Extracting fabric hand information from visual representations of flared skirts

Abstract

In online transactions of textile products, fabric hand was thought to be inaccessible to consumers. Recently, much effort has been made to study the feasibility of providing consumers with a real sense of fabric through a virtual experience. The current paper proposes to extract fabric hand information from the perspective of visual perception. Two sensory experiments are conducted according to the standardized sensory evaluation procedures on a set of representative textile fabrics by two trained panels. The first experiment is aimed to measure how much fabric hand can be perceived through fabrics’ visual displays. On the basis of the positive results obtained, the second experiment is carried out to further investigate the interactive mechanism between samples’ visual features and their tactile properties. A novel algorithm based on rough set theory and fuzzy set theory is proposed in order to quantitatively measure relations between different sensory information.

Keywords

Fabric hand visual perception sensory analysis inclusion degree

Fabric hand (i.e., fabric tactile properties) is an important aspect of fabric assessment that is related to the comfort, style and appearance of an apparel item. In the textile industry, new materials with different touch feelings are produced to satisfy the ever-growing market demands.

Generally, fabric hand evaluation can be carried out in two ways: (1) objective evaluation and (2) sensory evaluation. The objective evaluation is performed by measuring a set of mechanical and physical parameters on fabrics. For this purpose, many fabric characterization devices and systems have been developed, such as the ‘Kawabata Evaluation System for fabrics’ (KES-FB)¹ and the ‘Fabric Assurance by Simple Testing’ (FAST),² which are, to date, most commonly practiced, and a more recent measurement system, FAMOUS, which applies similar principles.³ The most attractive strength of objective evaluation is that it can lead to precise and repeatable numerical data. However, its main weakness lies in that it is, in fact, an indirect measurement of fabric hand. Thus its interpretation with respect to real human feeling requires further exploitation. In this situation, as another branch of study on fabric hand, sensory evaluation has been, recently, found to be more important since it directly characterizes the human sense of touch on fabrics. Compared with FOM (Fabric Objective Measurement), sensory evaluation has a high capacity in dealing with human linguistic data, and producing interpretable results. Practically, a sensory experiment is carried out by a panel of trained subjects according to normalized evaluation procedures.^4,5 For example, sensory descriptors are selected through standardized methods; sensory evaluation techniques are scientifically designed and unified among the assessors; the implementation of sensory evaluation is under supervision; and so on. It was stated by Fritz that, ‘people are capable of making objective, quantitative, and repeatable assessments of their sensations’. The sensory evaluation can be considered as an objectified subjective assessment, and has been extensively applied in fabric hand studies in recent years.^6–9

In real applications, since objective measurements can provide more direct and practical guidance for textile production, a great amount of research work has been found in the modeling of relations between objective and sensory evaluations of fabric hand using classical or intelligent methods.^10–16 Many successful attempts are bringing this branch of research to a certain kind of maturity. However, the available mechanical studies are based on real fabric measurements. They seem to lack solutions when fabrics are presented in a virtual environment, such as the Internet.

Nowadays, e-shopping has emerged and is becoming a generally accepted purchasing mode due to its economical and convenient features. With an ever larger market being accessible to manufacturers and retailers, e-commerce is steadily growing, which makes the interaction between consumers and business expand and flourish to an unforeseen height. In practice, as one of the most life-related goods, apparel products enjoy an indispensable large share in the overall number of online transactions. According to Internet Retailer, 34% of British consumers shopped online for clothes in 2010, up from 25% in 2009. In addition, the UK’s online apparel sales were estimated to grow 60% by 2015.¹⁷

Although e-shopping is becoming popular, there still exists consumers’ strong need to fully ‘feel’ textile items before making a final purchasing decision. One big barrier making the current online apparel shopping still far from satisfactory is the products’ intangibility during the transaction. The virtual fitting room is a recent approach to resolve consumers’ demand to make choices from various clothes styles and then examine the physical fitness of the preferred items to their own body shapes.^18,19 Despite that this fitting system can simulate with some success people’s try-on behavior as it is in a real store, fabric hand properties of clothes still remain as a mist in this virtualized environment.

Recently, much research work has been dedicated to exploring the possibility of providing a user with a completely reliable sense of fabric through a virtual experience. One remarkable progress is the development of haptic force feedback devices, such as the Cyberglove, the PHANTom and, more recently, the HAPTEX system,²⁰ etc. These inventions are aimed to make an interaction between users and virtual fabrics using simulation techniques. However, until now, few of them have the sensitivity required for an accurate simulation of fabric hand at the research level, not to mention reaching the market. One reason, as was mentioned previously, is that the attempts made by these devices are based on objective measurements of physical properties on real fabrics.^1,21–23 But so far, there is not a commonly accepted FOM that is capable of truly and comprehensively quantifying people’s complicated sense of touch.^24–26

In fact, according to the previous description about sensory evaluation and the definition put forward by the Textile Institute in 1975 that fabric hand is ‘the subjective assessment of a textile material obtained from the sense of touch’,²⁷ there are enough reasons to believe that the most reliable and effective way to study fabric tactile properties might be from the perspective of human natural perception. It is especially the case when a textile product is to be evaluated in a virtual experience.

Apart from fabric hand evaluation, visual evaluation is another important aspect of textile sensory study. A large number of papers have been found in the aesthetic and behavioral studies on the color and the structural appearance of textile products.^28–30 Some researches deal with the quantification of garment visual information. For example, Kawabata et al.³¹ developed a model to predict the quality of the appearance of men’s suit from fabric mechanical properties; more recently, Gersak³² and Pavlinic and Gersak³³ have developed an intelligent system to study the correlations between fabric mechanical parameters and the grade of garment appearance quality. However, in all these fields, few studies have tried to explore fabric tactile properties from perceived visual information.

Since visual information is the main accessible medium to consumers during online transactions, in the current study we are considering the possibility to study fabric hand from the point of view of our visual perception. The multisensory relation in our study is proved to exist based on the so-called memory association mechanism according to many studies in cognitive psychology.^34–37 Briefly speaking, cumulated touching experience accompanied by visual perceptions would make the already-familiar or similar visual information a stimulus next time in a non-haptic environment so as to recall associated tactile memory.

In the present study, discussions were carried out around two sensory experiments on a set of flared skirts made of different textile fabrics. In Experiment 1, a trained panel was organized to evaluate fabric hand of the samples in three independent scenarios, that is, real-touch, video and image scenarios. The aim of this experiment is to measure the extent to which tactile information could be interpreted through fabric’s visual representations. On the basis of Experiment 1, a second experiment (Experiment 2) was conducted using another group of panelists to describe the visual features of the skirts through video displays. This part is aimed to investigate in depth the mechanism of how the samples’ external visual features have an impact on the fabrics’ internal tactile properties in video representations.

A novel method based on the so-called ‘fuzzy inclusion degree’ is proposed to quantitatively study the relations between different sensory information. Unlike classical correlation analysis-based methods, this method computes the information inclusion of tactile properties in fabric visual representation by combining the use of rough sets theory³⁸ and the fuzzy inference system.³⁹ Rough sets have high capacity in the analysis of data structure, while the fuzzy inference system⁴⁰ is good at processing uncertain information, such as human expert reasoning, to produce interpretable results. In addition, the proposed algorithm can effectively detect the non-linear structure lying beneath the sensory data, while being safe to include a comparatively small number of samples.

The results obtained from this study will make contributions to at least two aspects as follows. Firstly, it can be a significant step forward on the visual interpretation of fabric tactile properties. The positive results will prove the feasibility of visual representations being a reliable medium through which fabric hand information can be perceived during distant transactions. Secondly, upon the current work, an interactive mechanism between fabrics’ visual features and tactile properties will be established. Fabric hand will, for the first time, become manipulable through a virtual experience.

