=(xi,yi,li) it is layer and position coordinates where i-th of filter, the score of goal hypothesis calculates as follows:
Wherein Fi'.φ(H,pi) be i-th of filter score, φ (H, pi) be feature pyramid H feature vector,
Fi' it is to connect weight vectors in i-th of filter and the vector that obtains, (dxi,dyi)=(xi,yi)-(2(x0,y0)+vi) give
Displacement of i-th of filter position relative to its anchor point position is gone out, optimal location, root is obtained by goal hypothesis score
According to the comprehensive response score of each position of best position calculation:
Multiple examples that target is detected by the comprehensive response score of each position, obtain testing result.
As shown in figure 3, distinguish in the present invention training Fashionable women dress image T' after extracting parts detection and it is to be sorted when
4 kinds of still HOG, LBP of women's dress image I', color histogram and boundary operator low-level image features, when training after obtaining feature extraction
Still women's dress image T " and Fashionable women dress image I " to be sorted.
Dimensionality reduction is carried out to feature using PCA dimension reduction method, calculates the mean value of feature vector in each dimension first, and each
Characteristic value in a dimension subtracts mean value.Then covariance matrix and the feature vector and characteristic value of the matrix are solved, and is guaranteed
Feature vector is unit vector, then will be high-dimensional under feature vector work as principal component, go out corresponding spy according to characteristics extraction
Levy vector.Suitable principal component coating ratio is finally selected, in order to guarantee information loss minimum, deletes the feature of relative distribution
Point increases whole confidence level.It is 94% that usually setting, which retains percentage value, can keeping characteristics information to greatest extent.
As shown in table 1, table 2 and Fig. 4, the particular content of step 3 is that Fashionable women dress is divided into upper body women's dress and the lower part of the body first
Women's dress, wherein upper body clothes are divided into 14 kinds of styles, and lower body garment is divided into 6 kinds of styles, and whole body clothes are divided into 3 kinds of styles；According to when
Still women's dress different attribute (such as collar, sleeve, sleeve type, color, type pattern) carries out attribute labeling.
1 Fashionable women dress style table of table
2 Fashionable women dress attribute list of table
Secondly, as shown in table 3, being carried out by defining visual signature descriptor to the style and attribute of Fashionable women dress image
Description, the descriptor are divided into upper body visual signature descriptor, lower part of the body visual signature descriptor and global characteristics descriptor.
Then the 4 kinds of low-level image features extracted visual signature descriptor and step2 carry out characteristic matching.
For different styles and attribute, the present invention defines a series of visual signature descriptors to the money of Fashionable women dress image
Formula and attribute are described, and are divided into upper body visual signature descriptor, lower part of the body visual signature descriptor and global characteristics descriptor.
Wherein, upper body feature descriptor is divided into 3 kinds of collar type, sleeve type, and lower part of the body feature descriptor is divided into length type, fold class
3 kinds of type, width type, global characteristics descriptor have a kind of style characteristics.By visual signature and low-level image feature in characteristic extraction procedure
Matching is got up, and the validity of feature extraction is improved.
3 Fashionable women dress visual signature descriptor table of table
Wherein, τ indicates trunk,In m indicate detection collar corner quantity, AτIndicate the pixel number on trunk τ
Amount,Middle D (Ik,Ig) it is different colours pixel Ik, IiBetween color distance measurement,Middle RcIndicate neck collar end,InIndicate the normal place of j-th of collar corner being detected,Middle nAIndicate the arm detected
The pixel quantity in region, flMiddle llIndicate the lower length filled,WithRespectively indicate the length of left and right leg, frMiddle nwIndicate lower dress
The quantity for the pixel that wrinkles, AlExpression fills down the whole pixel quantities detected, ftMiddle nvIndicate the quantity of lower dress vertical line pixel,In It is the lower width for filling three parts, w respectivelyω
It is the width of lumbar region.
