Objective Fitting Evaluation Model for Dressing Fit Based on Wrinkle Index of Dressing Image

ZHANG Mengmeng()， ZHUANG Meiling()*， ZHANG Xiaofeng()

1 College of Textiles and Clothing， Qingdao University， Qingdao 266071， China 2 College of Computer Science and Technology， Qingdao University， Qingdao 266071， China

Abstract： An effective model(image to wrinkle， ITW) for garment fitting evaluation is presented. The proposed model is to improve the accuracy of garment fitting evaluation based on dressing image. The ITW model is an objective evaluation model of fitting based on the wrinkle index of dressing image. The ITW model consists of two main steps， the gray curve-fitting(GCF) threshold segmentation algorithm and Canny edge detection algorithm. In the ITW model， three types of wrinkle trends are defined. And the network dressing image is evaluated and simulated by three quantitative indexes： wrinkle number， wrinkle regularity and wrinkle unevenness. Finally， the fitness of three kinds of dress effects(tight， fit and loose) is quantified by objective fitting evaluation model.

Key words： objective fitting evaluation model； image to wrinkle(ITW)； dressing image； wrinkles index

Introduction

The contradiction between visual fit of dress image and actual dressing fit has become one of the main factors restricting the development of the garment electronic business. In the virtual environment of online shopping， consumers can’t try on the size of garments， can’t intuitively distinguish the color and luster of garment’s fabrics， and can’t touch the feel quality of fabrics[1]. The comfort of garments can only be judged by the business copy display and customer purchase feedback on purchases. They can only judge the fitness of garment through the feedback of the merchants and customers. Shaoetal.[2]ranked and quantified the details of perceptual clues according to weight value and importance degree based on the eye movement testing. They also analyzed the impact of color， layout， pictures， text and other elements on consumers’ online shopping experience. It confirmed that images are more important than texts in garments’ online shopping. In other words， in the online shopping platform of garment， consumers are more inclined to make purchase prediction by observing the dressing image. Furthermore， the research on the influence of image display[3]affects consumers’ impulsive purchasing intention， and also proves the importance of image in garments’ online shopping. Therefore， this paper takes images as the research object. In recent years， in the field of textile and garment， most of the research on images mainly focuses on two aspects：(1) the recognition of information from single image；(2) retrieval and classification of target images in the virtual system. Based on the influencing factors of garment fitting[4]， the former adopts image processing technology[5]to recognize and extract the quality， pattern， color， texture， style， modality and other information of garment and fabric in the image[6].Among them， fabric composition and properties are identified based on garment image[7-8]. Many scholars based on back propagation(BP) neural network， Fourier transform and Gabor transform algorithm recognize the surface texture and structure characteristics of fabrics[9]. At the same time， the identification of fabric defects was studied[10-12]. Javieretal.[13]and Leonardoetal.[14]objectively extracted and evaluated the wrinkles in different parts of garment images. Bossardetal.[15]proposed an algorithm to extract structural technical parameters such as chest circumference of garment based on standard working sketch and identify style image. Mengetal.[16]and Xingetal.[17]respectively analyzed the composition of three elements of color and extracted the main colors of national costumes based on algorithms.The latter is mostly based on the attitude estimation， scale-invariant feature transform(SIFT) features and other algorithms. The algorithm establishes feature labels to retrieve and classify garment images[18-24]， such as the garment color， garment length， sleeve length， dress posture and other feature labels. In conclusion， current researchs are mostly limited to image recognition and target image retrieval. Therefore， an objective fitting evaluation model(image to wrinkle, ITW) based on the wrinkle index of dressing image is proposed to solve the contradiction in online garment purchasing. Furthermore， it improves the accuracy of consumers’ perception of actual dress fitting based on dress images， and reduces the return rate.

1 ITW Model Framework Design

Wrinkles are not only a direct index of the fitness of dressing， but also the reflectin of the wearer’s physiological and psychological comfort states. Therefore， this research presents an objective fitting evaluation model of dressing fit in virtual environment， which is called ITW. And the proposed model takes the wrinkle as its object of study. The ITW model framework design process is shown in Fig. 1.

Fig. 1 ITW model framework

Step 1 Image acquisition： download the dressing image on the garment website.

Step 2 Image processing： the downloaded image is intercepted and processed by the software of Matlab， which can realize the function of segmentation of the target image and noise reduction of the original image. In this process， GCF threshold segmentation algorithm， Canny edge detection and gray scale curve are adopted.