Experiments

The current study is based on two sensory experiments, Experiment 1 and Experiment 2. Experiment 1 is aimed to verify the feasibility of interpreting fabric tactile properties through samples’ visual representations. With the positive answer obtained from Experiment 1, Experiment 2 will be carried out to further investigate the interpretative mechanism of visual features to fabric tactile properties.

Experimental preparation

Sample skirts

Eighteen textile fabrics with various tactile properties were selected and made into two-piece flared skirts as our experimental samples. The material and constructional details of these 18 fabrics are shown in Table 1.

Table 1.

Fabric details of the 18 samples

Sample	Fabric Content	Weave structure	Areal density (g/m²)
S1	100% Cotton	Twill	406.6
S2	55% Cotton, 45% linen	Plain	201.9
S3	50% Polyester, 50% Acrylic	Plain	120.1
S4	100% Polyester	Twill	108.3
S5	100% Polyester	Twill	224.6
S6	100% Silk	Satin	66.30
S7	100% Cotton	Twill	421.3
S8	55% Ramie, 45% Cotton	Plain	189.5
S9	100% Polyester	Plain	31.70
S10	100% Cotton	Twill	410.3
S11	97% Cotton, 3% Spandex	Plain	209.6
S12	100% Polyester	Plain	373.2
S13	65% Linen, 35% Cotton	Plain	199.4
S14	100% Polyester	Satin	110.1
S15	100% Silk	Satin	68.90
S16	85% Cotton, 15% Polyester	Plain	452.1
S17	100% Cotton	Plain	157.8
S18	100% Polyester	Plain	30.20

In the current study, the simple style, two-piece flared skirt was chosen for two reasons. Firstly, simply styled apparel products without many structural variations can display more clearly the tactile properties of the corresponding textile material. Secondly, simply styled samples are easier for accurate reproducing, which is especially important when fabric difference should be kept as the only variable between different samples.

All these skirts are of the same design and production parameters. The specifications and basic pattern of each skirt are shown in Table 2 and Figure 1, respectively.

Figure 1.

Basic pattern of one piece of the skirt.

Table 2.

Specifications of the flared skirt

Specification	cm
Waist	68
Length	60

Creation of visual representations

A girl, whose body size is best fit into the sample skirt, was invited as the mannequin in our experiments. Two types of visual materials were created for each sample skirt: a series of static photos and a dynamic video clip.

Creation of image materials

In this part, the 18 skirts were put on the mannequin one after another. A DSLR (digital single lens reflex) camera whose maximum resolution is 5616 × 3744 pixels was used to take photos of each skirt from eight different directions, according to the angle contained between the mannequin’s front and the lens, 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315°. Taking sample 1 as an example, Figure 2 shows its image representation consisting of eight photos from standardized angles.

Creation of video materials

Figure 2.

Image representation of sample 1.

A camcorder whose maximum resolution is 1920 × 1080 pixels was used to make video clips for all the samples. In each clip, the mannequin was required to walk to-and-fro (with pauses) and make various postures according to a predefined standard that is aimed to show comprehensively both the static and dynamic effects of the skirts. Before the real recording, a posture-training session was taken. Figure 3 shows a screenshot during the video recording of sample 1.

Figure 3.

Video representation of sample 1.

Sensory terms

Two experiments were carried out in the current study. One is about the fabric hand evaluation of the sample skirts in three different sensory scenarios. The other is aimed to characterize samples’ external features through visual representations. As a premise, the determination of evaluation criteria is crucial in sensory experiments. Here, a standardized procedure was carried out to generate descriptive terminologies corresponding to the two experiments.

Fabric hand descriptors

‘Brainstorming’

A so-called ‘brainstorming’^41,42 was launched among some professionals from the textile industry to produce an exhaustive list of linguistic descriptors to depict fabric tactile properties from a general point of view. During the discussion, experts were free to speak out words that came to their mind when they thought about the tactile properties of textile products in our daily life. There was no restriction on the form of the words. They can be adjectives, nouns or even verbs. After the discussion, about 220 words were collected.

Screening

In this step, the words that had been collected in the ‘brainstorming’ session were screened for the first time through another discussion among the same panel of experts. During this discussion, experts were asked to remove those words that express hedonic preferences such as ‘pleasant’, ‘uncomfortable’, etc. The words that would easily lead to confusion in understanding were also removed from the list. In addition, those words that cannot be used to describe the specific samples in our experiment were eliminated as well. Finally, adjective words were determined to be more preferable in the study, and the words in other forms that were regarded as important were turned into adjectives by discussion. After this session, over 50 words remained on the list.

Literature study

A final screening of the descriptive words was done on the basis of the information gathered from the literature.^43,44 Among the over 50 words obtained from the previous session, those words that were delivered in a less normalized way were removed or replaced by the ones from the literature. For example, the descriptor ‘weak’ was replaced by ‘flimsy’. Finally, 22 pairs of descriptors were determined as the tactile evaluation criteria in our study, which are listed in Table 3. For a better understanding, the 22 tactile descriptors have been classified by experts into three major categories, mechanical, surface and constructional, respectively. This classification is based on the semantic understanding of the experts on fabric hand. Some of them are close to mechanical measures but not the same. For example, to assess the descriptor pair ‘Stiff–pliable’ (D1), one is supposed to hold the fabric between the fingers and feel the resistance by folding it forward and back, which is similar to a mechanical bending test. Thus, D1 is reasonably classified into ‘Bending’ under the category of ‘Mechanical’. What is worth mentioning is that due to the complexity of human senses and the richness of our language, the selected linguistic descriptors are not uniquely defined according to the above classification. Some of them can cover several property aspects. For example, the assessment of the descriptor pair ‘Flimsy–firm’ (D8) is related to a fabric’s weave density and resistance to deformation. According to this definition, D8 should fall into ‘Bending’, ‘Tensile and shearing’ under the category ‘Mechanical’, and at the same time the category ‘Constructional’. The classification (and sub-classification if any) of the selected 22 descriptors is shown in Table 4.

Table 3.

Twenty-two bipolar descriptors

Nm	Descriptor pair	Nm	Descriptor pair	Nm	Descriptor pair
n
D1	Stiff–pliable	D9	Thin–thick	D17	Bumpy–non-bumpy
D2	Crisp–limp	D10	Light–heavy	D18	Prickly–non-prickly
D3	Dead–lively	D11	Soft–hard(in compression)	D19	Fuzzy–non-fuzzy
D4	Draped–non-draped	D12	Non-springy–springy	D20	Non-slippery–slippery
D5	Crumply–wrinkle-resistant	D13	Non-full–full	D21	Harsh–soft
D6	Non-stretchy–stretchy	D14	Rough–smooth(overall feeling)	D22	Warm–cool
D7	Loose–tight	D15	Grainy–non-grainy
D8	Flimsy–firm	D16	With ridges–without ridges

Table 4.

Classification of tactile descriptors

Category	Mechanical			Surface	Constructional
Category	Bending	Tensile and shearing	Compression	Surface	Constructional
Tactile descriptors	D1, D2, D3, D4, D5, D8	D3, D4, D5, D6, D7, D8	D11, D12, D13	D14–D22	D7, D8, D9, D10, D13, D22

For each descriptor pair, a detailed explanation to both its definition and the corresponding assessing gestures was determined by referring to the literatures^7,46 and, in particular, by carrying out a discussion among the experts. Initial tests were performed to decide whether the evaluation techniques were understood and could easily be applied by the panelists. One example is shown as follows.

Stiff–pliable:

Definition: A fabric that is ‘stiff’ is not easily bent, and is rigid and inflexible. ‘Pliable’ expresses the opposite meaning of ‘stiff’.

Gestures: To assess this attribute, the panelist holds the fabric between the thumb and the other four fingers of his/her most used hand; while moving the fabric back and forth, he/she assesses the resistance.