It is carried out respectively finally by random forest (RF) and multiclass SVM algorithm according to the different styles and attribute of defined mistake
Supervised learning establishes fine grit classification device model.Random forest is the set of T decision tree, and wherein each tree is trained to every
A node rank maximizes information gain, is quantified as following form:
Wherein, H (x) is the entropy of sample set x, and t is that x is divided into subclass xlAnd xrBinary system test, class prediction is by putting down
Equal leaf distributionClass execute, L=(l1,......lT) it is leaf node on all trees.This
Invention uses the distinction learner of strong binary system SVM as division decision function t, if x ∈ RdBe a d dimension input to
Amount, w is trained SVM weight vector.SVM node is by wTAll sample decompositions of x < 0 are left side, by every other sample point
It is not divided into the child node on right side.In training, what several binary class subregions were randomly generated.For each grouping, linearly
SVM is trained for randomly selected feature channel.Finally, the division of Multi-level information gain L (x, w) is maximized, measurement selection
True label is as division function, to obtain trained Fashionable women dress style fine grit classification device.
In addition, being trained to each fine granularity attribute with one-vs-all method in multiclass SVM supervised learning
According to 47 binary classifiers of the 47 of definition kinds of Fashionable women dress attribute constructions, wherein h-th of classifier is similar remaining each i-th
Class demarcates, and h-th of classifier takes in training set h class to be positive class when training, remaining classification point be negative class be trained for
One data x for needing to classify determines that the classification of x assumes that classifier h predicts data x by the mode of ballot is used,
If obtain be positive class result, then with classifier h to x classify the result is that: x belongs to h class, if class h obtain a ticket
What is obtained is negative class result, then therefore x belongs to other class. other than h class, and each class in addition to h obtains a ticket and finally unites
Meter who gets the most votes's class is the generic attribute of x, to train Fashionable women dress attribute fine grit classification device.
The training Fashionable women dress image T " after feature extraction is carried out by random forest and Multiclass SVM method supervised learning
Training, obtains the fine grit classification device model of style and attribute.
As shown in figure 5, being realized to the Fashionable women dress image I " of feature extraction thin by the fine grit classification device after training
Grain-size classification, exports the classification results of Fashionable women dress image, and testing result shows that style and attribute are with list in the form of detection block
Only different labels are shown in classification results.
Above in conjunction with attached drawing, the embodiment of the present invention is explained in detail, but the present invention is not limited to above-mentioned
Embodiment within the knowledge of a person skilled in the art can also be before not departing from present inventive concept
It puts and makes a variety of changes.
Claims (3)
1. a kind of Fashionable women dress image fine grit classification method based on component detection and visual signature, it is characterised in that: including
Following steps:
Step1 is to the training Fashionable women dress image T and Fashionable women dress image I to be sorted of input, using improved DPM model pair
Human body under different postures and visual angle carries out component detection；Firstly, to training Fashionable women dress image T and fashion to be sorted
Women's dress image I extracts HOG and obtains DPM feature after being normalized；Secondly, according to human body attitude, visual angle adjustment DPM human body inspection
Model is surveyed, human testing model is divided into root model and partial model；Then, root model and portion are calculated separately according to DPM feature
The response score of part model carries out response transform and calculates goal hypothesis score, obtains optimal location to calculate target each
The comprehensive response score of position, finally obtains testing result；
Step2 extract respectively detection after training Fashionable women dress image T' and Fashionable women dress image I' to be sorted HOG, LBP,
4 kinds of low-level image features of color histogram and boundary operator, training Fashionable women dress image T " after obtaining feature extraction and to be sorted
Fashionable women dress image I "；
Step3 matches the visual signature descriptor of definition with the 4 kinds of low-level image features extracted, is supervised using multiclass SVM
Learning training fine grit classification device model；Fashionable women dress is divided into upper body women's dress and lower part of the body women's dress first, wherein upper body clothes point
For 14 kinds of styles, lower body garment is divided into 6 kinds of styles, and whole body clothes fashion is divided into 3 kinds of styles, carries out attribute mark according to different attribute
Note；Secondly, the style and attribute of Fashionable women dress image are described by defining visual signature descriptor, then by vision
4 kinds of low-level image features that feature descriptor and step2 are extracted carry out characteristic matching；Finally by random forest and the multiclass side SVM
Method supervised learning is trained the training Fashionable women dress image T " after feature extraction, obtains the fine grit classification of style and attribute
Device；
Step4 realizes fine grit classification by the fine grit classification device after training, to the Fashionable women dress image I " of feature extraction,
Export the classification results of Fashionable women dress image.