Step 3 Index extraction and quantification： extract the area， trends， width and depth unevenness of the wrinkle are used to quantify the wrinkle number， wrinkle regularity and wrinkle unevenness， respectively.

2 Setting-Up ITW Evaluation Model

2.1 GCF threshold segmentation algorithm

GCF algorithm is an image processing algorithm proposed in this study. It is an image processing algorithm based on GCF. The purpose of the image processing is to reduce the noise of the original image， remove the noise and blur caused by illumination， scene and other factors， so as to improve the resolution and clarity of the image. Let f be the original color image， t be the gray image，t(x) be the gray histogram， g is the binary image， the background part isbw=1， the target part isbw=0， andxis the unknown gray value，T*is the optimal threshold. The GCF threshold segmentation algorithm is used to optimize the threshold. The flow diagram of the GCF threshold segmentation algorithm is shown in Fig.2. And the steps are as follows.

Fig. 2 Flow diagram of GCF threshold segmentation algorithm

Step 1 The background gray polynomial curveP1(x) and the target gray polynomial curveP2(x) are fitted respectively based on the background and target gray data.

Step 2 The probability of background pixel occurrence isP1， the probability of target pixel occurrence isP2，P1+P2=1.xis the gray value， and the image can be expressed as mixed probability density function by Eq.(1).

P(x)=P1P1(x)+P2P2(x).

(1)

Step 3 Searche for the thresholdx=T*， achieve the segmentation of the target and the background.

The error probabilityE1(T) of dividing the target pixels into the background is expressed by Eq.(2).

(2)

The error probabilityE2(T) of dividing the background pixels into the target is expressed by Eq.(3).

(3)

The total error probabilityE(T) is expressed by Eq.(4).

E(T)=P2E1(T)+P1E2(T).

(4)

(5)

WhenP1=P2is taken， the optimal thresholdT*of GCF threshold segmentation algorithm is obtained.

2.2 Canny edge detection

Edge detection is a feature detection method， including Canny， Sobel， Roberts， Log and other edge detection operators[25]. In this study， the edge detection method is used to obtain wrinkle contour of clear and complete， so as to define three trends of wrinkle. Therefore， only the effect of edge detection on target contour recognition is considered， and other differences such as the noise of four kinds of edge detection are not considered. This research intends to use Canny edge detection algorithm to recognize the wrinkle contour in dressing image.

2.3 Quantification of wrinkle indexes

(1) Number index

The number index is characterized by the area of wrinkle. It is expressed byWs. For the processed binary image， the area of the wrinkle is calculated by Eqs.(6) and(7).

s=∑n(i， j)，

(6)

(7)

where，n(i， j)=0， ∑n(i， j)represents the accumulation of black pixels， that is the target part， andNarepresents the area of the binary image g，Wsrepresents the area of wrinkle.

(2) Regularity index

According to Canny edge detection image， combined with three shirt dress images of tight， fit and loose， this study defines three trends of wrinkle to represent the regularity index， as shown in Fig. 3.

Fig. 3 Three wrinkle trends of the index of wrinkle regularity

1) PG

The force centre of the trend is in the chest point， shoulder point and other points of human body structure， as shown in Fig. 3(a). In the dress effect of tightening， most of wrinkles are PG.

2) PLG

The force centre of the trend is on the central line， and the wrinkles parallel to the central line are divergent， as shown in Fig. 3(b). On the edge of the surface of the human body， it is accumulated by the allowance of the fabric. In the dressing effect of looseness， most of wrinkles are PLG.

3) LG

The force centre of the trend is also the central line， but the wrinkle extends along the central line in all directions， as shown in Fig. 3(c). In the dressing effect of fitting， most of wrinkles are LG.

(3) Unevenness index

The wrinkle distribution of armhole， shoulder and chest is the main part of this study. In the intercepted image， the linear direction ofL1 andL2 is extended to recognize the gray change and generate the gray change curve.L1 represents 1/2 of the distance from the bottom of the armhole to the shoulder point.L2 represents the line in the bottom direction of the elongated armhole. The positions ofL1 andL2 in the structure diagram are shown in Fig. 4.Horizontal lines are drawn atL1 andL2， respectively. Calculate the depth unevenness(YVH) and width unevenness onL1 andL2 horizontal lines. The undulating variation of wrinkles is expressed by the depth unevenness， which reflects the lateral evenness of the wrinkles. The width unevenness is used to represent the space of wrinkles， which reflects the uniformity of wrinkles.