Visual feature descriptors

The aim of this part is to produce an exhaustive list of descriptors in order to cover as comprehensively as possible the visual features of the concerned skirts. These descriptors are supposed to be directly captured through vision and express the premier and basic information about the samples’ external features. To be specific, they should be composed of two aspects. One is the appearance characteristics about the skirts, while the other is the visual characteristics about the fabrics.

Skirts appearance characteristics

To generate descriptors of the first aspect, a procedure similar to the one used for the generation of fabric hand descriptors ((1) Fabric hand descriptors in the section Sensory terms) was designed and implemented. Some redundant descriptors were screened out after several discussions. Meanwhile, a literature study was necessary to make sure that the selected descriptors are generally in consensus with the commonly accepted terminology. In this way, 28 descriptors concerning both the static and dynamic effects of the sample skirts are determined (from E1 to E28 in Table 5).

Visible fabric characteristics

Table 5.

Visual features of sample skirts

Position	Nm	Feature		Position	Nm	Feature
Waist line	E1	Outline of pleats	Clear–Fuzzy		E20	Drape	Badly–well
Waist line	E2	Outline of pleats	Rigid–soft		E21	Overall outline	Rigid–soft
Abdomen and hip	E3	Fitness to body shape	Unfit–fit	Dynamic	E22	Balance	Badly–well
	E4	Pleat size	Small–big		E23	Following capacity	Badly–well
	E5	Pleat distribution	Uneven–even		E24	Clinging	Badly–well
	E6	Appearance	Unnatural–natural		E25	Ethereality	Badly–well
Lower part of skirt	E7	Expending extent from below the hip	Not expanding–expanding		E26	Wave flowability	Badly–well
	E8	Wave size	shallow–deep		E27	Swinging range	Small–big
	E9	Wave smoothness	Non-smooth–smooth		E28	Rhythm	Badly–well
	E10	Wave distribution	uneven–even	Fabric	E29	Rough–smooth(overall feeling)
	E11	Wave-height consistency	Inconsistent–consistent		E30	Grainy–non-grainy
	E12	Bottom-edge evenness	Uneven–even		E31	With ridges–without ridges
	E13	Curling extent of bottom edge	Non-curling–curling		E32	Bumpy–non-bumpy
Color	E14	Brightness	dark–light		E33	Prickly–non-prickly
	E15	Vividness	Muddy–vivid		E34	Fuzzy–non-fuzzy
	E16	Temperature	Cool–warm		E35	Non-slippery–slippery
luster	E17	Light intensity	Weak–strong		E36	Harsh–soft
Silhouette	E18	Fitness	Unfit–fit		E37	Thin–thick

As the other aspect of the skirts’ visual features, the descriptors about fabrics’ visual characteristics are decided as follows.

In fact, the selected 22 tactile descriptors can be divided into two categories, visible and invisible, according to their direct accessibility by vision. In the category of visible, we firstly, include eight descriptors concerning fabric surface properties, that is, from D14 to D21. The different surface textures perceived by the panelists during touch originate from the fabrics’ geometric characteristics, such as natural convolution of fibers, cross-sectional shapes of fiber, twists of yarns, surface fluff and so on. Actually, these characteristics can also be perceived by our eyes. For different fabric samples, incident light beams are scattered at different strengths in different directions according to the specific surface geometry. When a panelist views a fabric sample, the diffusely reflected light would stimulate the panelist’s eyes and provide two-dimensional color images on the panelist’s retinas as a description of the sample’s surface characteristics, at which point the panelist registers a three-dimensional image by way of recognizing memories of experiences with fabrics. This is a typical process of the so-called memory association we mentioned in the previous section. In this sense, we call these eight descriptors visible. In addition, there are another two properties that can also be considered as visible. One is the mechanical property ‘Draped–non-draped’ (D4); the other is the constructional property ‘Thin–thick’ (D9). Similarly, these two properties are directly measurable via vision. The evaluation of the drape of a fabric sample is highly related to the silhouette of the skirt it is made into. The thickness of a fabric can easily be judged through direct visual observation.

The remaining 12 tactile descriptors concerning fabric mechanical properties (D1, D2, D3, D5, D6, D7, D8, D11, D12 and D13), certain constructional and surface properties (D10 and D22) fall into the category of invisible due to the fact that they cannot be directly measured by visual observation.

For the purpose of this part, along with the 28 skirt appearance descriptors, the 10 visible tactile descriptors are involved as another aspect into the visual features about the samples. Since the descriptor ‘Draped–non-draped’ is included in both visual aspects, 37 visual descriptors are finally determined in this study. As is shown in Table 5, these descriptors are categorized according to their descriptive positions. For each descriptor, a detailed definition with a graphic illustration (shown in Figure 4) was available to the panelists.

Figure 4.

Graphic illustration of some parts of the skirt.

Evaluation scale

An 11-point scale numbered from 0 to 10 was applied to the sensory evaluations in both Experiment 1 and Experiment 2. In order to prevent mis-scoring, every point on the scale was well defined, as is shown in Figure 5. Taking the fabric hand evaluation as an example, for any descriptor pair (e.g., Stiff–pliable), according to the scale, the score ‘5’ signifies ‘medium’ in sensory intensity, and as it decreases to ‘0’ and increases to ‘10’, the sense is intensified to the two poles, ‘extremely stiff’ and ‘extremely pliable’. For example, if a sample was considered to be very stiff, then its score on the ‘Stiff–pliable’ scale should be 1; if it was considered very pliable, then the score should be 9, and so forth.

Figure 5.

Evaluation scale and semantic explanation.

Sensory experiments

Experiment 1 (fabric hand evaluation)

Panel

In this part, the fabric hand evaluations were carried out by experts with a textile background. A panel of 30 female and 12 male members between 23 and 55 years were recruited. They are either university professors (including lecturers, researchers or research assistants from two textile colleges), professionals in the textile industry (working mainly on fabric design and fashion design), PhD students or postgraduate students (working or studying in textile schools).

All of these panelists have previously participated in at least two subjective tests concerning the assessment of textiles’ tactile properties. In order to prevent cross-scenario effects during the experiments, the 42 panelists were randomly put into three groups to assess the fabric hand independently in real-touch, video and image scenarios, respectively.

Training

A training session was organized for the panelists in all the three scenarios. This is a real-touch training, the aim of which is to strengthen the panelists’ evaluation-related knowledge. In this session, some training samples (different from the 18 experimental samples) were used to help the panelists become familiar with all the descriptors, including gestures as well as the evaluation scales. This session took about 6 hours.

Procedure

Real-touch scenario

Before the experiments, all the samples were conditioned for a minimum of 24 hours under the standard atmospheric condition (20 ± 2℃ temperature, 65 ± 2% relative humidity). Fourteen panelists took part in the real-touch scenario. They were allowed to both see and touch the fabric during the evaluation, which is in accordance with our real-life experience. Before getting started, the panelist would be asked to wash and dry his/her hands with the non-moisturizing soap and paper towel provided. The evaluation should be carried out individually for each panelist.

Video scenario

Another 14 panelists participated in the video assessment scenario, where the video clips of the sample skirts were displayed one by one on a computer screen. During the evaluation, the panelists were free to control the playback of the video clips and make pauses wherever they needed. The panelists were also required to conduct the tests individually.

Image scenario

The remaining 14 panelists took part in this scenario, in which the multi-angle images of the 18 samples were shown one by one on a computer screen. During the process, the panelists were free to control the display of the images by either changing the displaying order or zooming in/out the photo. The evaluations were carried out individually for each panelist.

Experiment 2 (visual feature evaluation)

On the basis of Experiment 1, a second experiment was launched to further investigate the mechanism of interaction between samples’ visual features and the corresponding fabric hand through video displays.

Panel

Five professionals were recruited from the apparel industry to evaluate skirts’ visual features (or visual features). All these panelists have profound experience in evaluating the appearance of apparel products according to standard criteria.

Training

A training session was organized to help the panelists become familiar with the descriptors as well as the scoring method to be concerned in the evaluation.

