CN109993213A - A kind of automatic identifying method for garment elements figure - Google Patents
A kind of automatic identifying method for garment elements figure Download PDFInfo
- Publication number
- CN109993213A CN109993213A CN201910169562.1A CN201910169562A CN109993213A CN 109993213 A CN109993213 A CN 109993213A CN 201910169562 A CN201910169562 A CN 201910169562A CN 109993213 A CN109993213 A CN 109993213A
- Authority
- CN
- China
- Prior art keywords
- point
- garment elements
- measured
- feature
- edge
- 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
- 238000000034 method Methods 0.000 title claims abstract description 86
- 238000010586 diagram Methods 0.000 claims abstract description 80
- 238000000605 extraction Methods 0.000 claims abstract description 18
- 230000008569 process Effects 0.000 claims abstract description 14
- 238000007781 pre-processing Methods 0.000 claims abstract description 5
- 239000011159 matrix material Substances 0.000 claims description 48
- 230000009466 transformation Effects 0.000 claims description 26
- 230000014509 gene expression Effects 0.000 claims description 20
- 238000013139 quantization Methods 0.000 claims description 19
- 238000012545 processing Methods 0.000 claims description 18
- 238000012935 Averaging Methods 0.000 claims description 15
- 230000011218 segmentation Effects 0.000 claims description 13
- 238000003708 edge detection Methods 0.000 claims description 12
- 230000008859 change Effects 0.000 claims description 8
- 238000004458 analytical method Methods 0.000 claims description 7
- 238000006243 chemical reaction Methods 0.000 claims description 7
- 238000003860 storage Methods 0.000 claims description 7
- 238000012706 support-vector machine Methods 0.000 claims description 4
- 230000008901 benefit Effects 0.000 claims description 2
- 238000007689 inspection Methods 0.000 claims description 2
- 238000013461 design Methods 0.000 abstract description 12
- 238000011161 development Methods 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 11
- 238000010606 normalization Methods 0.000 description 9
- 238000012549 training Methods 0.000 description 9
- 239000000284 extract Substances 0.000 description 8
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000011160 research Methods 0.000 description 7
- 238000012360 testing method Methods 0.000 description 7
- 238000005452 bending Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 238000013178 mathematical model Methods 0.000 description 4
- 238000001514 detection method Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000001914 filtration Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000003909 pattern recognition Methods 0.000 description 2
- 238000013519 translation Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000001174 ascending effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000004883 computer application Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000005194 fractionation Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 238000003707 image sharpening Methods 0.000 description 1
- 230000002401 inhibitory effect Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 238000000513 principal component analysis Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
- 230000007306 turnover Effects 0.000 description 1
- 230000000007 visual effect 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
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- 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
-
- 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
- G06V10/48—Extraction of image or video features by mapping characteristic values of the pattern into a parameter space, e.g. Hough transformation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
Abstract
The present invention relates to the technical fields of image procossing, disclose a kind of automatic identifying method for garment elements figure, comprising the following steps: Step 1: establishing the garment elements chart database of multiple types；Step 2: pre-processing to garment elements figure to be measured, binaryzation garment elements figure is obtained；Step 3: carrying out vectorized process to the binaryzation garment elements figure, corresponding vector part figure to be measured is obtained；Step 4: carrying out feature extraction to the vector part figure to be measured, then the feature extracted is normalized, Classification and Identification then is carried out to normalized vector part figure to be measured, combined data library obtains matching component diagram, completes automatic identification.The present invention constructs the frame of garment elements figure identification, provides thinking for automatic identification, software implementation development and application, also designs for intelligent clothing and saves a large amount of manpower and time, promotes the design efficiency of clothes early period.
Description
Technical field
The present invention relates to the technical field of image procossing more particularly to a kind of automatic identification sides for garment elements figure
Method.
Background technique
Since Design System of Garment digitlization, intelligentized demand and image procossing have in computer application technology
Good application prospect and market value, some scholars have some research achievements at present, below mainly from clothes fashion or component
The method for extracting shape features and classifying identification method of figure, to summarize domestic and international present Research.
(1) Shape Feature Extraction of garment elements figure
The outer profile of clothes shows the difference of apparel modeling feature to varying degrees, it is to discriminate between a weight of clothes
Feature is wanted, while its resemblance is depicted, therefore to identify garment elements figure or garment elements classification, will be examined first
Survey clothes edge contour.Extract dress planar component diagram Edge Feature Points as Li Ke is fearful et al., with algorithm judge node and
Functional relation, then the garment elements of generation are carried out mathematical description, to realize clothing department by automanual fractionation clothes profile
The identification of edge line feature in part figure.Ji Xiaoyan et al. proposes the components such as clothing collar, sleeve, front and rear panel are main with theirs
Control point and backbone graph-based, establish garment elements knowledge base, and the designer of serve future, which goes to splice new clothes, to be set
Meter figure etc..But these researchs require in advance to analyze the style or component information to be identified, establish characteristic model,
Then these features are extracted, cannot quickly and effectively be described.Later period have scholar merge other contour feature extracting methods into
Row research, An et al. are extracted garment elements map contour using morphology, describe son with Wavelet Fourier and carry out Expressive Features, from
And realize the Classification and Identification of garment elements figure.Bending moment and Fourier (Fourier Descriptor) be not special with Hu by Hou et al.
It levies to extract and describe the edge line feature of clothes.Bending moment Expressive Features are not retouched by the Fourier that is used in the above research, Hu
Son is stated, can be used as the feature of Classification and Identification component diagram, however they cannot but show the concave-convex geometric characteristic to go beyond the scope
With the curvature of curve.
(2) garment shape tagsort recognition methods
Needs are classified and identified to obtained feature to after clothes fashion or component diagram extraction feature, about classification
Method by initial constructor model and clustering, then develop to the existing pattern classification with computer technology
Device.Wang et al. carrys out Classification and Identification with fuzzy theory and goes out image of clothing, is progressively scanned by the method and obtains clothes edge shape
The key point of shape constructs the fuzzy clustering matrix of seven characteristic values based on these key points to Classification and Identification.Meanwhile Wang Xiu
Treasure et al. analyzes the characteristic of Bodice structure point, constructs print data model, obtains about print characteristic, will be to be measured
Print and standard sample of photo carry out match cognization.Above research be with measure formulas, contour line more to be measured with establish
Database contour line between diversity factor, to identify garment elements figure, recognition accuracy is heavily dependent on edge
The detection of contour line, entire identification process be not cumbersome and time-consuming long.An et al. trains classification with ExtremeLearningMachine classifier
Carried out Wavelet Fourier description of component diagram edge detection.Hou is in the research that image of clothing is retrieved, with European measurement
Function compares the diversity factor between two images.The efficiency and classification adaptability of these types of pattern-recognition, all need further to verify.
To sum up, also under study for action with computer image processing technology and the garment elements figure Classification and Identification of pattern-recognition, do not have also currently
Have that adaptability is good, classification accuracy is high and the comprehensive method of Feature Descriptor classifies to dress planar style or component diagram
And identification.
Summary of the invention
The present invention provides a kind of automatic identifying methods for garment elements figure, and it is numerous to solve existing recognition methods process
It is trivial, time-consuming, extracts Feature Descriptor not comprehensively, the problems such as accuracy rate is poor.