Fig. 4 Location marking of L1 and L2 in the garment prototype structure

Depth unevenness can be estimated by Eq.(8).

(8)

The width unevenness(YVD) can be estimated by Eq.(9).

(9)

3 Experiment and Simulation

3.1 Single image simulation

3.1.1ImageprocessingofGCFthresholdsegmentationalgorithm

Step 1 Download shirt image from application software(APP)and capture the left chest and shoulder parts of the image as the original image f， as shown in Fig. 5(a).

Step 2 Turn f to gray image t as shown in Fig. 5(b)， calculate the gray distribution of t， and get the histogramt(x). In addition， the contrast adjustment was made to the gray distribution oft(x) to make the image clearer， and the adjusted histogram was obtained， as shown in Fig. 5(c).

Step 3 The least square method was used to polynomial fit the bimodal curve in the gray histogram.

The GCF is shown in Fig. 6，and the curve expression is as follows.

(a) (b) (c) (d)

Fig. 6 Function curves of P1(x) and P2(x) in GCF threshold segmentation algorithm

(1) Background function can be estimated by Eq.(10).

P1(x)=

-0.203x2+52.64x-2 391.x∈[0， 255].

(10)

(2) Objective function can be estimated by Eq.(11).

P2(x)=

-0.308x2+121.8x-11 478.x∈[0， 255].

(11)

(3) Mixed probability density function. The probability of background pixel occurrence isP1， the probability of target pixel occurrence isP2，P1+P2=1.xrepresents the gray value， and the image can be expressed as mixed probability density function by Eq.(12).

P(x)=

P1P1(x)+P2P2(x)=-(0.203P1+0.308P2)x2+

(52.64P1+121.8P2)x-(2 391P1+11 478P2).

(12)

(4) Optimal thresholdT*

The error probabilityE1(T) of dividing the target pixels into the background is expressed by Eq.(13).

(13)

The error probabilityE2(T) of dividing the background pixels into the target is expressed by Eq.(14).

(14)

The total error probabilityE(T) can be calculated by Eq.(15)， andE(T) image is shown in Fig. 7.

(15)

The optimal thresholdT*is the value whenE(T) reaches the minimum valueE(T)min， that is to calculate

(16)

(17)

T*=181， that is the optimal threshold of GCF segmentation algorithm for this image is 181.

WhenT*=181, the binary image’s target and background are clearly segmented. Therefore, the optimal threshold value of this image is 181 by GCF threshold segmetation algorithm.

Fig. 7 First derivative image of E(T) function

3.1.2Imagecannyedgedetection

Canny edge detection combines the three criteria of signal-to-noise ratio， positioning accuracy and single edge response to obtain the optimal detection operator. From the results of edge detection of lady’s shirt image in Fig. 8， we can see that the Canny edge detection Fig. 8(a) can better achieve the purpose of clear and complete contour of wrinkle. Using Canny edge detection algorithm to recognize wrinkle contour in dressing image is suitable for this research.

(a) Canny (b) Log

(c) Roberts (d) Sobel

3.1.3Quantificationofwrinkleindexes

(1) Wrinkle number

The area of the wrinkle in Fig. 5(d) is calculated by Eqs.(6) and(7).

Ws=0.279,

(18)

whereWsrepresents the area of wrinkle. Therefore， according to Eq.(18)， the area of wrinkles accounts for 27.9% of the total area of the image.

(2) Wrinkle regularity

As shown in Fig. 9， the wrinkles in the red elliptical region are distributed in a line-gathering mode with the red solid line as the central line. This trend mainly focuses on the stress on the chest， shoulder and armhole to the central line. The yellow solid line(wrinkle line) is distributed in a point-gathering mode with the yellow dot as the centre， and the trend is mainly concentrated on the chest point and acromion point. The trend of the point is that the wrinkles in the blue box area are the blue lines with a parallel line. Among them， the regularity of parallel-line-gathering is the highest.