2. the Fashionable women dress image fine grit classification method according to claim 1 based on component detection and visual signature,
It is characterized by: improved DPM model is made of a root model and several partial models in the Step1, the object of n component
Body Model is expressed as (n+2) tuple (F0,P1,...Pi,...Pn, b), wherein F0It is root filter, PiIt is i-th of component
Model, b is a deviation loss coefficient, in l0Scale layer, with (x0,y0) be anchor point response score are as follows:
Wherein, Ro,l0(x0,y0) be root model response score, viIt is a bivector, for specifying the anchor of i-th of filter
Coordinate of the point position relative to root position,For the response score of n partial model, λ is in spy
Levy the number of levels of the Feature Mapping calculated in pyramid with twice of resolution ratio；
After calculating response score, the response of converting member filter simultaneously considers spatial location laws, and response transform calculation formula is such as
Under:
Wherein, (x, y) is the ideal position of i-th partial model in scale layer, and l is the number of levels of feature pyramid H, (dx,
It dy is) offset of opposite (x, y), Ri,l(x+dx, y+dy) is matching score of the partial model at (x+dx, y+dy), di.
φd(dx, dy) is the score for deviating (dx, dy) and being lost, φd(dx, dy)=(dx, dy, dx2,dy2) it is DPM feature, diIt is inclined
Move loss coefficient, when model initialization, di=(0,0,1,1) deviate loss be offset relative ideal position Euclidean away from
From；
Each goal hypothesis specifies position of each filter in feature pyramid H in model: z=(p0,...,pn),
pi=(xi,yi,li) it is layer and position coordinates where i-th of filter, the score of goal hypothesis calculates as follows:
Wherein Fi'.φ(H,pi) be i-th of filter score, φ (H, pi) be feature pyramid H feature vector, Fi' it is to connect
The vector for connecing the weight vectors in i-th of filter and obtaining, (dxi,dyi)=(xi,yi)-(2(x0,y0)+vi) give i-th
Displacement of a filter position relative to its anchor point position, obtains optimal location by goal hypothesis score, according to optimal
Position calculates the comprehensive response score of each position:
Multiple examples that target is detected by the comprehensive response score of each position, obtain testing result.
3. the Fashionable women dress image fine grit classification method according to claim 1 based on component detection and visual signature,
It is characterized by: the visual signature descriptor in the Step3 is divided into upper body visual signature descriptor, the description of lower part of the body visual signature
Symbol and global characteristics descriptor, and characteristic matching is carried out with 4 kinds of low-level image features in Step2 accordingly；
The upper body feature descriptor is for describing collar and sleeve, the percentage including corner in neck collar endThe x variable of all corners in neck collar endAll sides in neck collar end
The y variable at angleThe percentage of arm regions pixelThese four feature descriptors
All matched with HOG, Roberts boundary operator feature；
The lower part of the body feature descriptor is for describing length, fold and width, including the long ratio with lower dress length of legThe percentage f of lower dress region wrinkler=(nw/Al), the percentage f of lower dress region vertical linet=(nv/
Al), the ratio of lower dress and lumbar region width These four
Feature descriptor is all matched with HOG, Roberts boundary operator feature；
The global characteristics descriptor is for describing pattern, the density including corner in regionColor in region
The overall salience of varianceThe density matching LBP feature of corner in region, color variance in region
Overall salience matching color histogram feature；
Wherein m indicates the quantity of the collar corner of detection, RcIndicate neck collar end,InIt indicates to be detected for j-th
The normal place of the collar corner measured, nAIndicate that the pixel quantity of the arm regions detected, τ indicate trunk, AτIndicate trunk
Pixel quantity on τ, llIndicate the lower length filled,WithRespectively indicate the length of left and right leg, nwWrinkle pixel is loaded under expression
Quantity, AlExpression fills down the whole pixel quantities detected, nvIndicate the quantity of lower dress vertical line pixel,It is the lower width for filling three parts, w respectivelyωIt is lumbar region