The present invention can be achieved through the following technical solutions:
A kind of automatic identifying method for garment elements figure, comprising the following steps:
Step 1: establishing the garment elements chart database of multiple types；
Step 2: pre-processing to garment elements figure to be measured, binaryzation garment elements figure is obtained；
Step 3: carrying out vectorized process to the binaryzation garment elements figure, corresponding vector part figure to be measured is obtained；
Step 4: carrying out string feature extraction to the vector part figure to be measured, then the string feature extracted is subjected to normalizing
Then change processing carries out Classification and Identification to normalized vector part figure to be measured, combined data library obtains matching component
Figure completes automatic identification.
Further, the method that vector part figure to be measured is obtained in the step 3 includes:
Step I carries out edge for the binaryzation garment elements figure using the edge detection method based on Canny operator
Detection；
Step II parses the edge detected, and stores the point on edge at point set sequence according to parsing sequence
Column carry out feature judgement to all the points inside the sequence of point sets using Hough transform, if meeting conllinear feature, to it
It is labeled, otherwise, does not mark；
Step III carries out straightway vector quantization to the point of mark, to the point not marked using segmentation three bezier curve
Carry out vector quantization.
Further, the analytic method in the step II includes: firstly, being looked for sequence analytic method the edge detected
To first marginal point as starting point, then eight neighborhood analytic method is taken all to parse all marginal points, then according to
Sequential storage is parsed into sequence of point sets.
Further, the step 4 specifically includes the following steps:
Step I, the string feature that each vector part figure to be measured is extracted using string feature extracting method, obtain three string spies
Levy matrix, respectively averaging projection's length matrix PM, outer chord length matrix OM and inside and outside chord length absolute value of the difference matrix IODM；
Step II, the normalized that three string eigenmatrixes are translated respectively, rotated and scaled；
Step III is normalized all parts figure in database using the method for step I and step II, then
Classify using based on support vector machine svm classifier method to all parts figure after the normalized in database, and sentences
Break the generic of each vector part figure to be measured；
Step IV accurately identifies vector part figure to be measured using nearest neighbor method 1NN in generic, find out with
The most matched component diagram of vector part figure to be measured completes automatic identification.
Further, enabling the sequence of point sets is set C={ Pi(xi,yi), i=1 ..., N }, wherein N=2T, and T takes just
The function expression of integer, the binaryzation component diagram is set asX is respectively indicated with y
The transverse and longitudinal coordinate of pixel in component diagram, D indicate region of the edge of component diagram in component diagram,
The expression formula of the outer chord length matrix OM is arranged are as follows:
The expression formula setting of the inside and outside chord length absolute value of the difference matrix IODM are as follows:
The expression formula of averaging projection's length matrix PM is arranged are as follows:
Wherein,Indicate outer chord length,String in indicating
It is long, A=(ys+i-yi)/ls,i, B=(xs+i-xi)/ls,i, C=(xiys+i-yixs+i)/ls,i, δ (Δ) expression Dirac function, x1
And x2Respectively indicate xiAnd xi+sMinimum value and maximum value, y1And y2Respectively indicate yiAnd yi+sMinimum value and maximum value, s indicate
Unit length number, value 21,22,...,2T-1,It indicates from the point P on edgei
(xi,yi) set out, reach another point P counterclockwise along edgei+s(xi+s,yi+s) passed through chord length, Ax+By+C=0 indicates
Point PiAnd Pi+sThe normal equation of the straight line determined,Averaging projection head of the expression arc to string
Degree,Indicate stringPoint P on corresponding arci+tTo the projector distance of the string.
Further, to the normalized of scaling, with averaging projection length matrix PM, outer chord length matrix OM and inside and outside chord length
The maximum value of every a line in absolute value of the difference matrix IODM carries out normalizing operation to each element of locating row；Rotation is returned
One change processing, is carried out using Fourier transformation.
Further, by averaging projection length matrix PM, outer chord length matrix OM and inside and outside chord length absolute value of the difference matrix IODM
Every a line be all used as an one-dimensional discrete signal, one-dimensional discrete Fourier change then is carried out to the one-dimensional discrete signal
It changes, the row before replacing transformation with the sequence of the mould of Fourier Transform Coefficients, wherein the coefficient expressions of Fourier transformation are arranged such as
Under:
Further, the preprocess method of the step 2 includes: using the Laplace operator based on second-order differential
Laplace operator is sharpened processing to garment elements figure to be measured, and setting threshold value is the garment elements to be measured after 0.6 corresponding sharpening
Figure carries out binary conversion treatment.
The present invention is beneficial to be had the technical effect that
By being pre-processed to garment elements figure to be measured, binary conversion treatment, vectorized process chord feature extraction, to mentioning
The string feature taken is normalized, and the vector part figure to be measured after normalization is carried out Classification and Identification, combined data library is obtained
Matching component diagram, complete automatic identification, to construct the frame of garment elements figure identification, be automatic identification,
Software implementation development and application provides thinking, also designs for intelligent clothing and saves a large amount of manpower and time, promotes clothes early period
Design efficiency.Line vectorization based on HOG transformation and the curve vector based on three sections of Bezier curves, carry out edge point
Vector quantization storage, can be applied not only to the electronics original text component of the software designs such as CAD, also can be applied to manuscript component, expand
The application range of method of the invention, the extracting method based on string eigenmatrix CFM being capable of accurate description component diagram wheels
Convex and concave feature, the bending features of exterior feature, anti-noise ability is strong, to comprehensively describe the geometrical characteristic of profile, mentions for subsequent identification
For basis, the thick identification based on svm classifier method and accurately identifying based on 1NN classifying identification method find Feature Points Matching
Metric form in the process improves the accuracy rate of Classification and Identification.
Detailed description of the invention
Fig. 1 is flow chart of the invention；
Fig. 2 is segmentation three bezier curve figure of the invention；
Fig. 3 is the pixel point set figure of component one side of something collar of the invention；
Fig. 4 is P on three bezier curve controlling polygon and edge-of-part of the inventionmPnThe corresponding relationship of section；
Fig. 5 is after crew neck neck of the invention extracts eigenmatrix progress Fourier transformation, and choosing scale grade S respectively is 4
Preceding 5,10,18,32,40 description are reconstructed, the figure after obtaining this few component statement reconstruct；
Fig. 6 is after turndown collar of the invention extracts eigenmatrix progress Fourier transformation, and choosing scale grade S respectively is 4
Preceding 5,10,18,32,40 description are reconstructed, the figure after obtaining this few component statement reconstruct；
Fig. 7 is low frequency coefficient number and classification accuracy relational graph after use Fourier transformation normalization of the invention；
Fig. 8 is for use of the invention based on support vector machine svm classifier method to the three of turndown collar, common lapel and flat collar
The assorting process and classification results figure that a string eigenmatrix OM, IODM, PM classify, wherein mark a is indicated to turndown collar
Assorting process figure, mark b indicate that the assorting process figure to common lapel, mark c indicate the assorting process figure to flat collar, mark
D presentation class result figure；
Fig. 9 obtains three string features after a certain crew neck feature normalization after operation program in MATLAB to be of the invention
Matrix data distribution map, wherein mark A indicates the data profile of string eigenmatrix IODM, and mark B indicates string eigenmatrix OM
Data profile, mark C indicate string eigenmatrix PM data profile.