Fig. 9 Three wrinkle trends by Canny edge detection

(3) Wrinkle unevenness

The depth unevennessYVHand width unevennessYVDof the wrinkle onL1 andL2 are calculated. It can be seen from Fig. 10 that the depth unevenness ofL2 is less thanL1 and the width unevenness is greater thanL1.L2 passes through the chest circumference， and the chest is a smooth curved surface with almost no wrinkles. The gray curve is relatively smooth， so the depth unevenness is lower thanL1. The right side ofL2 passes through the bottom of the armhole. The wrinkles at the bottom of the armhole are larger and the gray value decreases obviously. Therefore， the width of the wrinkle changes greatly and the unevenness is greater thanL1.

Fig. 10 L1 and L2 unevenness of wrinkles(L1 curve： YVH=24.2%， YVD=11.5%； L2 curve： YVH=22.4%， YVD=12.7%)

3.2 Multi-image of different fitness recognition simulation

Based on the ITW objective evaluation model， simulation experiments were carried out with three different images： tight， fit and loose of dressing fit. And wrinkle indexes were extracted and quantified， and also the relationship between wrinkle indexes and dressing fit was established. Among them， the images of tight， fit and loose dress effects were recorded asT，F(xiàn)andLrespectively. The experimental results are shown in Figs. 11-12 and Table 1.

(a)

(b)

(c)

Fig. 11 Algorithmic simulation of three shirt dress images： (a) images in the tight state；(b) images in the fit state；(c) images in the loose state

(a1)

(a2)

(b1)

(b2)

(c1)

(c2)

Fig. 12 Gray-scale curve of three shirt dress images：(a1) and(a2) represent theL1andL2curves of gray image of the tight state， respectively；(b1) and(b2) represent theL1 andL2 curves of gray image of the fitting state， respectively; (c1) and(c2) represent theL1 andL2 curves of gray images of the loose state, respectively

The result of wrinkle number can be calculated by Formula(19).

FWs

(19)

It has been proved that the wrinkle area is the smallest when the dressing is a fitting state. When the dressing is a loose state， although the number of wrinkles is small， the fabric allowance accumulates on the side of the body， resulting in a larger area of wrinkles. When the dressing is a tight state， the wrinkle area is relatively large after accumulating several small areas of wrinkles. The result of wrinkle regularity can be defined by Formula(20).

Tr

(20)

whererrepresents the index of wrinkle regularity.Trrefers to the wrinkle unevenness when the dress is a tight.Frrefers to the wrinkle unevenness when the dress is fitting.Lrrefers to the wrinkle unevenness when the dress is loose. When the dressing is tight state， PG mode is the main trend of wrinkle， and the wrinkle regularity is the worst. When the dressing is a fitting state， the surface of the garments is smooth， and the wrinkles are mostly presented on the armhole by LG, and the wrinkles are more regularity. When the dressing is a loose state， the garment has margins. These margins are concentrated on the side of the garment and armhole. The wrinkles are mostly the PLG mode， and the wrinkle regularity is the highest.

Table 1 Wrinkle index comparison of tight， fit and loose dressing effects

The result of wrinkle unevenness can be calculated by Formulas(21) and (22).

TYVH

(21)

TYVD

(22)

It is proved that when the dressing effect is tight， the wrinkle unevenness of depth and width is the smallest. When the dressing effect is fitting， the garment presents the smooth surface. When the dressing effect is loose， the unevenness of wrinkles is the highest. At the same time， the wrinkle index along the bottom of the armhole changes more than the wrinkle index at 1/2 of the distance from the bottom of the armhole to the shoulder point.

Three shirt dressing images are processed by ITW objective fitting evaluation model， and the conclusion can be drawn：(1) the index of wrinkle number：FWs

4 Conclusions

This paper proposes the ITW objective fitting evaluation model of dressing fit. In the ITW objective fitting evaluation model， wrinkle information of dressing image is the subject of the study. Image processing technology is adopted， and GCF threshold segmentation algorithm and Canny edge detection algorithm are designed to extract the wrinkle area quantization wrinkle number index. Define three wrinkle trends to quantify the wrinkle regularity index, and the wrinkles unevenness is quantified by depth unevenness and width unevenness. Finally， the fitness of three kinds of dress effects(tight， fit and loose) are quantified by objective evaluation model. There are some shortcomings in this research, such as lack of physical verification. In the follow-up study， different samples will be designed through orthogonal experiments to verify the actual dressing fit. At present， this study can only compare the objective quantitative results of three dress states， and choose the clothes with the best fit in the comparison. Further quantification is needed in the following research， and the quantization range of the three dressing effects is given.