Width, D (Ik,Ig) it is different colours pixel Ik, IgBetween color distance measurement.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810784023.4A CN109145947B (en) | 2018-07-17 | 2018-07-17 | Fashion women's dress image fine-grained classification method based on part detection and visual features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810784023.4A CN109145947B (en) | 2018-07-17 | 2018-07-17 | Fashion women's dress image fine-grained classification method based on part detection and visual features |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109145947A true CN109145947A (en) | 2019-01-04 |
| CN109145947B CN109145947B (en) | 2022-04-12 |
Family
ID=64800777
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810784023.4A Active CN109145947B (en) | 2018-07-17 | 2018-07-17 | Fashion women's dress image fine-grained classification method based on part detection and visual features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109145947B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110136100A (en) * | 2019-04-16 | 2019-08-16 | 华南理工大学 | The automatic classification method and device of CT sectioning image |
| CN110738233A (en) * | 2019-08-28 | 2020-01-31 | 北京奇艺世纪科技有限公司 | Model training method, data classification method, device, electronic equipment and storage medium |
| CN111508042A (en) * | 2019-01-31 | 2020-08-07 | 斯特拉德视觉公司 | Method and apparatus for generating bicycle rider safety garment pattern |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102819566A (en) * | 2012-07-17 | 2012-12-12 | 杭州淘淘搜科技有限公司 | Cross-catalogue indexing method for business images |
| CN104978762A (en) * | 2015-07-13 | 2015-10-14 | 北京航空航天大学 | Three-dimensional clothing model generating method and system |
| CN105069466A (en) * | 2015-07-24 | 2015-11-18 | 成都市高博汇科信息科技有限公司 | Pedestrian clothing color identification method based on digital image processing |
| CN105373783A (en) * | 2015-11-17 | 2016-03-02 | 高新兴科技集团股份有限公司 | Seat belt not-wearing detection method based on mixed multi-scale deformable component model |
| CN105488490A (en) * | 2015-12-23 | 2016-04-13 | 天津天地伟业数码科技有限公司 | Judge dressing detection method based on video |
| CN106022375A (en) * | 2016-05-19 | 2016-10-12 | 东华大学 | HU invariant moment and support vector machine-based garment style identification method |
| CN106021603A (en) * | 2016-06-20 | 2016-10-12 | 昆明理工大学 | Garment image retrieval method based on segmentation and feature matching |
| WO2016168235A1 (en) * | 2015-04-17 | 2016-10-20 | Nec Laboratories America, Inc. | Fine-grained image classification by exploring bipartite-graph labels |
| CN106203313A (en) * | 2016-07-05 | 2016-12-07 | 昆明理工大学 | The clothing classification of a kind of image content-based and recommendation method |
| CN106295693A (en) * | 2016-08-05 | 2017-01-04 | 深圳云天励飞技术有限公司 | A kind of image-recognizing method and device |
| CN107368832A (en) * | 2017-07-26 | 2017-11-21 | 中国华戎科技集团有限公司 | Target detection and sorting technique based on image |
| CN107729908A (en) * | 2016-08-10 | 2018-02-23 | 阿里巴巴集团控股有限公司 | A kind of method for building up, the apparatus and system of machine learning classification model |
| US9959480B1 (en) * | 2015-06-03 | 2018-05-01 | Amazon Technologies, Inc. | Pixel-structural reference image feature extraction |
-
2018
- 2018-07-17 CN CN201810784023.4A patent/CN109145947B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102819566A (en) * | 2012-07-17 | 2012-12-12 | 杭州淘淘搜科技有限公司 | Cross-catalogue indexing method for business images |
| WO2016168235A1 (en) * | 2015-04-17 | 2016-10-20 | Nec Laboratories America, Inc. | Fine-grained image classification by exploring bipartite-graph labels |
| US9959480B1 (en) * | 2015-06-03 | 2018-05-01 | Amazon Technologies, Inc. | Pixel-structural reference image feature extraction |
| CN104978762A (en) * | 2015-07-13 | 2015-10-14 | 北京航空航天大学 | Three-dimensional clothing model generating method and system |
| CN105069466A (en) * | 2015-07-24 | 2015-11-18 | 成都市高博汇科信息科技有限公司 | Pedestrian clothing color identification method based on digital image processing |
| CN105373783A (en) * | 2015-11-17 | 2016-03-02 | 高新兴科技集团股份有限公司 | Seat belt not-wearing detection method based on mixed multi-scale deformable component model |
| CN105488490A (en) * | 2015-12-23 | 2016-04-13 | 天津天地伟业数码科技有限公司 | Judge dressing detection method based on video |
| CN106022375A (en) * | 2016-05-19 | 2016-10-12 | 东华大学 | HU invariant moment and support vector machine-based garment style identification method |
| CN106021603A (en) * | 2016-06-20 | 2016-10-12 | 昆明理工大学 | Garment image retrieval method based on segmentation and feature matching |