Procedure

In the present experiment, the video clips of the sample skirts were displayed the same as they were in the fabric hand evaluation. During the evaluation, the panelists were free to control the playback of the video clips and make pauses wherever they needed.

Mathematical methodology

Mathematically, the aim of the current research is to study the relations between different sensory data sets. Many statistical methods are available for this purpose, such as linear regressions⁴⁶ and factorial analysis (principal component analysis (PCA), canonical correlation analysis (CCA), etc.).^47,48 These methods are efficient in solving many problems in sensory evaluation due to their good capacity of studying linear patterns of different information and then discovering correlations therein from a large base of data. However, when the above classical tools are applied to problems dealing with human knowledge, many drawbacks would come into being: (1) linear statistical techniques are not capable of treating non-linear or complex relations between data sets; (2) most of these methods are not good at describing uncertain and imprecise linguistic data and numerical results obtained from the classical methods cannot, sometimes, lead to significant physical interpretation; (3) there are always strict constraints on the size and distribution of the database during the application of classical analysis methods.

Therefore, in the current study, a novel algorithm is proposed to measure the relations between different sensory information. This algorithm is based on a fuzzy inclusion defined according to rough sets theory and fuzzy techniques. Compared with the existing methods, it is capable of treating relations between variables or attributes from a small number of learning data.^49,50 Our approach is constructed upon two major indices, the Classification Consistency (CCons*) based on the concept of fuzzy inclusion degree, and the Ranking Consistency (RCons) obtained from the non-parametric ranking coefficient (Kendall’s τ). Then, in order to generate a criterion to measure the general consistency (denoted by GCons) between different data sets in the condition that this criterion should be both robust to noise and easy for interpretation, a fuzzy inference system is developed to integrate the previous two indices, CCons* and RCons.

The general framework of our approach is illustrated in Figure 6. The following discussion consists of three parts. The first two parts illustrate the respective principles of these two indices. The third part describes the development of a fuzzy inference system in order to generate the General Consistency (GCons) as a fusion measure of the previous two indices.

Figure 6.

General framework for our approach.

Classification consistency (*CCons**)

In this part, the ‘Classification consistency (CCons*)’ between different sensory data sets (i.e., visual and real-touch perceptions of fabric tactile properties) is formulated. Firstly, let us have some basic ideas about the rough set philosophy.

Problematic

Rough set theory was proposed by Pawlak in the 1980s.⁵¹ It is a relatively new soft computing tool for analyzing a vague description of an object. A rough-set-based data analysis starts from a data table, called an information system. The information system contains data about objects of interest, characterized by a finite set of attributes. It turns into a decision table when its condition attributes and decision attributes are distinguished.

The formalization of the present problem is given below. Let U = {e₁, e₂,…, e₆} be the set of samples. The corresponding evaluation scores for any specific pair of tactile descriptors, such as ‘Stiff–pliable’ have been obtained from the visual (either image or video) and real-touch perceptions on all the samples, denoted as C = (c(e₁)… c(e₆)) ^T and D = (d(e₁)… d(e₆)) ^T , respectively. All the evaluation scores c(e_i) and d(e_i) vary between 0 and 10. In the following discussion, the visual perception C is taken as condition variable and the real-touch perception D as a decision variable. According to rough sets philosophy, the knowledge acquisition is in fact a process of knowledge classification. Different knowledge would generate different partitions of data. From the previous two vectors of C and D, we obtain two partitions for the visual and real-touch results, that is, U/C = {X₀, X₁,…, X₁₀} and U/D = {Y₀, Y₁,…, Y₁₀}. $X_{i} \in U / C (i = 0, 1, \dots, 10)$ (or $Y_{j} \in U / D (j = 0, 1, \dots, 10)$ ) is the class of samples in which the evaluation scores are all i (or all j). In practice, some X_i (or Y_j) can be empty if its index i (or j) does not exist in the evaluation scores of all the samples e_k values.

Thus, the aim of this part of the algorithm is to compute the extent to which the partition of the condition set (visual perception) is consistent with that of the decision set (real-touch perception).

Classical classification consistency (CCons)

Inclusion degree (Inc)

Let

S = (U, C \cup D)

be a complete decision table, U the collection of experimental samples,

X \in U / C

an equivalence class representing the visual perceptions of the samples on a specific tactile property and

U / D = {[e] D : e \in U}

representing the corresponding real-touch perceptions of the samples. For any sample

e k \in U

, the inclusion degree of X_i with respect to

[e j] D

is denoted by

\begin{matrix} Inc ([e k] D / X i) = \frac{| X i \cap [e k] D |}{| X i |}, \\ (i = 0, 1, \dots, 10; k = 1, 2, \dots, 6) \end{matrix}

(1)

where

0 \leq Inc ([e k] D / X i) \leq 1

$[e k] D$ is the set of samples that are classified into the same group with e_k according to the decision attribute.

In fact, this formulation of inclusion degree is in agreement with that of the rough membership function of e_k in X_i, that is, $δ X i (e k) = \frac{| X i \cap [e k] D |}{| X i |}$ ,⁵² defined according to rough sets theory. It is evident that, if $Inc ([e k] D / X i) = 1$ , then X_i can be said to be consistent with respect to $[e k] D$ , or one has $X i \subseteq [e k] D$ (which is called a complete inclusion).

Construction of classical classification consistency

Based on the above inclusion degree, we can define the classification consistency (CCons) of the visual evaluation results (obtained from video or image scenarios) with respect to the real-touch ones on any fabric tactile property. Mathematically, it expresses the percentage of objects that can be correctly classified into decision classes of

U / D

by the condition attribute set C.

Let $S = (U, C \cup D)$ be a complete decision table, $U = {e 1, e 2, \dots, e 6}$ the set of samples, $U / C = {X 0, X 1, \dots, X 10}$ and $U / D = {[e] D : e \in U}$ the equivalence classes of the condition attribute and the decision attribute, respectively. Notably, the aim of the current study is to analyze, under different experimental conditions, the panelists’ perceptual difference on the same fabric tactile property. As a consequence, the condition attribute is determined as the same as the decision attribute. Here, a consistency measure of an equivalence class X_i of the condition part $U / C$ with respect to the decision part $U / D$ is defined as

\begin{matrix} Cons (X i, D) = 1 - \frac{4}{| U |} \sum_{k = 1}^{| U |} Inc ([e k] D / X i) (1 - Inc ([e k] D / X i)), \\ (i = 0, 1, \dots, 10; k = 1, 2, \dots, 6) \end{matrix}

(2)

where

0 \leq Cons (X i, D) \leq 1

Inc ([e k] D / X i)

is the inclusion degree of X_i into

[e k] D

for sample e_k.

Then, on the above basis, the classification consistency measure (CCons) of C with respect to D is defined as

\begin{matrix} CCons (C, D) = \sum_{i = 0}^{10} \frac{| X i |}{| U |} (1 - \frac{4}{| U |} \sum_{k = 1}^{| U |} Inc ([e k] D / X i) \\ \times (1 - Inc ([e k] D / X i))), (k = 1, 2, \dots, 6) \end{matrix}

(3)

where

Inc ([e k] D / X i)

is the inclusion degree of X_i into

[e k] D

for sample e_k.