Specific embodiment
With reference to the accompanying drawing, specific embodiments of the present invention are further elaborated.
Attached drawing 1 is flow chart of the invention, and the present invention provides a kind of automatic identifying method based on garment elements figure, packets
Include following steps:
Step 1: establishing the garment elements chart database of multiple types.
The sample database of garment elements figure of the invention derives from two parts: a part is provided by Shanghai PGM company
The garment elements figure made carry out component segmentation and obtain；Another part is the garment elements figure made by inventor.
The style of each component of clothes is ever-changing, by taking collar as an example, collar can be divided into lapel, turn over stand-up collar, collar, turndown collar, stand-up collar,
Connect stand-up collar, without neck etc., wherein the collar of many types can be subdivided into below the collar of every kind of rough classification again, it is each type of
Collar component makes 60 respectively, identifies for subsequent feature extraction and part classification.
Computer Image Processing is usually to use various image processing algorithms, to each pixel in the image of being handled
It is operated, to reach certain desired effect.Due to needing each pixel to image to grasp during image processing
Make, the characteristic information of image itself is formed to the processing of pixel, convenient for the image recognition in later period, therefore in image preprocessing
Multinomial pretreatment operation, such as denoising, enhancing, gray processing, recovery, segmentation, edge detection would generally be carried out to image.With
The component diagram of method identification of the invention, without texture, color and Lighting information, the mainly profile information of clothes, so
In image preprocessing, it is only necessary to be sharpened to component diagram, binaryzation, edge detection and vector quantization, mainly pass through following steps
Two and three realize, to prepare for subsequent Feature extraction and recognition.The marginal information of acquisition is the identification of garment elements figure
Key, the smooth and order of accuarcy of boundary curve will affect the accuracy of subsequent shapes feature extraction and description.
Step 2: being pre-processed to garment elements figure to be measured, binaryzation garment elements figure is obtained, is specifically first carried out sharp
Change, then carries out binary conversion treatment.
From the point of view of the source of component diagram sample, component diagram for identification is not the same size specification, can generate pressure
Image resolution ratio is different during compression process or preservation, and it is not high that this will lead to some component diagram resolution ratio, it is therefore desirable into
The sharpening of row garment elements figure, the purpose of sharpening are that original fuzzy figure is allowed to become more fully apparent, and visual effect is more preferable.
The Robert that image sharpening method common at present is mainly based upon first differential is sharpened, Sobel is sharpened,
Prewitt Edge contrast operator and Laplace operator based on second-order differential sharpen, in their template, each coefficient and all
It is zero.This shows that these operators in the response in the constant region of gray scale are original image in 0, that is, image after Edge contrast
The smooth region of picture is substantially black, and in original image object edge, details and Gray Level Jump point all can in black background with
High gray portion highlights.After running program in MATLAB, effect of the component diagram under the processing of different sharp filtering operators
It is compared as follows table.
Upper table can be seen that Sobel operator and Prewitt operator can enhance in sharp filtering highlight component diagram edge letter
Breath, plays sharpening effect, but the part edge of component diagram shows unclear, such as collar, the profile of waist two sides, does not reach increasing
Strong effect.Prewitt operator widens component diagram edge, is almost distorted；And the characteristics of Sobel operator is a symmetrical scale
Point, certain smoothing effect is played to center weighting, the lines of outline after Sobel operator processing component figure is calculated compared to Prewitt
Lines of outline after subprocessing slightly broadens.Compared to first two sharp filtering operator, Laplace operator obscures component diagram
Part sharpened, the boundary of collar and waistline two sides is more obvious, with Laplace can with the details of strengthening part figure,
Its edge is found, therefore, the present invention selects Laplace operator Laplace to be sharpened component diagram.
Due to testing component diagram inherently black white image used, so not needing to carry out ash to the component diagram after sharpening
Degree processing, directly progress binary conversion treatment, the purpose of binary conversion treatment are for the subsequent image vector based on HOG transformation
Precondition.Common binaryzation function is im2bw in MATLAB, and it is 1,0.6,0.2 that threshold value, which is respectively set, and operation program obtains
It is as shown in the table to different threshold binarization images.
As can be seen from the above table when threshold value is 1, garment elements map contour is thicker, and many noises are taken as target figure
As processing, the place that the circle in table image as above is demarcated will affect subsequent edge detection in this way；When threshold value takes 0.2,
Profile line thinning, but the unconspicuous local contour line in the boundaries such as collar, muffs is intermittent, the circle institute in table image as above
The place of calibration, profile are no longer closure lines, this can directly hinder component segmentation.When threshold value takes 0.6, it can be seen that profile
Line is clearly demarcated and is all the lines of closure, and noise does not also occur.Therefore present invention setting threshold value 0.6 carries out binary conversion treatment.
Step 3: carrying out vectorized process to above-mentioned binaryzation garment elements figure, multiple vector part figures to be measured are obtained.
Garment elements figure is sharpened with after binaryzation, needs to be partitioned into each component from entire garment elements figure with that.Clothes
The segmentation result of component diagram is filled as edge extracting, the basis of component identification and vector quantization, the quality of component segmentation decides side
Edge contours extract quality and vector quantization effect accurately whether.Below to be partitioned into collar and sleeve from whole garment elements figure
For, concrete operations are as follows:
Step I carries out edge detection to component diagram using the edge detection method based on Canny operator.
Profile extracted to component diagram, the detection of profile point and extraction are the important steps of vector quantization, contours extract it is accurate
Degree will directly influence the precision of subsequent vector, and edge detection can reduce the data volume of pixel, and retain component
Edge shape information, it is contemplated that the advantage and disadvantage of various edge detection methods, the present invention using based on Canny operator edge inspection
Survey method obtains the marginal information of all parts after segmentation.
Step II parses the edge detected, and stores the point on edge at point set sequence according to parsing sequence
Column carry out feature judgement to all the points inside above-mentioned sequence of point sets using Hough transform, if meeting conllinear feature, to it
It is labeled, otherwise, does not mark, specific as follows:
After carrying out edge detection to component, what is obtained is the marginal information of component, wherein 0 indicates edge letter for white
Breath, 1 indicates background information for black.Component wants vector quantization, it is necessary to sequential storage is carried out to edge sequence of point sets, component diagram
Point set storage, the edge contour of component is actually stored into sequence of point sets in the order of connection, be stored as sequence of point sets it
Before need to parse component.There are commonly sequence parsing and neighborhood parsing, binding sequence parsing of the present invention and eight neighborhood solution
The method of analysis uses sequential search method to find first marginal point as starting point, then in the component diagram being partitioned into first
Eight neighborhood method is taken all to parse all marginal points, and by parsing sequential storage at sequence of point sets.