| CN106203313A (en) * | 2016-07-05 | 2016-12-07 | 昆明理工大学 | The clothing classification of a kind of image content-based and recommendation method |
| CN106295693A (en) * | 2016-08-05 | 2017-01-04 | 深圳云天励飞技术有限公司 | A kind of image-recognizing method and device |
| CN107729908A (en) * | 2016-08-10 | 2018-02-23 | 阿里巴巴集团控股有限公司 | A kind of method for building up, the apparatus and system of machine learning classification model |
| CN107368832A (en) * | 2017-07-26 | 2017-11-21 | 中国华戎科技集团有限公司 | Target detection and sorting technique based on image |
Non-Patent Citations (4)
| Title |
|---|
| PAN HUO 等: "A Part-Based and Feature Fusion Method for Clothing Classification", 《PCM 2016: ADVANCES IN MULTIMEDIA INFORMATION PROCESSING》 * |
| SI LIU 等: "Street-to-Shop: Cross-Scenario Clothing Retrieval via Parts Alignment and Auxiliary Set", 《2012 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| 王安琪 等: "面向个性化服装推荐的判断优化模型", 《计算机工程与应用》 * |
| 黄冬艳 等: "联合分割和特征匹配的服装图像检索", 《计算机辅助设计与图形学学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111508042A (en) * | 2019-01-31 | 2020-08-07 | 斯特拉德视觉公司 | Method and apparatus for generating bicycle rider safety garment pattern |
| CN111508042B (en) * | 2019-01-31 | 2023-08-18 | 斯特拉德视觉公司 | Method and apparatus for generating a bicycle rider safety garment pattern |
| CN110136100A (en) * | 2019-04-16 | 2019-08-16 | 华南理工大学 | The automatic classification method and device of CT sectioning image |
| CN110738233A (en) * | 2019-08-28 | 2020-01-31 | 北京奇艺世纪科技有限公司 | Model training method, data classification method, device, electronic equipment and storage medium |
| CN110738233B (en) * | 2019-08-28 | 2022-07-12 | 北京奇艺世纪科技有限公司 | Model training method, data classification method, device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109145947B (en) | 2022-04-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106778584B (en) | A kind of face age estimation method based on further feature Yu shallow-layer Fusion Features | |
| Opelt et al. | Learning an alphabet of shape and appearance for multi-class object detection | |
| CN103942577B (en) | Based on the personal identification method for establishing sample database and composite character certainly in video monitoring | |
| CN106022343B (en) | A kind of clothes fashion recognition methods based on Fourier descriptor and BP neural network | |
| Gall et al. | Hough forests for object detection, tracking, and action recognition | |
| Johnson et al. | Clustered pose and nonlinear appearance models for human pose estimation. | |
| CN103824051B (en) | Local region matching-based face search method | |
| CN103514456B (en) | Image classification method and device based on compressed sensing multi-core learning | |
| CN106022375B (en) | A kind of clothes fashion recognition methods based on HU not bending moment and support vector machines | |
| Ghiasi et al. | Parsing occluded people | |
| CN104992142A (en) | Pedestrian recognition method based on combination of depth learning and property learning | |
| Hu et al. | Exploring structural information and fusing multiple features for person re-identification | |
| CN106056132B (en) | A kind of clothes fashion recognition methods based on Fourier descriptor and support vector machines | |
| CN109145947A (en) | A kind of Fashionable women dress image fine grit classification method based on component detection and visual signature | |
| CN104834941A (en) | Offline handwriting recognition method of sparse autoencoder based on computer input | |
| Khan et al. | Gait recognition using motion trajectory analysis | |
| CN106599785A (en) | Method and device for building human body 3D feature identity information database | |
| CN106778489A (en) | The method for building up and equipment of face 3D characteristic identity information banks | |
| CN106897669A (en) | A kind of pedestrian based on consistent iteration various visual angles transfer learning discrimination method again | |
| CN104050460B (en) | The pedestrian detection method of multiple features fusion | |
| CN108595558A (en) | A kind of image labeling method of data balancing strategy and multiple features fusion | |
| Li et al. | Multideep feature fusion algorithm for clothing style recognition | |
| Tong et al. | Cross-view gait recognition based on a restrictive triplet network | |
| CN109002850A (en) | The method and device of fuel value of food in a kind of calculating image | |
| CN115294660B (en) | Body-building action recognition model, training method of model and body-building action recognition method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |