CN104616291A - Sparse coding-based fabric appearance flatness evaluation method - Google Patents

Abstract

The invention provides a sparse coding-based fabric appearance flatness evaluation method. The method consists of four parts, namely low level feature detection, middle level feature design, pooling and classification. Firstly, feature point vector quantization is performed on an image by using scale invariant feature transform; secondly, the middle level feature is constructed by using a sparse coding method and a bag-of-word model, and classification of different types of fabrics is realized by using a support vector machine. According to the method, automatic rating of fabric wrinkles in an unsupervised state can be effectively realized.

Description

A kind of appearance of fabrics flatness evaluation method based on sparse coding

Technical field

Under the present invention relates to fabric pincher state, wrinkle grade machine vision evaluates problem, is specifically related to the establishment of fabric pincher feature character code storehouse, wrinkle sorting algorithm, realizes polytype appearance of fabrics flatness automatic measure grading.

Background technology

Flat appearance degree, as the important aesthetic parameter of one, fabric, becomes fabric quality and controls, the design of washing machine, dryer and automatic pressing ironing equipment, washing agent wash the important evaluating basis protecting effect assessment etc.The crease intensity of user satisfaction and clothes, fabric is comparatively close.Fabric pincher degree uses the signs such as flatness, crease-resistant grade, recovery angle usually, and wherein flatness embodies the fluctuating quantity of fabric face, usually in order to characterize the appearance after fabric nursing.

Woven fabric, knitted fabric and supatex fabric, macroscopically tow sides are all planes.Processes such as wearing experience consumer, wash, be dry, be subject to mechanical stretching, compression, bending, shearings, friction etc., fibrous inside stress and tension force continuity continuous decrease in time, shows creep, relaxes, form plastic yield, fabric face deforms.Because fabric respectively produces contraction in various degree in local, fabric, by producing the concavo-convex of entirety, presents the concavo-convex difference of centimetre-sized.Fabric pincher degree evaluation is a traditional problem, usually adopts artificial vision to evaluate with reference to relevant criterion (as AATCC 124).The method belongs to subjective assessment, is subject to the impact of individual physiology, psychology and environmental factor, and precision is low, repeatability is poor, is difficult to meet current detection demand.Therefore need by setting up relevant quick, objective, accurate evaluation method.

Accurate and effective objective method is the effective way evaluating appearance of fabrics flatness.Way common is at present from some angle, extracts local feature, carries out mathematical description, convert, obtain characteristic index to original spatial domain or frequency-domain information a few class physical propertys of image model, carries out comprehensive and evaluation on this basis.The index of current application is still difficult to the crease intensity evaluating fabric exactly, certain difference is still there is with actual wrinkle, though there is certain correlativity between every evaluation index and wrinkle grade, but because the fold morphology of template surface is totally different, fold distribution more complicated, the wrinkle grade of fabric reality only effectively can not be evaluated by the simple combination of several index, be difficult to simulate the sensory evaluation reaching people, between subjective and objective, consistance is poor.Merely extract some characteristic indexs and carry out comprehensive evaluation, there will be the deviation with true wrinkle effect.Some index truly may not reflect fabric property.Therefore propose to adapt to this special object of fabric and to be satisfied with the evaluation index of human body vision particularly important.

In expert opinion process, be based upon on overall, comprehensive, fuzzy basis to the assessment of sample, even have references to the whole evaluation that historical experience is carried out, the non-minutia by a certain classification determined.Therefore how to hold and utilize Given information, comprehensively going out judgment basis and can realize above judgement, seeming particularly important, and comprehensively this, should be effective representative of textile image or three-dimensional spatial information.Human visual system can process a large amount of visual information rapidly, biological explanation mammal generates in long-term evolution can be quick, accurately, low-cost ground represents the ability of the optic nerve aspect of natural image, rebuilds and store piece image by less cost.Therefore, digital picture can be proposed by similar neural coding model, the linear superposition noise model of things target image by basis function can be described.

Summary of the invention

The object of this invention is to provide a kind of evaluation method of the crease intensity to fabric of the sensory evaluation close to people.

In order to achieve the above object, technical scheme of the present invention there is provided a kind of appearance of fabrics flatness evaluation method based on sparse coding, comprises the following steps:

Step 1, extract the low-level image feature of each textile image, comprise the following steps:

Step 1.1, structure metric space, realize metric space extreme value;

Step 1.2, the extreme point of searching image on each yardstick are as unique point, and the extreme point under current scale is in current scale and bilevel 26 fields, the maximum or minimum value found out;

Step 1.3, by the three-dimensional quadratic function of matching to determine position and the yardstick of each unique point, remove low contrast and unstable edge respective point simultaneously, to strengthen coupling stability, improve noise resisting ability, then give each unique point assignment direction parameter;

Step 1.4, the positional information by unique point, residing dimensional information and directional information, determine corresponding SIFT feature region, by carrying out piecemeal to the SIFT feature region around unique point, computing block inside gradient histogram, thus generate unique SIFT feature vector, each SIFT feature vector is a SIFT descriptor x, thus every width iamge description is the set X=[x of SIFT descriptor ₁..., x _m] ^t∈ R ^{m × D}, in formula, M is total number of SIFT descriptor, and D is the dimension of feature space;

Step 2, set X=[x to the SIFT descriptor that step 1.4 obtains ₁..., x _m] ^t∈ R ^{m × D}utilize sparse coding algorithm to obtain one group of base vector and more efficiently represent X, namely dictionary, each base of recycling feature coding to dictionary distributes weight the most suitable, expression sample characteristics that can be correct;

The method of step 3, use support vector machine is divided into the wrinkle classification preset to the dictionary that step 2 obtains, thus sets up the character code storehouse of different wrinkle grade;

After step 4, acquisition textile image, realize the wrinkle grade assessment of different fabric based on character code storehouse.

Can carry out automatic Evaluation to the flatness of fabric by method provided by the invention, and the opinion rating obtained is more consistent with artificial vision's evaluation result.

Accompanying drawing explanation

Fig. 1 is the character code storehouse schematic diagram based on sparse coding;

Fig. 2 is the spatial extrema process schematic based on difference of Gaussian;

Fig. 3 is the rarefaction representation schematic diagram of textile image;

Fig. 4 is the visual vocabulary schematic diagram after training based on 30 1 grade of images;

Fig. 5 A is the visual vocabulary schematic diagram of 1.5 grades;

Fig. 5 B is the visual vocabulary schematic diagram of 2 grades;

Fig. 5 C is the visual vocabulary schematic diagram of 2.5 grades;

Fig. 5 D is the visual vocabulary schematic diagram of 3 grades;

Fig. 5 E is the visual vocabulary schematic diagram of 3.5 grades;

Fig. 5 F is the visual vocabulary schematic diagram of 4 grades;

Fig. 5 G is the visual vocabulary schematic diagram of 4.5 grades;

Fig. 5 H is the visual vocabulary schematic diagram of 5 grades.

Embodiment

For making the present invention become apparent, hereby with preferred embodiment, and accompanying drawing is coordinated to be described in detail below.Should be understood that these embodiments are only not used in for illustration of the present invention to limit the scope of the invention.In addition should be understood that those skilled in the art can make various changes or modifications the present invention, and these equivalent form of values fall within the application's appended claims limited range equally after the content of having read the present invention's instruction.

In order to make those of ordinary skill in the art understand the present invention better, first some making an explanation property of concept are wherein illustrated:

(1) scale invariant feature conversion: scale invariant feature conversion (Scale-invariant featuretransform, SIFT) is a kind of descriptor for image processing field.This description has scale invariability, can detect key point in the picture.It is a kind of local description.

(2) sparse coding: sparse coding is a kind of Artificial Neural Network simulating mammalian visual systems main visual cortex V1 district simple cell receptive field.The method has the band general character of the locality in space, directivity and frequency domain, is a kind of adaptive image statistics method.

(3) sparse coupling: the characteristic matching of image is also known as sparse coupling, and main employing gray scale matching by cross correlation carries out the coupling of marginal point or angle point.

In one exemplary embodiment of the present invention, propose a kind of appearance of fabrics wrinkle assessment method based on sparse coding.This method is made up of low-level image feature detection, middle level features design, pond and classification four parts.The character code storehouse study schematic diagram of Fig. 1 institute embodiment of the present invention.First, (English full name is Scale-invariant feature transform to adopt scale invariant feature to change, referred to as SIFT) unique point vector quantization is carried out to image, adopt the mode of sparse coding and word bag model to carry out the structure of middle level features, use the method for support vector machine to realize the classification of dissimilar fabric.Below these parts are specifically described:

Step one: low-level image feature extracts

First build metric space, realize metric space extreme value.At this, be L (x, y, σ) by graphical rule definition space, calculate by formula (1) and obtain:

L(x，y，σ)＝G(x，y，σ)*I(x，y) (1)

$G(x,y,\mathrm{\σ})=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}{e}^{-(x^2+y^2)/2{\mathrm{\σ}}^2}---\left(2\right)$

Wherein G (x, y, σ) refers to the variable Gaussian function of yardstick, and (x, y) is volume coordinate, and the size of σ decides the smoothness of image, and I (x, y) represents input picture.

In order to find the extreme point of metric space, each pixel of image will compare with comparison diagram image field and scale domain consecutive point.As shown in Figure 2, extreme point is found out.The searching of extreme point, is exactly in this yardstick and bilevel 26 fields, finds out maximum or minimum value, just thinks for being the unique point of image under this yardstick.Conveniently find stable unique point, adopt Gaussian difference scale space (DOG scale-space) to carry out the Gaussian difference pyrene of different scale and image convolution generates.Use DoG to find Min-max than faster with simple, formula is as follows:

D(x，y，σ)＝(G(x，y，kσ)-G(x，y，σ))*I(x，y)

(3)

＝L(x，y，kσ)-L(x，y，σ)

By the three-dimensional quadratic function of matching to determine position and the yardstick (reaching sub-pixel precision) of unique point, remove low contrast and unstable edge respective point simultaneously, to strengthen coupling stability, improve noise resisting ability.More than obtain the unique point in every width figure, then give unique point assignment one 128 dimension direction parameter, utilize the gradient direction of unique point field pixel to be divided into for each unique point assigned direction parameter, its operator has rotational invariance.Adopt formula (7) and (8) to calculate and obtain direction parameter:

$m(x,y)=\sqrt{{(L(x+1,y)-L(x-1,y))}^2+{(L(x,y+1)-L(x,y-1))}^2}---\left(4\right)$

θ(x，y)＝tan ^-1((L(x，y+1)-L(x，y-1))/(L(x+1，y)-L(x-1，y))) (5)

M (x, y) and θ (x, y) is respectively modulus value and the direction formula of SIFT feature point (x, y) place gradient, and L () is the yardstick for each SIFT feature point place separately.Unique point has three information: position, and residing yardstick, direction have so far detected complete all, can determine a SIFT feature region thus.By carrying out piecemeal to the SIFT feature region around unique point, computing block inside gradient histogram, thus generate unique vector, SIFT vector.Each SIFT feature vector is a SIFT descriptor x, thus every width iamge description is the set X=[x of SIFT descriptor ₁..., x _m] ^t∈ R ^{m × D}, in formula, M is total number of SIFT descriptor, and D is the dimension of feature space.Make image form different blocks by segmentation, intensive or random acquisition, key point or the mode such as stabilized zone, marking area, and obtain the feature at each piece of place.

Step 2: middle level features design learns with visual vocabulary

Middle level features designs based on SIFT descriptor statistic histogram.

Loosen if we and must input equal restriction output, utilize the concept of base in linear algebra, i.e. O=a1* Φ 1+a2* Φ 2+....+an* Φ n, Φ i is base, and ai is coefficient simultaneously, and we can obtain such optimization problem:

Min|I-O| (6)

Wherein I represents input, and O represents output.A signal is expressed as the linear combination of one group of base, and requires only to need less several bases just signal can be showed.Sparse coding algorithm is a kind of unsupervised learning method, and it is used for searching one group " super complete " base vector and more efficiently represents sample data.

Feature coding is exactly distribute weight (sparse) the most suitable, expression sample characteristics that can be correct to each base of dictionary.Given a series of samples pictures, we need study obtain one group of base [Φ 1, Φ 2 ...], namely dictionary.Assuming that the set of X a kind of SIFT descriptor that is step, for set X=[x ₁..., x _m] ^t∈ R ^{m × D}constraint function is as follows:

$\underset{P,Q}{\min }\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{\left|\right|x_m-p_{m}Q\left|\right|}^2+\mathrm{\λ}\left|p_m\right|$

Wherein: $p_m\≥0;\left|\right|Q_k\left|\right|\≤1,\∀k=\mathrm{1,2},...,K---\left(7\right)$

In formula (7), Q is code set, p _mfor m the projection coefficient of character code Q.Training process is exactly the process of an iteration, by said above, and the change p that we replace _mmake this objective function above minimum with Q.

Each iteration is in two steps: a) fixing dictionary Q, then adjusts p _m, make formula (7), namely objective function is minimum.B) then p is fixed _m, adjustment Q, make formula (7), namely objective function is minimum.Continuous iteration, until convergence.So just can obtain the base that a group well can represent this series of x, namely dictionary.Be a regional area/descriptor characteristic set by every width iamge description.Feature coding is exactly distribute weight (sparse) the most suitable, expression sample characteristics that can be correct to each base of dictionary.

Based on block pyramid, adopt the mode of random pool, utilize probability coefficent to try to achieve final character code.P _ifor weight is sparse, pond formula is:

$z=\underset{i\∈R}{\mathrm{\Σ}}{p}_i{u}_m---\left(8\right)$

Vectorial z ∈ R in formula (8) ^m, u _mfor Pixel Information.

In this example, 9 classes are selected, every class 50 wrinkle images.Every class image has the screening of detection expert evaluation.Extract wherein each 30 images and train use as visual vocabulary, another 20 × 9 images are used for test.Often open image and adopt 2000 SIFT descriptors, these descriptors are extracted.Each shell is of a size of 8 × 8 pixels, amounts to 16 × 16 shells.Each shell is all adopted and is realized gray processing process.By 400 iteration, after visualization process, as shown in Figure 4.

Step 3: fabric pincher is classified

Support vector machine (SVM) is a kind of classification and homing method of having supervision.In a n-dimensional space, input two class data, support vector machine constructs a lineoid within this space and is used for differentiation two class data, and the border of these lineoid distance two class data is maximum.For calculating the distance of two class data, build in the both sides of classifying face the lineoid that two are parallel to classifying face, the data on these two lineoid are called support vector.

K sorter (k=9), m class and remaining class are separated by m sorter, and be that is 1 by m class again label, other class labels are-1.Complete this process need and calculate k quadratic programming, according to label by each sample separately, what finally export is that binary classifier exports as that maximum class.

$k(z_i,z_j)=z_i^T{z}_j=\underset{l=0}{\overset{2}{\mathrm{\&Sigma;}}}\underset{s=1}{\overset{2^l}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{2^l}{\mathrm{\&Sigma;}}}<z_i^l(s,t),z_j^l(s,t)>---\left(9\right)$

$f\left(z\right)={\left(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{z}_i\right)}^{T}z+b=w^{T}z+b---\left(10\right)$

In formula (9) and formula (10), z _irepresent image i maximum pondization statistics description encoding, z _jrepresent image j maximum pondization statistics descriptor, k (z _i, z _j) represent the above two dot product, s and t be spatial domain scope, represent that image i is under 1 level, the maximum pondization of (s, t) scope add up description encoding, represent that image j is under 1 level, the maximum pondization statistics description encoding of (s, t) scope, f (z) represent that two class SVM express, a _iexpression SVM coefficient, b represent regulated value.

Unconfined convex optimization problem:

In formula (9) and formula (10), w _crepresent floor coefficient, herein with two norms expression, J (w _c) represent unconfined convex Optimal Expression, represent penalty coefficient, z that user specifies _ithe same, expression constraint condition, C are constant.

In this example, 180 images are used for test, and accuracy of the mean reaches 95.4%, and standard variance is that 0.053. is more consistent with artificial vision's evaluation result, can realize the automatic measure grading of fabric pincher.

Claims (1)

1., based on an appearance of fabrics flatness evaluation method for sparse coding, comprise the following steps:

Step 1, extract the low-level image feature of each textile image, comprise the following steps:

Step 1.1, structure metric space, realize metric space extreme value;

Step 1.2, the extreme point of searching image on each yardstick are as unique point, and the extreme point under current scale is in current scale and bilevel 26 fields, the maximum or minimum value found out;

Step 1.3, by the three-dimensional quadratic function of matching to determine position and the yardstick of each unique point, remove low contrast and unstable edge respective point simultaneously, to strengthen coupling stability, improve noise resisting ability, then give each unique point assignment direction parameter;

Step 1.4, the positional information by unique point, residing dimensional information and directional information, determine corresponding SIFT feature region, by carrying out piecemeal to the SIFT feature region around unique point, computing block inside gradient histogram, thus generate unique SIFT feature vector, each SIFT feature vector is a SIFT descriptor x, thus every width iamge description is the set X=[x of SIFT descriptor ₁..., x _m] ^t∈ R ^{m × D}, in formula, M is total number of SIFT descriptor, and D is the dimension of feature space;

Step 2, set X=[x to the SIFT descriptor that step 1.4 obtains ₁..., x _m] ^t∈ R ^{m × D}utilize sparse coding algorithm to obtain one group of base vector and more efficiently represent X, namely dictionary, each base of recycling feature coding to dictionary distributes weight the most suitable, expression sample characteristics that can be correct;

The method of step 3, use support vector machine is divided into the wrinkle classification preset to the dictionary that step 2 obtains, thus sets up the character code storehouse of different wrinkle grade;

After step 4, acquisition textile image, realize the wrinkle grade assessment of different fabric based on character code storehouse.