After the completion of sequence of point sets storage, vector quantization can be carried out to it, vector quantization is exactly with simple figure or to use number
The form of equation is learned to indicate garment elements, generallys use straight line and free curve to indicate.Straightway mathematical model is simply easy
Understand, can be indicated with following parametric equation:
P (t)=P1t+P2(1-t)
Wherein, P1、P2Respectively indicate line segment starting point and ending point, parameter t ∈ [0,1].
For curved section mathematical model, there are commonly curvilinear interpolations in production apparel construction figure and design style system
Method, B-spline curve, Bezier curve etc..Since Bezier curve is the shape with control point come controlling polygon, make song
The design and modification of line are very intuitive, and being segmented indicates that measurement accuracy can be made higher.So the present invention uses segmentation three times
Bezier curve indicates the free curve of component diagram edge-of-part, realize from geometric meaning with component diagram curved edge etc.
Effect, segmentation three bezier curve is as shown in Fig. 2, its matrix indicates as follows:
P (t)=TMBQ
Wherein, T=[t3 t2T 1],T ∈ [0,1], Q are that control is polygon
The control point of shape.
Hough transform be find straightway feature a kind of efficient algorithm, principle be by rectangular coordinate system point (x,
Y) point (r, θ) being transformed into polar coordinate system, using all the points, that is, collinear points in rectangular coordinate system on same straight line (r,
Same this feature of point, the judgement of Lai Jinhang straight line θ) are shown as in plane.Straight line expression formula y=mx+b in rectangular coordinate system
It is as follows with polar coordinate representation:
ri=xicosθi+yisinθi
To judge whether the sequence of point sets of a certain component diagram is straight line, r- θ plane can be quantized into several small lattice, according to
After quantized value after (x, y) point substitution θ known to each, r is calculated by above formula, then knows that (r, θ) value is in a certain grid,
The quantity of this small lattice is added 1 at this time, when all (x, y) point after aforesaid operations, counts the quantity of small lattice, statistic compared with
Big small lattice just correspond to collinear points, and this point (r, θ) is the pole coordinate parameter of required straight line, using this parameter as judging part
Part marginal point whether be straightway parameter.
Step III carries out straightway vector quantization to the point of mark, to the point not marked using segmentation three bezier curve
Carry out vector quantization.
It include below straight line P with a width0Pm, straight line P0PnWith curve PmPnCollar component diagram for illustrate the mistake of vector quantization
Journey, Fig. 3 are the pixel point set figure of component one side of something collar.If to collar edge pixel is obtained after the edge detection process of collar component
Sequence of point sets is as follows:
P0Pm={ P0,P1,P2,...,Pm-2,Pm-1,Pm}；
PmPn={ Pm,Pm+1,Pm+2,...,Pn-1,Pn}；
P0Pn={ Pn,Pn+1,Pn+2,...,Pn+k-2,Pn+k-1,Pn+k}；
The feature of collar pixel sequence of point sets is sought using Hough transform, wherein P0PmSection and P0PnSection respectively meets conllinear spy
Sign, makes marks to its reference point, PmPnSection does not meet conllinear feature, does not make marks.Straightway is carried out to the pixel point set to make marks
Vector quantization, to P in the pixel point set such as Fig. 3 not made marksmPnSection carries out vector quantization with segmentation three bezier curve.
During free curve vector quantization, since Bezier curve is determined by the polygon that control point determines,
And the known only neck collar end point of the present invention, the parameter of Bezier curve is obtained, with regard to all control point Q must be found out.For
Ensure that be fitted Bezier curve can be by the point in neck collar end, it need to be with the partial dot on edge-of-part come reverse control
The control point of polygon, Fig. 4 are P on three bezier curve controlling polygon and edge-of-partmPnThe corresponding relationship of section,
Shown in mathematical model following equation:
Wherein, P1、P2、P3、P4For collar marginal point, Q1、Q2、Q3、Q4For the control point of controlling polygon, as shown in Figure 4,
P1、P2、P3、P4And Q1、Q4For known point, Q2、Q3For unknown point, t1、t2For known point P2、P3Corresponding parameter value, they
Value is by taken P2、P3Position determine.If P1、P2、P3、P4Spacing is equal, then can find out t1、t2.Solving above-mentioned equation group can ask
4 control points of controlling polygon out, so that three bezier curve is obtained, thus vector quantization free curve.
Step 4: carrying out feature extraction to each vector part figure to be measured, then place is normalized in the feature extracted
Then reason carries out Classification and Identification, combined data library, to obtain matching component to normalized vector part figure to be measured
Figure completes automatic identification.
Component diagram feature extraction of the invention is mainly the extraction of geometric characteristic, the edge contour shape base of component diagram
This is all the shape feature for being made of straight line and free curve, therefore using component diagram edge lines, wherein containing edge lines
The shape features such as curvature, concavity, convexity, feature description is carried out to component diagram to be measured.At present for image object identification
Have there are a variety of methods in profile curves feature extraction and description, wherein the method based on curvature feature extraction is that relatively tool represents
Property, it is a kind of shape feature for more intuitively showing image object contour line bending degree.
The string eigenmatrix that Wang Bin was proposed in 2016 describes sub- CFM to carry out Classification and Identification to leaf, has obtained very well
Recognition effect, leaf and component diagram extract feature and have in common that it is all the geometry for extracting target image, therefore
The method that the present invention selects feature extraction and the description of Wang Bin, the feature of each component of component diagram is described based on string eigenmatrix,
This method is the method for the description neck collar end line convex and concave feature of indirect, is compared in component diagram component feature extraction process
Other methods are advantageously.Feature about indirect describes, at present useful chord length, the point on edge to farthest point apart from etc.
As the parameter of description, these parameters reflect the flexural property of object edge line, but edge line indirectly to some extent
Concavity and convexity cannot also describe well.The string feature extracting method that the present invention uses describes each feature of component diagram, such
Method is also referred to as CFM, and this character description method can describe the concavity and convexity of collar component outline, wherein by extracting multiple rulers
Feature Descriptor of the inside and outside chord length of the string of grade as style component is spent, and collar edge-of-part is depicted in inside and outside chord length well
Concavity and convexity, so that the convex and concave feature of more accurate description component outline curve, this method not only depict neck collar end
Concavity and convexity, and the arc introduced, to averaging projection's length parameter of string, this parameter attribute can portray each collar fashion portion
The curvature of part curved edge can completely describe edge-of-part wheel by portraying for the above convexity, concavity and flexural property
Wide geometric characteristic.
Specifically includes the following steps:
Step I, the feature that each vector part figure to be measured is extracted using string feature extracting method, obtain three string features
Matrix, respectively averaging projection's length matrix PM, outer chord length matrix OM and inside and outside chord length absolute value of the difference matrix IODM.
Enabling sequence of point sets is set C={ Pi(xi,yi), i=1 ..., N }, wherein N=2T, and T takes positive integer, binaryzation
The function expression of component diagram afterwards is set asX and y respectively indicate the picture in component diagram
The transverse and longitudinal coordinate of vegetarian refreshments, D indicate region of the edge of component diagram in component diagram,
The expression formula of outer chord length matrix OM is arranged are as follows:
The expression formula of inside and outside chord length absolute value of the difference matrix IODM is arranged are as follows:
The expression formula of averaging projection length matrix PM is arranged are as follows:
Wherein,Indicate outer chord length,String in indicating
It is long, A=(ys+i-yi)/ls,i, B=(xs+i-xi)/ls,i, C=(xiys+i-yixs+i)/ls,i, δ (Δ) expression Dirac function, x1
And x2Respectively indicate xiAnd xi+sMinimum value and maximum value, y1And y2Respectively indicate yiAnd yi+sMinimum value and maximum value, s indicate
Unit length number, value 21,22,...,2T-1, unit length be defined as the point on edge therewith between most consecutive points away from
From, the distance is related with the size of sequence of point sets,It indicates from the point P on edgei(xi,
yi) set out, reach another point P counterclockwise along edgei+s(xi+s,yi+s) passed through chord length, Ax+By+C=0 indicates point Pi
And Pi+sThe normal equation of the straight line determined,Indicate arc to string averaging projection's length,Indicate stringPoint P on corresponding arci+tTo the projector distance of the string.
From the mathematical model of CFM it is found that three matrixes are all (T-1) × N, s often take a value just have a scale with
Correspondence, that is, description of string eigenmatrix has (T-1) a scale.It follows that the string eigenmatrix of the method is every
A line all corresponds to the distinctive scale grade of every a line.Wherein the 1st row of eigenmatrix is exactly smallest dimension grade i.e. s=21, matrix
Last line is exactly out to out grade i.e. 2T-1, wherein T=log2N.The geometric characteristic information of different scale grade feature description
Difference, the small characteristic information portrayed of scale is more careful, and the shape information of the big description of scale is relatively coarse, from string feature square
From the point of view of battle array building, it is from 21Scale, increase to 2 at Geometric SequenceT-1, thus can be the shape of whole part profile
Shape feature, which is comprehensively depicted, to be come.In addition, the present invention is identified to garment elements figure, the spy between collar component contour line
It is very big to levy similarity, so many minutias will be extracted to be distinguished, such as much without neck, crew neck, V neck, U-shaped neck etc.
Deng in the salient point of these component outlines, concave point, inflection point, the regions such as extreme point just need small size features to portray.
It can be seen that, after marginal point determines, it is more careful that the smaller entire marginal information of scale describes from following table；Scale
The monnolithic case for only describing profile greatly, may miss minutia when profile details are more；When scale-value is identical, marginal point
Quantity more multiple features are more, and edge details describe more complete, but if marginal point is excessive, Riming time of algorithm is more
It is long.
Step II, the normalized that above three string eigenmatrix is translated respectively, rotated and scaled.
Since the geometry of flat components figure is under translation, rotation, the transformation of scaling, geometry be will not change,
Therefore before the classification and identification to collar component, need to by the string Feature Descriptor normalized of garment elements, avoid because
These geometric transformations occur causes to describe sub- difference, finally causes part classification mistake, identification inaccuracy.
(1) component diagram translates
Value, interior exterior string from above-mentioned OM, IODM, PM formula it can be found that when component translates, in outer chord length matrix
Value in long absolute value of the difference and average three matrixes of projected length be to maintain it is constant, so not having to do the variation of translation to return
One change processing.
(2) component diagram scales
When component diagram scales, the value of string will change.Assuming that the string characteristic function that component is constituted is by g (x, y)
Become g (ax, ay), wherein a > 0 is zoom factor.With Feature Descriptor OM, every a line in tri- matrixes of IODM, PM is most
Big value operates each element normalizing of the row.By constituting it is found that the characteristic element for being in same a line is all corresponding for eigenmatrix
Same scale grade, the corresponding different scale grade of different rows, therefore the present invention respectively operates every a line normalizing, so that it may guarantee
The feature of T-1 scale grade, that is, each row feature has same contribution in component identification, the wheel because of certain a line would not occurs
Characteristics of Profile value is excessive and makes other lesser failures of row contour line characteristic value.
(3) component diagram rotates
When the collar component diagram of experiment is rotated, the string on edge-of-part line also all generates same rotation, according to
Before the expression formula of OM, IODM, PM are able to know that the inside and outside chord length of edge line, averaging projection's length of arc to string with rotation.But
It is that the rotation of collar component diagram can allow the initial position of the edge line of extraction to generate variation, and then allow the three of string Feature Descriptor
Each row generation of a matrix is shifted serially.If t indicates translational movement, wherein 1≤t≤N is translational movement, then the m in three matrixes
Row, the characteristic value of kth column will be moved to m row, kth+t column.For this problem, the present invention uses one-dimensional discrete Fourier
Transformation makees normalized to the rotation of collar component diagram.
Firstly, all regard every a line in OM, IODM, PM as an one-dimensional discrete signal, then carry out it is one-dimensional from
Fourier transformation is dissipated, the row before replacing transformation with the sequence of the mould of Fourier Transform Coefficients.The wherein description system of Fourier transformation
Number formula is as follows:
Wherein, u=0,1,2 ..., M-1, at the beginning of signal, the value a (0) after Fourier transformation is referred to as Fu
In leaf transformation DC component.By the theory of Fourier transformation it is found that if component is rotated, corresponding to one-dimensional signal also has phase
Should translate, the mould of acquired Fourier transformation coefficient | a (u) | it is constant, so postrotational string eigenmatrix and not rotating
Component is the same through the matrix that Fourier is converted.The present invention obtains after making Fourier transformation to the every row element of string eigenmatrix
The vector length obtained is N, to highlight the characteristic quantity of main style component and inhibiting noise jamming, chooses preceding M low frequency coefficient, M
Value will be determined in subsequent experimental.Because the coefficient taken is more, although style component feature can be described preferably, can increase
Add the susceptibility to noise, and Classification and Identification calculation amount can be increased；Conversely, if the low frequency coefficient extracted is less, then extracting
Characteristic quantity can not accurately distinguish the shape of component.Three matrixes after Fourier transformationWithThe value of every a line all done a Fourier transformation, transformed coefficient of correspondence value has M, divides this M value
Other modulus respectively corresponds each transformed value to get the value to three every a line of matrix: | a (0), | a (1) | ..., | a (M-1)
|.Finally, the string eigenmatrix of component all will become the matrix that size is (T-1) × M from the matrix that size is (T-1) × N.
Figures 5 and 6 are respectively that crew neck and turndown collar extract after eigenmatrix carries out Fourier transformation, and choosing scale S respectively is 4
Preceding 5,10,18,32,40 Fourier descriptors are reconstructed, the figure after obtaining this few component statement reconstruct.From figure
As can be seen that if the length component of the Fourier descriptor of selection is more, effect picture after reconstruct, can be with closer to original image
See when length component number is 18, it is substantially identical with original image, if but the length component of Fourier descriptor that uses
Excessively, although the shape after reconstruct is finer, calculation amount when Classification and Identification is also increase accordingly, whereas if choose
It is very coarse to the description of its collar shape if length component is fewer, so in subsequent component diagram Classification and Identification, need into
Row many experiments determine suitable low frequency coefficient number.
Step III is normalized all parts figure in database using the method for step I and step II, then
Classify using based on support vector machine svm classifier method to all parts figure after the normalized in database, and sentences
Break the generic of each vector part figure to be measured.
In more classification problems of SVM, multi-categorizer is usually the building that combined by multiple two-value sub-classifiers.
The present invention belongs to the more classification problems of SVM for classifying to eight kinds of collar components it is necessary to construct eight two-value subclassification moulds
Type, about i-th of disaggregated model, the present invention is by the i-th class collar component as one kind, i.e., positive class, remaining seven class collar component
Collar component feature x to be sorted is sequentially input to each SVM two-value subclassification model in classification as another kind of, i.e., negative class
In, which SVM output valve maximum compared, if k-th of output valve is big, collar component to be sorted is just classified as kth class.
The specific implementation steps are as follows:
(1) step of SVM training: the eigenmatrix of respectively clothes collar portion part to be trained is inputted；The suitable input of selection is empty
Between kernel function；Solving optimization equation obtains supporting vector and corresponding Lagrangian；Utilize clothing collar component feature library
In a supporting vector, bring discriminant function into, obtain class label.
(2) the step of svm classifier: garment elements feature to be measured is inputted；It is calculated using Lagrange trained in step (1)
Son, deviation and input space kernel function, solve optimal discriminant function；Classification is exported according to the value of sgnf (X).
SVM classifier Kernel Function of the invention chooses RBF kernel function, after feature normalization, obtains OM, IODM, PM tri-
A eigenmatrix is not that a numerical value is constituted, but is made of the two-dimensional matrix of (T-1) × M.Because every in training sample
A line is as a sample, so needing for the amount of three eigenmatrixes to be sequentially connected into a line constitutes a component feature amount.It is first
It is first, it is then, IODM and PM is each as characteristic quantity that the T-1 row of OM matrix is sequentially connect into one full line of composition behind every a line
From T-1 row sequentially connect behind every a line constitute a full line as characteristic quantity, finally, the full line feature that IODM is obtained
It is connected to behind OM, by being connected to behind IODM for PM, finally obtained each component sample has the characteristic quantity of (T-1) * M*3.By
The quantity and scale of point on above-mentioned edge are tested it is found that the detail edges point variation of component diagram is not special frequent, institute
With default marginal point N=64, i.e. T=log of the invention2N=6.
In order to choose the low frequency coefficient number of Fourier transform normalization appropriate, that is, the value of M, the present invention carries out
Preliminary experiment: it is random from eight kinds of styles of collar component library respectively to extract 40 collar component diagrams, it is pre- to carry out above-mentioned component diagram
Processing and string eigenmatrix is extracted, to feature normalization, wherein based on Fourier transformation come when being normalized, characteristics of low-frequency
Coefficient number selects 1,2,3,4,5 respectively ..., and 60, as shown in fig. 7, carrying out component diagram training and class test, classification with this
Accuracy rate is as shown in Figure 7.As shown in Figure 7, when coefficient number is 16, recognition accuracy basically reaches highest, and what is chosen is
If number is more, program runtime also can be longer accordingly, therefore, low frequency coefficient M selection 16 of the invention.
By that can obtain above, each string eigenmatrix is the two-dimensional matrix of 5*16 now, so when input SVM training, often
A kind of string eigenmatrix has 80 characteristic quantities, obtains every kind of component by 240 spies in conjunction with three string eigenmatrixes OM, IODM, PM
Sign amount is constituted, and then carries out principal component analysis to feature set, and choosing the principal component that contribution rate of accumulative total is 95% is feature vector, right
Respective string feature and comprehensive feature carry out Classification and Identification respectively, obtain the accuracy rate under respective Feature Descriptor.It randomly selects
The feature vector of 66.7% sample is training set in collar part library, and the feature vector of remaining 33.3% sample is test set,
In each component quantity it is consistent with overall ratio, test data and training data are in addition to the mark that feature vector also includes this style
Label, see the table below.
By three string eigenmatrix OM, IODM of turndown collar, common lapel and flat collar, PM, for carry out experiment test,
Assorting process and the following Fig. 8 of classification results have found that turndown collar classification is substantially all correct, common lapel and flat collar from Fig. 8
There are three classification errors, but whole recognition accuracy has still accomplished 95%, and wherein turndown collar recognition accuracy is 100%.
Step IV accurately identifies vector part figure to be measured using nearest neighbor method 1NN in generic, to look for
Out with the most matched component diagram of vector part figure to be measured.
It is good to component diagram part classification effect based on the Classification and Identification of SVM, but it can not be same by style component to be measured
The similarity size discrimination of each individually component comes in library, that is, when identify which kind of belongs to when, in intelligent clothing
In Design modelling not enough, it is also necessary to the style in will be similar distinguishes, this just need by nearest neck Classification and Identification method into
One step will be similar in similarity size sequence distinguish.
Nearest neighbor method 1NN is a kind of Nonparametric Identification Method, does not need to provide prior probability in advance and class conditional probability is close
The knowledge such as function are spent, but directly obtained component feature matrix is operated, particularly suitable for more classification problems.
Assuming that there is M ω1,ω2,...,ωMThe collar component of classification identifies that target, every one kind have the sample of mark classification
NiA (i=1,2 ..., M), can specify that ωiThe discriminant function of class isWherein,Footmark i
Indicate ωiClass, k indicate ωiClass NiK-th in a sample.Decision rule indicates are as follows:
It is to classify with SVM classifier to component sample database, but have certain defect above, in contrast, KNN is calculated
Method principle is fairly simple, it does not have to train template and test identification based on template, is directly trained instead, above
On the basis of component diagram through point good classification, component diagram to be identified is calculated with the distance between known class component diagram, then
It is accumulative that the K the smallest ballot mark that carries out is chosen in these distances, is seen in K nearest component diagrams shared by which base part
Number is more, which kind of is just classified as.But KNN algorithm is also defective in identification, when middle number of components of all categories is not equal, if a certain
There are many collar number of components of class, and when other class quantity are seldom, it is possible to cause when testing collar component to be identified, it should
Quantity shared by that a kind of component diagram in K closer distance of component more than capacity is more[53].In order to avoid this problem, this hair
It is bright use nearest neighbor method Classification and Identification, if compare at a distance between the component chart database of N number of known class, and decision with from
Its nearest component is similar, and this differentiates the style classification nearest from it, referred to as 1NN classifying identification method.Because 1NN is identified
Before need to know in advance known style classification, so need to classify to feature database before carrying out 1NN identification, that is,
Need to carry out SVM training classification before carrying out 1NN identification.
When carrying out 1NN Classification and Identification to component diagram shape, before use classes discriminant function, the present invention then will be wait sentence
Other component diagram shape carries out Diversity measure with the shape in known features part library, then carries out size ratio to otherness
Compared with processing.After the feature normalization operation and SVM classifier processing of front, trained feature database classification is it is known that string is special
Sign matrix has met feature normalization, if judging whether two style components belong to same category, it is only necessary to comparing unit
Feature i.e. their string feature, thus the similitude to compare them.Assuming that having extracted 3 squares of the string feature CFM of lapel A
Battle array OM, IODM, PM areWith3 matrixes of the string feature CFM of lapel B to be identified
OM, IODM, PM are respectivelyWithThen their diversity factor following formula[42]Carry out table
Show:
Wherein, N indicates the point number on edge, takes 64, T=log2N=6 can be seen from the foregoing M=16.So with
Upper three eigenmatrixes are the matrix of 5*16.Fig. 9 is after obtaining a certain crew neck feature normalization after running program in MATLAB
Three string eigenmatrix data profiles, wherein the rectangle frame in figure is size of data range, the lines pair inside rectangle frame
The value answered is the mean value of each scale value of series, and as seen from Figure 8, this crew neck characteristic low-frequency data is substantially less than 1.
Following algorithm flow is obtained by the principle of 1NN algorithm:
(1) training component feature set and classification are initialized；
(2) data and style component feature data to be identified of known class are stored；
(3) setup parameter k, k takes 1 here；
(4) category feature data to be measured are calculated with the diversity factor of each component feature data of known class；
(5) ascending sort is made to training component figure characteristic data set according to diversity factor size；
(6) the smallest first k trained style component of diversity factor size is chosen, them is then counted and adds up in of all categories
Number；
(7) the accumulative maximum classification of number is returned, need to only choose the smallest classification of diversity factor here, then collar to be measured
Component just belongs to the classification where the component；
(8) L is a apart from lesser known class serial number before taking out, these classifications may be used for subsequent design and develop.
Picture library to be prepared in advance below, each component for therefrom choosing collar is tested, when running 1NN,
Diversity factor size identification result between the parts to be tested figure and the component diagram of the data store internal of point good class is as shown in the table,
In, the leftmost side is collar component diagram to be measured in table, and the corresponding numerical value below the component diagram of data store internal is their difference
Degree.It can be seen that, the diversity factor size between collar component diagram and the component diagram of data store internal to be identified is with it from table
Shape have close connection, in intelligent clothing Design modelling, designer can be according to required selection target component.
The present invention will identify that the classification based on SVM is by unknown multidimensional to garment elements figure in conjunction with both modes
The parts data library of degree multiple features classification carries out the identification that classification is used for sample to be tested, and subsequent 1NN Classification and Identification is mainly
Then known class calculates with a distance from component style of the tested sample from each classification and judges it is which style portion belonged to
Part, the latter can more specifically identify most matched component diagram, facilitate designer rapidly and accurately to find out required style, mention
High design efficiency, the identification of the identification of other garment elements with the above collar component.In addition it can suitable according to diversity factor size
Sequence arrangement part figure, the selection for having multiplicity in clothes intelligent design in this way provide.
Although specific embodiments of the present invention have been described above, it will be appreciated by those of skill in the art that these
It is merely illustrative of, without departing from the principle and essence of the present invention, a variety of changes can be made to these embodiments
It more or modifies, therefore, protection scope of the present invention is defined by the appended claims.
Claims (8)
1. a kind of automatic identifying method for garment elements figure, it is characterised in that the following steps are included:
Step 1: establishing the garment elements chart database of multiple types；
Step 2: pre-processing to garment elements figure to be measured, binaryzation garment elements figure is obtained；
Step 3: carrying out vectorized process to the binaryzation garment elements figure, corresponding vector part figure to be measured is obtained；
Step 4: carrying out string feature extraction to the vector part figure to be measured, then place is normalized in the string feature extracted
Then reason carries out Classification and Identification to normalized vector part figure to be measured, combined data library obtains matching component diagram,
Complete automatic identification.
2. the automatic identifying method according to claim 1 for garment elements figure, it is characterised in that in the step 3
The method for obtaining vector part figure to be measured includes:
Step I carries out edge inspection for the binaryzation garment elements figure using the edge detection method based on Canny operator
It surveys；
Step II parses the edge detected, and stores the point on edge at sequence of point sets, benefit according to parsing sequence
Feature is carried out to all the points inside the sequence of point sets with Hough transform to judge, if meeting conllinear feature, it is marked
Note, otherwise, does not mark；
Step III carries out straightway vector quantization to the point of mark, is carried out to the point not marked using segmentation three bezier curve
Vector quantization.
3. the automatic identifying method according to claim 2 for garment elements figure, it is characterised in that in the step II
Analytic method include: firstly, use sequence analytic method to find first marginal point as starting point at the edge detected, then
Eight neighborhood analytic method is taken all to parse all marginal points, then according to parsing sequential storage at sequence of point sets.
4. the automatic identifying method according to claim 2 for garment elements figure, it is characterised in that the step 4 tool
Body the following steps are included:
Step I, the string feature that each vector part figure to be measured is extracted using string feature extracting method, obtain three string feature squares
Battle array, respectively averaging projection's length matrix PM, outer chord length matrix OM and inside and outside chord length absolute value of the difference matrix IODM；
Step II, the normalized that three string eigenmatrixes are translated respectively, rotated and scaled；
Step III is normalized all parts figure in database using the method for step I and step II, then uses
Classified based on support vector machine svm classifier method to all parts figure after the normalized in database, and judged each
The generic of a vector part figure to be measured；
Step IV accurately identifies vector part figure to be measured using nearest neighbor method 1NN in generic, find out with it is to be measured
The most matched component diagram of vector part figure completes automatic identification.
5. the automatic identifying method according to claim 4 for garment elements figure, it is characterised in that: enable the point set sequence
It is classified as set C={ Pi(xi,yi), i=1 ..., N }, wherein N=2T, and T takes positive integer, the function of the binaryzation component diagram
Expression formula is set asX and y respectively indicates the transverse and longitudinal coordinate of the pixel in component diagram, D
Indicate region of the edge of component diagram in component diagram,
The expression formula of the outer chord length matrix OM is arranged are as follows:
The expression formula setting of the inside and outside chord length absolute value of the difference matrix IODM are as follows:
The expression formula of averaging projection's length matrix PM is arranged are as follows:
Wherein,Indicate outer chord length,Chord length in indicating, A=
(ys+i-yi)/ls,i, B=(xs+i-xi)/ls,i, C=(xiys+i-yixs+i)/ls,i, δ (Δ) expression Dirac function, x1And x2Point
It Biao Shi not xiAnd xi+sMinimum value and maximum value, y1And y2Respectively indicate yiAnd yi+sMinimum value and maximum value, s indicate unit
The number of length, value 21,22,...,2T-1,It indicates from the point P on edgei(xi,
yi) set out, reach another point P counterclockwise along edgei+s(xi+s,yi+s) passed through chord length, Ax+By+C=0 indicates point Pi
And Pi+sThe normal equation of the straight line determined,Indicate arc to string averaging projection's length,Indicate stringPoint P on corresponding arci+tTo the projector distance of the string.
6. the automatic identifying method according to claim 4 for garment elements figure, it is characterised in that: to the normalizing of scaling
Change processing, with each in averaging projection length matrix PM, outer chord length matrix OM and inside and outside chord length absolute value of the difference matrix IODM
Capable maximum value carries out normalizing operation to each element of locating row；To the normalized of rotation, using Fourier transformation into
Row.
7. the automatic identifying method according to claim 6 for garment elements figure, it is characterised in that: by averaging projection head
Degree matrix PM, outer chord length matrix OM and inside and outside chord length absolute value of the difference matrix IODM every a line be all used as one it is one-dimensional discrete
Then signal carries out one-dimensional discrete Fourier transformation to the one-dimensional discrete signal, with the sequence of the mould of Fourier Transform Coefficients
Column replace the row before transformation, and wherein the coefficient expressions of Fourier transformation are provided that
8. the automatic identifying method according to claim 1 for garment elements figure, it is characterised in that the step 2
Preprocess method includes: to be carried out using the Laplace operator Laplace operator based on second-order differential to garment elements figure to be measured
Edge contrast, setting threshold value are that the garment elements figure to be measured after 0.6 corresponding sharpening carries out binary conversion treatment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910169562.1A CN109993213A (en) | 2019-03-06 | 2019-03-06 | A kind of automatic identifying method for garment elements figure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910169562.1A CN109993213A (en) | 2019-03-06 | 2019-03-06 | A kind of automatic identifying method for garment elements figure |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109993213A true CN109993213A (en) | 2019-07-09 |
Family
ID=67130136
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910169562.1A Pending CN109993213A (en) | 2019-03-06 | 2019-03-06 | A kind of automatic identifying method for garment elements figure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109993213A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110610499A (en) * | 2019-08-29 | 2019-12-24 | 杭州光云科技股份有限公司 | Method for automatically cutting local detail picture in image |
| CN111858997A (en) * | 2020-06-23 | 2020-10-30 | 浙江蓝天制衣有限公司 | Clothing pattern generation method based on cross-domain matching |
| CN112488174A (en) * | 2020-11-26 | 2021-03-12 | 江苏科技大学 | Intelligent retrieval method for similar parts |
| CN113505410A (en) * | 2021-07-13 | 2021-10-15 | 高瑞彤 | Intelligent garment design system |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102184413A (en) * | 2011-05-16 | 2011-09-14 | 浙江大华技术股份有限公司 | Automatic vehicle body color recognition method of intelligent vehicle monitoring system |
| CN104268543A (en) * | 2014-10-04 | 2015-01-07 | 上海工程技术大学 | Pilling image collection method for complex-texture fabric |
| CN105117739A (en) * | 2015-07-29 | 2015-12-02 | 南京信息工程大学 | Clothes classifying method based on convolutional neural network |
| CN106056132A (en) * | 2016-05-19 | 2016-10-26 | 东华大学 | Garment style identification method based on Fourier descriptor and support vector machine |
| CN106384126A (en) * | 2016-09-07 | 2017-02-08 | 东华大学 | Clothes pattern identification method based on contour curvature feature points and support vector machine |
| CN106897724A (en) * | 2015-12-18 | 2017-06-27 | 南京财经大学 | A kind of plant leaf identification method based on contour line shape facility |
-
2019
- 2019-03-06 CN CN201910169562.1A patent/CN109993213A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102184413A (en) * | 2011-05-16 | 2011-09-14 | 浙江大华技术股份有限公司 | Automatic vehicle body color recognition method of intelligent vehicle monitoring system |
| CN104268543A (en) * | 2014-10-04 | 2015-01-07 | 上海工程技术大学 | Pilling image collection method for complex-texture fabric |
| CN105117739A (en) * | 2015-07-29 | 2015-12-02 | 南京信息工程大学 | Clothes classifying method based on convolutional neural network |
| CN106897724A (en) * | 2015-12-18 | 2017-06-27 | 南京财经大学 | A kind of plant leaf identification method based on contour line shape facility |
| CN106056132A (en) * | 2016-05-19 | 2016-10-26 | 东华大学 | Garment style identification method based on Fourier descriptor and support vector machine |
| CN106384126A (en) * | 2016-09-07 | 2017-02-08 | 东华大学 | Clothes pattern identification method based on contour curvature feature points and support vector machine |
Non-Patent Citations (3)
| Title |
|---|
| 安立新: "服装款式图提取及其模式识别的研究", 《中国博士学位论文全文数据库》 * |
| 杨杰 等: "服装样片图像中直线与曲线矢量化的研究", 《天津工业大学学报》 * |
| 王斌 等: "弦特征矩阵：一种有效的用于植物叶片图像分类和检索的形状描述子", 《计算机学报》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110610499A (en) * | 2019-08-29 | 2019-12-24 | 杭州光云科技股份有限公司 | Method for automatically cutting local detail picture in image |
| CN110610499B (en) * | 2019-08-29 | 2020-10-20 | 杭州光云科技股份有限公司 | Method for automatically cutting local detail picture in image |
| CN111858997A (en) * | 2020-06-23 | 2020-10-30 | 浙江蓝天制衣有限公司 | Clothing pattern generation method based on cross-domain matching |
| CN111858997B (en) * | 2020-06-23 | 2024-04-16 | 浙江蓝天制衣有限公司 | Cross-domain matching-based clothing template generation method |
| CN112488174A (en) * | 2020-11-26 | 2021-03-12 | 江苏科技大学 | Intelligent retrieval method for similar parts |
| CN113505410A (en) * | 2021-07-13 | 2021-10-15 | 高瑞彤 | Intelligent garment design system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yue et al. | Automatic detection of feather defects using Lie group and fuzzy Fisher criterion for shuttlecock production | |
| CN110021028A (en) | A kind of automatic clothing method based on garment fashion drawing | |
| CN109993213A (en) | A kind of automatic identifying method for garment elements figure | |
| CN105389593B (en) | Image object recognition methods based on SURF feature | |
| Alsmadi et al. | Fish recognition based on robust features extraction from size and shape measurements using neural network | |
| Guo et al. | Efficient center voting for object detection and 6D pose estimation in 3D point cloud | |
| CN106384126B (en) | Clothes fashion recognition methods based on contour curvature characteristic point and support vector machines | |
| CN101140623A (en) | Video frequency objects recognition method and system based on supporting vectors machine | |
| Berretti et al. | 3d mesh decomposition using reeb graphs | |
| CN109447100A (en) | A kind of three-dimensional point cloud recognition methods based on the detection of B-spline surface similitude | |
| Zhang et al. | A statistical approach for extraction of feature lines from point clouds | |
| Xie et al. | Fabric defect detection method combing image pyramid and direction template | |
| Efraty et al. | Facial component-landmark detection | |
| Pham et al. | Accurate junction detection and characterization in line-drawing images | |
| Deng et al. | Efficient 3D face recognition using local covariance descriptor and Riemannian kernel sparse coding | |
| CN110008920A (en) | A kind of facial expression recognizing method research | |
| Liu et al. | A novel rock-mass point cloud registration method based on feature line extraction and feature point matching | |
| Chen et al. | Mesh convolution: a novel feature extraction method for 3d nonrigid object classification | |
| CN106778491B (en) | The acquisition methods and equipment of face 3D characteristic information | |
| CN105760828A (en) | Visual sense based static gesture identification method | |
| CN109886091A (en) | Three-dimensional face expression recognition methods based on Weight part curl mode | |
| Artola et al. | Glad: A global-to-local anomaly detector | |
| Rathore et al. | CBISC: a novel approach for colon biopsy image segmentation and classification | |
| CN115588178B (en) | Automatic extraction method for high-precision map elements | |
| Lin et al. | In defense of iterated conditional mode for hyperspectral image classification |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190709 |
|
| WD01 | Invention patent application deemed withdrawn after publication |