Modified classification consistency (*CCons**)

Fuzzy inclusion degree (FInc)

In practice, an inclusion degree based on the previous crisp partition of samples might lead to serious information loss. For any specific class, it does not make a distinction between samples close to it and those far from it. For any sample e_k not belonging to a class X_i, we have its membership value equaling zero, indicating that its adhesion to this equivalence class is regarded as null. Evidently, in the definition of the inclusion degree, it is more reasonable to consider the degrees of adhesion of the samples to the equivalence classes so that samples close to a class are more important than those far from it. According to this idea, we modify the previous inclusion degree using the concept of fuzzy partition.³⁹

The following illustrates the modified inclusion degree $FInc ([e k] D / X i)$ :

FInc ([e k] D / X i) = \sum_{l = 1}^{6} min (μ X i (e l), μ [e k] D (e l)) / \sum_{l = 1}^{6} μ X i (e l)

(4)

in which, X_i denotes the ith condition set and

[e k] D \subset U / D = {Y 0, Y 1, \dots, Y 1 0}

is the decision set where the sample e_k belongs. The classes X_i and Y_j constitute two fuzzy partitions (sets) for the attributes C and D and they are characterized by the fuzzy membership functions

μ X i (e)

and

μ Y j (e)

, defined as

μ X i (e) = 1 - h | i - c (e) |

(5)

μ Y j (e) = 1 - h | j - d (e) |

(6)

They are triangular functions centered on i and j, respectively. h is the coefficient controlling the sensitivity of these functions. In the current study, we assign 0.2 to h as a general case.

Notably, (1) the decision set where a sample e_k might belong is determined according to the maximum membership principle, that is, e_k is believed to belong to a decision set Y_j when the corresponding fuzzy membership degree $μ_{Y_{j}} (e_{k})$ reaches the highest among all the decision sets; (2) given the present problem, any $μ_{X_{i}} (e)$ or $μ_{Y_{j}} (e)$ whose value is negative would be taken as zero for further computation. Figure 7 depicts the membership function of the descriptor pair ‘Stiff–pliable’ as an example.

Construction of modified classification consistency

Figure 7.

Membership function of ‘Stiff– pliable’.

On this basis, the classification consistency of X_i with respect to D has been modified as

\begin{matrix} Cons * (X i, D) = 1 - \frac{4}{\sum_{i = 0}^{10} \sum_{k = 1}^{6} μ X i (e k)} \sum_{k = 1}^{6} FInc ([e k] D / X i) \\ \times (1 - FInc ([e k] D / X i)) \end{matrix}

(7)

where

FInc ([e k] D / X i)

is the fuzzy inclusion degree of X_i into

[e k] D

for sample e_k. Similar to the consistency degree defined in Equation (2), we have

0 < Cons * (X i, D) < 1

. This is due to the following facts: (1) for each sample e_k, there always exists a Y_j whose value j is the closest to d(e_k) (the difference is smaller than 0.5), that is,

μ Y j (e k) \geq 0.9

; (2)

0 \leq FInc ([e k] D / X i) (1 - FInc ([e k] D / X i)) \leq \frac{1}{4}

holds for any X_i and Y_j.

Hitherto, the final classification consistency of the condition attribute C with respect to the decision attribute D can be constituted as

CCons * (C, D) = \sum_{i = 0}^{10} \frac{\sum_{k = 1}^{6} μ X i (e k)}{\sum_{i = 0}^{10} \sum_{k = 1}^{6} μ X i (e k)} Cons * (X i, D)

(8)

where

FInc ([e k] D / X i)

is the fuzzy inclusion degree of X_i into

[e k] D

for sample e_k.

Ranking consistency (RCons)

According to the algorithm introduced above, the fuzzy inclusion degree I(C, D) quantifies the extent to which the partition of the condition set (or, the visual result set) is consistent with that of the decision set (or the real-touch result set). However, the ordinal consistency between the sets is not well taken into consideration. Actually, in an information system, the element’s different positions in the two sets may lead to large differences between the knowledge to be represented. Therefore, it is significant to include another index to measure the ordinal difference between the data obtained from different sensory modalities. For this purpose, Kendall’s rank coefficient is introduced in our study.⁵²

As a non-parametric measure of rank correlation, Kendall’s τ depends upon the number of inversions of pairs of objects that would be needed to transform one rank order into the other. Still taking the first experiment as an example, for the sample set $U = {e 1, e 2, \dots, e 18}$ , C = (c(e₁) … c(e₆)) ^T and D = (d(e₁) … d(e₆)) ^T are observations corresponding to the visual evaluation (i.e., for the video or image scenarios) and the real-touch evaluation (i.e., for the real scenario), respectively. Kendall’s rank correlation coefficient τ or RCons(C, D) is computed from

\begin{matrix} RCons (C, D) = τ (C, D) = 1 - \frac{2 d Δ (C, D)}{n (n - 1)} \\ - 1 \leq τ (C, D) \leq 1 \end{matrix}

(9)

where n is the number of samples.

d Δ (C, D)

denotes the symmetric difference between two attribute sets C and D, and it is obtained from the following formula:

\begin{matrix} d Δ (C, D) = (number of concordant pairs) \\ - (number of discordant pairs) \end{matrix}

(10)

Notably, any pair of observations $(c (e_{i}), d (e_{i}))$ and $(c (e_{j}), d (e_{j}))$ are considered to be concordant if the ranks for both elements agree: that is, if both $c (e_{i}) > c (e_{j})$ and $d (e_{i}) > d (e_{j})$ or if both $c (e_{i}) < c (e_{j})$ and $d (e_{i}) < d (e_{j})$ . On the other hand, they are considered to be discordant if $c (e_{i}) > c (e_{j})$ and $d (e_{i}) < d (e_{j})$ or if $c (e_{i}) < c (e_{j})$ and $d (e_{i}) > d (e_{j})$ . However, if $c (e_{i}) = c (e_{j})$ or $d (e_{i}) = d (e_{j})$ , then the pair is neither concordant not discordant.

General consistency measure

In this study, an aggregation criterion (AC) was proposed to integrate the previous two indices, CCons* and RCons, so as to constitute a general consistency measure (GCons) to investigate the overall correlation between the two sensory modalities. This criterion should be both robust to noise and easy for knowledge interpretation. For this purpose, a fuzzy inference system⁵³ is designed and illustrated in Figure 8. This system consists of three major parts responsible for the fuzzification of input data, fuzzy operation based on fuzzy rules and defuzzification to produce output data.

Figure 8.

Aggregation criterion constituted based on a fuzzy inference system.

Fuzzy rules are crucial for building a fuzzy inference system. In the current study, discussions were carried out among a panel of six experts to design a series of fuzzy rules. VS, S, M, L, VL denote the linguistic values of ‘Very small’, ‘Small’, ‘Medium’, ‘Large’ and ‘Very Large’, respectively. For example, a fuzzy rule can be described as the following linguistic expression:

IF CCons* is medium (M) and RCons is small (S), THEN GCons is small (S).

Each linguistic expression is equivalent to a numerical value ranging from 0 to 1 according to the fuzzy membership function defined for each input and output variable during the process of fuzzy inference. The membership function is illustrated in Figure 9. It is a commonly used function uniformly distributed on [0, 1].

Figure 9.

Fuzzy membership function for the input/output variable.

After applying the AC, a general consistency measure (GCons) was constituted to investigate the extent to which the tactile properties can be transmitted through different visual representations of a textile product, with both the classification consistency and the distribution similarity taken into consideration.

Results and discussion

Results from Experiment 1

GCons results are computed and illustrated as the so-called ‘Perceptual lines’ for both video and image observations in Figure 10. The solid line represents the video performance, while the dashed line represents the image performance. Out of the overall 22 attributes, 19 (over 86%) have their GCons values higher than 0.6 and 14 (about 64%) higher than 0.7 in both video and image scenarios, which indicates that an important part of the tactile properties can be well interpreted through some specific visual representations of the apparel products.

Figure 10.

GCons values on tactile descriptors.

It can be observed from the line shapes that on most of the descriptors, the shape of the solid line is more stable than the dashed line with fewer fluctuations, which indicates that the panelists in the video scenarios can perform more stably and with higher accuracy as compared with those in the image scenarios. This is consistent with the statistical results, shown in Table 6, that the video observations have a higher mean value and lower STDEV (standard deviation) value than the image observations.

Table 6.

Statistical results of visual observations

Index	Video	Image
MEAN	0.784	0.712
STDEV	0.135	0.157

Let us take a further look at the perceptual lines in Figure 10. For the left part of the figure, in which the descriptors concern the fabric’s mechanical and constructional properties, both the solid and dashed lines have generally more and larger fluctuations than for the descriptors on the right part of the figure, which implies that the panelists in both the video and image scenarios tended to encounter greater difficulties in perceiving these properties. This observation is reasonable, since in real-life experience, most of the properties falling in this group are evaluated through direct touch by hand such as stretching, grasping, bending, etc. However, when touch is deprived, as is the case in our visual experiments, the assessors are supposed to, in fact, make decisions based on the associated memory accumulated from previous touching experience. Thus, their judgment might be less accurate or even incorrect. For example, on the sixth descriptor pair ‘Stretchy–non-stretchy’ (D6), an extremely low value is detected in both video and image scenarios, which indicates that both the available video and image displays fail to well recall the panelists’ associated memory on this specific property.

What is worth mentioning is that, by comparing the shape of the two perceptual lines, we still can find that the panelists in video scenarios performed better than those in image scenarios on almost all the properties of this part. This observation is not hard to understand. Compared with static photos captured from a limited number of angles, video clips can record information about an object from every possible angle and in a continuous way just as the object is viewed in the real experience. As we have discussed previously, the more visual information about the skirts is available, the more the so-called associated memory, which plays a crucial role in the non-haptic evaluation of fabric tactile properties, will be recalled by our brain. In this situation, it is not surprising that the panelists tend to make more correct judgments on most of the samples’ mechanical and constructional properties through video clips than through static images.

On the other hand, for the descriptors concerning fabric surface characteristics (from D14 to D22), whose results are depicted on the right part of Figure 10, both perceptual lines (for video and image scenarios) are comparatively stable in shape and are situated above the level of 0.8. All these indicate that although the panelists could not really touch the fabric, they are still able to perceive most of its surface properties with certainty, which is in accordance with our daily life experience. Fabric surface characteristics originate from their weave, yarn thickness, yarn density and so on. Actually, these parameters are visible via diffusely reflected light that comes from (1) the surface layer of fibers and (2) between surfaces of internal fibers. As we view a real textile sample, the light reflected diffusely stimulates our eyes and provides two-dimensional color images on our retinas as an image of the woven construction, at which point our brain registers a three-dimensional image by way of recognizing memories of experiences with fabrics, which is a typical kind of memory association. According to our experimental results, in the current study, the panelists were not able to view the real samples, but the corresponding visual (video or image) representations are believed to have made the reflected light of importance be truly recorded.

Results about Experiment 2

Two major conclusions can be drawn from the results of Experiment 1. Firstly, most of the tactile properties can be transmitted through the fabric’s visual representations. Secondly and significantly, the panelists in video scenarios have generally better performance than those in image scenarios, which indicates that in a non-haptic environment, visual displays with dynamic effects can provide the most possible information about a fabric’s tactile properties.

On the above basis, the current experiment is aimed to discover the embedded relation between the 37 visual features and the 12 invisible tactile properties through samples’ video displays.

The method used in the first experiment is applied to the present experimental data. According to our proposed algorithm, let $U = {e_{1}, e_{2}, \dots, e_{18}}$ be the collection of samples. The results about the visible features and the invisible tactile properties from evaluating samples’ video representations are considered as the condition set and the decision set, respectively.

For each invisible tactile descriptor (D#), a set of GCons values are computed on all the 37 visual features. Notably, those visual features that have higher GCons values are believed to be important in interpreting the corresponding fabric tactile property. The results are ranked in descending order and shown as so-called ‘Impact lines’. Figure 11 shows an example of the descriptor pair ‘Crisp–limp’ (D2).

Figure 11.

Impact line for Crisp–limp (D2).

For each pair of tactile descriptors, there are initially 37 visual features that have impacts, to different extents, on revealing the corresponding tactile property. In this part of the research, for each tactile property, a stepwise screening has been taken to reasonably find out from such a large number of visual features the few ones that are of the highest significance. The first step is to remove the visual features that have an obvious low impact. Those features remaining are called ‘Relevant visual features’ (denoted as RFs in the following discussion). On this basis, the next step is to further select the visual features that are believed to have the closest relationship with the tactile property of interest. The features defined in this step are then called ‘Major visual features’ (denoted as MFs).

Selection of RFs

As was mentioned previously, for each tactile property, the ‘Impact line’ depicts the degrees of relevancy of the 37 visual features in descending order. To select the RFs for a specific tactile property is in fact to find out the visual features that have higher GCons values.

For each tactile property, the highest GCons value computed from the 37 visual features is called its Impact level or IL for short. According to the distribution of the GCons results obtained from the 12 tactile properties, it is determined that, for a specific tactile property Dj, the visual feature Ei can be taken as a RF when the following condition holds:

\begin{matrix} GCons (i) \geq IL (j) \times 70 % \\ (i = 1, 2, \dots, n; j = 1, 2, 3, 5, 6, 7, 8, 10, 11, 12, 13, 22) \end{matrix}

(11)

in which, n is the number of visual features (n = 37).

That is to say, any visual feature whose GCons value is higher than 70% (or has a decrease of less than 30%) of the impact level is regarded as a RF of the corresponding tactile property. Taking D2 as an example, its impact level is the GCons value of E20, 0.848. Then the visual features whose GCons values are higher than 0.594 (0.848 × 70%), or graphically, on the impact line shown in Figure 11, the visual features before (and including) E12 (GCons (12) = 0.643, and for the visual feature E11 that is after E12 on the impact line, GCons (11) = 0.531) are taken as the corresponding relevant visual features, or RFs. The RFs for D2 (i.e., from D20 to D12 on the Impact line) have been marked in Figure 11 by the horizontal parenthesis.

In this manner, a group of RFs is selected for each of the 12 tactile properties. For any visual feature, the number of tactile properties that regard it as the RF are counted and this value is called its Power. Figure 12 shows the 37 visual features ranked according to their strength of power in descending order. It is evident from this figure that almost every visual feature has a considerable impact on a range of tactile properties, which indicates that the 37 visual features are properly selected for the current study.

Figure 12.

Power of 37 visual features.

According to the classification in Table 4, in the following discussion, the 12 invisible descriptors are categorized into two groups. One is called ‘Mechanical’ which contains eight descriptors concerning fabric bending, compression, tensile and shearing properties (D1 (Stiffness), D2 (Crispness), D3 (Liveliness), D5 (Wrinkle resistance), D6 (Stretchiness), D7 (Tightness), D11 (Compressive softness) and D12 (Springiness)). Then, the other group is called ‘Constructional’, which includes the remaining four descriptors, that is, D8 (Firmness), D10 (Weight), D13 (Fullness) and D22 (Surface temperature).

Discussion on mechanical properties

Selection of Major visual features

Actually, for each tactile descriptor, there are several visual features, which have still higher GCons values among the RFs constituting the so-called ‘Major-impact list’. These visual features are called MFs. Here, we give the criterion for selecting the MFs as follows.

For a specific tactile property, let G = (g₁, g₂,…, g_t) ^T be the vector of GCons values of the RFs in a descending order. We define the decreasing rate dr_i of any RF from the previous one on the sequence as

{dr}_{i} = \frac{(g_{i - 1} - g_{i})}{g_{i - 1}}, (i = 2, 3, \dots, t)

(12)

where t is the number of RFs for the tactile property.

Thus, the average decreasing rate of the sequence is formulated as

\bar{dr} = \frac{1}{t} \sum_{i = 2}^{t} {dr}_{i}

(13)

On the RF sequence, we find the ith RF whose decreasing rate dr_i is the first to exceed the average $\bar{dr}$ . Then, we consider the RFs before the ith RF (not including it) as the MFs of the corresponding tactile property.

As an example, for D2, the corresponding average decreasing rate $\bar{dr}$ is 0.014. E8 is the first visual feature on the RF sequence whose decreasing rate dr₇ (=0.023) exceeds 0.014. Consequently, we select the first six visual features from the RF sequence, respectively E20, E21, E25, E7, E3 and E26, as the MFs of D2, which are marked by square brackets above the impact line in Figure 11.

According to this criterion, for the eight mechanical properties, there are overall 16MFs. For each MF, the number of mechanical properties under its power has been counted. The results for all the MFs are ranked in a descending order and shown in Figure 13.

Classification of Major visual features

Figure 13.

Power of Major visual features.

Actually, according to the specific definition of each MF (seen in Table 5), these 16MFs can be categorized into three major classes. The first class, which includes four MFs (E3, E7, E20 and E21) expressing skirts’ static outline features, is named ‘Macro-static features’ (MaSs). The second class contains seven MFs (E1, E2, E6, E8, E10, E11 and E13) concerning skirts’ detailed information, such as the size, shape and distribution of the pleats and waves. This class is called ‘Micro-static features’ (MiSs). The remaining five MFs (E23, E24, E25, E26 and E28) constitute the third class, ‘Dynamic features’ (Dyns).

As is obvious from Figure 13, among the ranked 16MFs, the first four are MaS features referring to the fitness at the hip (E3), the skirt’s expanding degree (E7), drape (E20) and outline shape (E21). As an individual, each feature has an impact on six out of eight fabric mechanical properties, which is called ‘impact-coverage’. Working together, they have an overall impact-coverage of 87.5% (seven out of eight). In addition, the MaS features have comparatively larger impact-coverage than the other two classes of features. From this, we can assume that skirts’ shaping effects are crucial elements among all the MFs affecting fabric mechanical properties.

According to the previous classification criteria, there are eight Dyn features in the overall 37 visual features, five among which appear as the MFs and have an impact on five out of eight fabric mechanical properties. This indicates that skirts’ dynamic effects are significant for reflecting fabric mechanical properties. Among these elements, the ethereality (E25) and wave flowability (E26), individually, are proved to be especially important, having impact-coverage of 62.5% and 37.5%, respectively.

Although, as individuals, the remaining seven MiS features have less impact-coverage than the MaS features, with the largest being 50% (E8, wave size), their total impact-coverage still reaches 62.5% (six of eight), which is the same as the Dyn features. In addition, the MiS features have an impact on fabric wrinkle resistance (D5), which both the MaS and Dyn features can hardly explain.

Visual interpretation mechanism

To be specific, Table 7 shows how these 16MFs have an impact on each fabric mechanical property. The impact levels for the eight mechanical properties are ranked in descending order and shown in Figure 14.

Table 7.

Major features for each fabric mechanical property

D1 (Stiff–pliable)	E21, E3, E20, E7	E24, E25, E26
D2 (Crisp–limp)	E20, E21, E7, E3	E25, E26
D3 (Dead–lively)	E20, E3, E7, E21	E23, E25	E8
D5 (Wrinkle resistance)	E11, E10
D6 (Stretchiness)	E20, E21, E7, E3	E8, E10, E2, E6, E13
D7 (Loose–tight)	E7, E21	E23, E28, E25, E26
D11 (Compressional softness)	E20, E3, E21, E7	E25	E8, E6, E2, E1
D12 (Compressional springiness)	E3,E20	E8, E2

Figure 14.

Impact levels for mechanical properties.

Some observations are obtained from Table 7 and Figure 14.

Above all, the first three mechanical descriptors (D1, D2 and D3) concerning fabric bending properties are mainly affected by the MaS and Dyn features. Their impact levels are highest among all the eight mechanical properties. Moreover, as is observed from Experiment 1, fabric bending properties can be better perceived in video scenarios than in image scenarios. These indicate that samples’ dynamic effects can add much to panelists’ evaluation accuracy on bending properties.

Then, for the descriptors concerning fabric tensile and shearing properties (D5, D6 and D7), the skirts’ dynamic features do not have strong relation with D5 (wrinkle resistance) and D6 (stretchiness). Although D7 (tightness) is affected by Dyn features, its impact level is very low. Together with the results obtained from Experiment 1 that the GCons values on these three descriptors are relatively low in both image and video scenarios, we can assume that the dynamic features have little impact on delivering fabric tensile and shearing properties.

Finally, we consider the compression properties, D11 (compressional softness) and D12 (compressional springiness), the skirts’ MiS and MaS features, such as the rigidness of pleat outlines (E2) and wave size (E8), fitness to body size (E3), drape (E20), and so on. In fact, as different from mechanical measurements, the daily deformation of our clothes is poly-directional. For example, when the skirt is draped, the main force comes from the vertical gravity, but the deformation of the fabric caused by this force might be bending, tensile, shearing and even compressive according to different stress-bearing positions. In this sense, it is not hard to understand that the pleats of a skirt are mainly caused by bending and compressive deformation. Therefore, it is reasonable to believe that fabric compression properties can be perceived to some extent by watching the static effects of the samples.

It seems that the samples’ dynamic features do not have much contribution to expressing fabric compression properties. However, according to the results in Experiment 1, although the GCons values of these two descriptors in video scenarios are not that high as compared with those of the bending descriptors, still they are much higher than those in image scenarios. From this interesting observation, we are assuming that although the specific dynamic features are not that capable of expressing some of the fabric mechanical properties, the dynamic display itself is a naturally better way of representation as compared with the static one. That is to say, more mechanical information of a fabric can be discovered through dynamic displays, which give a consecutive, comprehensive and vivid report about skirts’ all-around features. Another similar example is found in the observation about D5 (wrinkle resistance), where the GCons value in video scenarios is much higher than that in image scenarios, even though D5 is not affected by the Dyn features but instead the MiS features.

Discussion on constructional properties

Figure 15 shows the impact lines for the four constructional properties, ‘Flimsy–firm’ (D8), ‘Light–heavy’ (D10), ‘Non-full–full’ (D13) and ‘Warm–cool’ (D22). According to the method mentioned in the previous part (Equations (12) and (13)), the MFs for each property are selected and marked in brackets in Figure 15. It is obvious that these four properties are predominantly affected by the visual features concerning fabric surface attributes and thickness. To be specific, fabric firmness (D8) is overwhelmingly affected by their thickness (E37). Fabric weight (D10) is also largely correlated to their thickness, while the surface attributes, such as E35 (slipperiness), E32 (bumpiness) and E29 (overall roughness), exist as second important elements. This observation is somewhat puzzling, but there is still some logic behind it. According to our daily experience, fabrics with rough surfaces tend to give us an impression of being tightly woven with thick fibers. This can be a good proof of the so-called memory association. With regards to fabric fullness (D13), the most significant impact comes from their fuzziness (E34) on the surface, while thickness (E37) still plays an important role therein, which is in accordance with the original definition of the fullness. Finally, the descriptor concerning fabric temperature on the surface (D22) is reasonably connected to fabric thickness and surface properties, such as slipperiness (E35), overall roughness (E29) and fuzziness (E34). For example, a thick fabric with a fuzzy surface such as corduroy tends to give a warm feeling as it is touched by hand, while a thin satin with smooth surface would always produce a cool and refreshing hand.

Figure 15.

Impact lines for constructional properties (D8, D10, D13 and D22).

Discussion on fabric color and luster

Among the 37 visual features, there are four descriptors concerning fabric color (E14, E15 and E16) and luster (E17). It is observed from this experiment (Figures 12 and 15) that none of these four descriptors are taken as RF for any tactile property. Although E14 (color brightness) has a few impacts on perceiving fabric weight (D10), its strength of impact is relatively low as compared with other major visual features for D10. Besides, for E15 (color vividness), E16 (color temperature) and E17 (luster), their connection with any tactile property is found to be very weak. Therefore, it is assumed that color and luster have little influence on panelists’ perception of fabric tactile properties.

General conclusion

Today, the Internet has become a compelling channel for sales of textile products. To provide people with the most close-to-real sense of fabrics through a virtual experience is a strong wish of textile researchers. Since up to 80% of the daily information is perceived by our eyes, vision is believed to be the most widely used and reliable channel for information acquisition.⁵⁵ Moreover, compared with touch, vision can be obstacle-free in a virtual environment, for instance, in an online transaction. On these grounds, we propose to study fabric tactile properties from the angle of visual perception.

There are two major problems put forward in this study. The first is how much tactile information can be transmitted through textiles’ visual displays. If there exists the so-called compensatory effect from vision to touch, there comes the second problem of how this effect works. Two sensory experiments were implemented to answer these two problems. A novel algorithm based on the so-called ‘fuzzy inclusion degree’ was proposed to study the correlation between different sensory information.

The results obtained from the first experiment confirm that a large part of fabric tactile properties can be interpreted through some specifically designed visual displays (in either video or image forms). Comparatively, in the current study, through (and only through) the visual representations, the panelists can well perceive the samples’ surface properties while tending to encounter more difficulties in perceiving some mechanical properties whose evaluation predominantly depends on the direct handling of fabrics in our real-life experience (e.g., the stretchiness). Generally speaking, video displays can provide more comprehensive and accurate information about fabric tactile properties than image displays by being able to recall more memory association, which is established between the visual and haptic perceptions often happening simultaneously during our daily contact with fabrics.

On this basis, the second experiment was carried out to explore the interactive mechanism between samples’ visual features and their fabric hand through video displays. The results show that the visual features that have a major impact on fabric mechanical properties can be categorized into three classes concerning the samples’ shaping, detail and dynamic features. For example, the bending properties are mainly related to the shaping and dynamic features, while the compression properties are more affected by skirts’ detail characteristics. Although, for some mechanical properties, it appears that the dynamic features do not have a direct impact, the observation in this study testifies that dynamic displays are superior to static displays in transmitting fabric mechanical properties due to their better capacity in illustrating, for example, samples’ visual features in detail. On the other hand, fabric constructional properties are mainly affected by samples’ surface characteristics and thickness. Memory association is found in the perception of fabric weight, where moderate correspondence is detected from some surface features. This can be a proof of the existence of cooperation and compensation between multisensory information during the visual perception of fabric tactile properties in the present study. Finally, as another remarkable observation, color and luster are found to have little impact on panelists’ perception about samples’ tactile properties. This would help produce more straightforward results by largely reducing the system’s noise and meanwhile making the sensory evaluations more operable by simplifying the experimental design.

The mathematical framework proposed in the current research shows efficiency in exploring the interactive mechanism between fabric tactile properties and the corresponding visual representations. It has been generalized to be useful for a wide range of applications as long as the study of relations between different data sets is concerned. Moreover, due to its high capacity in dealing with data imprecision and uncertainty, this framework will be found particularly competent in solving sensory-related problems.

Footnotes

Funding

This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

References

Kawabata

. The standardization and analysis of hand evaluation, 2nd ed. Osaka: The Textile Machinery of Japan, 1980.

Minazio PG. Fast–fabric assurance by simple testing. International Journal of Clothing Science and Technology, 1995; 7(2/3): 43–48.

Stylios

. New measurement technologies for textiles and clothing. Int J Clothing Sci Technol 2005; 17: 135–149.

ISO 8586:1993. Assessors for sensory analysis, Part 1: Guide to the selection, training and monitoring of selected assessors, International Standard ISO 8586–1993 (F).

ISO 11035:1995. Sensory analysis - identification and selection of descriptors for establishing a sensory profile by a multidimensional approach, International Standard ISO 11035–1995 (F).

Bishop

. Fabric sensory and mechanical properties. Textil Prog 1996; 26: 5–27.

Philippe

Schacher

Adolphe

. Tactile feeling: Sensory analysis applied to textile goods. Textil Res J 2004; 74: 1066–1072.

Chollakup

Sinoimeri

Philippe

. Tactile feeling: sensory analysis applied to textile goods. Textil Res J 2004; 74: 1066–1072.

El-Ghezal

Babay

Sahnoun

. The tactile sensory evaluation of knitted fabrics: effect of some finishing treatments. J Sens Stud 2010; 25: 201–215.

10.

Vhen

Newton

. A psychophysical model for objective fabric hand evaluation: an application of Steven's law. J Textil Inst 1993; 84: 354–363.

11.

Park

Hwang

Kang

. Applying fuzzy logic and neural networks to total hand evaluation of knitted fabrics. Textil Res J 2000; 70: 675–681.

12.

El-Ghezal

Babay

Sahnoun

. Sensory and instrumental techniques evaluating the effect of structure parameters on the tactile properties of knitted fabrics. J Texture Stud 2010; 41: 714–735.

13.

El-Ghezal

Sahnoun

Babay

. Predicting compression and surfaces properties of knits using fuzzy logic and neural networks techniques. Int J Clothing Sci Technol 2011; 23: 294–309.

14.

Zeng

Koehl

. Representation of the subjective evaluation of fabric hand using fuzzy techniques. Int J Intell Syst 2003; 18: 355–363.

15.

Zeng

Koehl

Sanoun

. Integration of human knowledge and measured data for optimization of fabric hand. Int J Gen Syst 2004; 33: 243–258.

16.

Xue Z and Chen Y. Sensory evaluation of men’s suitings based on a fuzzy-genetic model. In: IEEE 10th international conference on computer-aided industrial design & conceptual design (CAID & CD 2009), Wenzhou, China, 26–29 November 2005, 2009, pp.2123–2128.

17.

Internet retailer, Online marketing trends, report, http://www.onlinemarketing-trends.com/2011/04/uk-online-apparel-sales-to-grow-60-by.html (2011). (accessed 9 April 2013).

18.

Protopsaltou D, Luible C, Arevalo M, et al. A body and garment creation method for an internet-based virtual fitting room. In: Vince J and Earnshaw R (eds) Advances in modeling, animation and rendering. Springer: London, 2002, pp.105–122.

19.

Pereira

Silva

Alves

. Virtual fitting room augmented reality techniques for e-commerce. Commun Comput Inf Sci 2011; 220: 62–71.

20.

Magnenat-Thalmann

Volino

Bonanni

. From physics-based simulation to the touching of textiles: the HAPTEX project. Int J Virtual Reality 2010; 6: 35–44.

21.

Kawabata S, Niwa M and Fumei W. Objective hand measurement of non-woven fabrics. I: Development of the equations. Textil Res J 1994; 64(10): 597–610.

22.

Sülar

Okur

. Objective evaluation of fabric handle by simple measurement methods. Textil Res J 2012; 78: 856–868.

23.

Gersak J. Mechanical and physical properties of textile materials / Mehanske in fizikalne Lastnosti Tekstilnih Materialov. Maribor: University of Maribor, Faculty of Mechanical Engineering, Department for textile material and design, 2006 (in Slovene).

24.

Kohko

Kawakami

. Study of texture appearance using computer graphics. J Illum Eng Inst Jpn 1995; 79: 16–23.

25.

Cherkassky

Weinberg

. Objective evaluation of textile fabric appearance Part 1: Basic principles, protrusion detection, and parameterization. Textil Res J 2010; 80: 226–235.

26.

Cherkassky

Weinberg

. Objective evaluation of textile fabric appearance Part 2: SET opti-grade tester, grading algorithms, and testing. Textil Res J 2010; 80: 135–144.

27.

Ali

Begum

. Fabric softeners and softness perception. Ergonomics 1994; 37: 801–806.

28.

Johnson

Downing

. Deindividuation and valence of cues: Effects on prosocial and antisocial behavior. J Pers Soc Psychol 1979; 37: 1532–1538.

29.

Soda

Murakami

. Relationship between preference of clothes color by college students and reason for color selection. J Jpn Res Assoc Textil End Uses 2008; 49: 881–888.

30.

Wang Y and Chen Y. The Kansei research on the style of women’s overcoats. In: IMACS multiconference on computational engineering in systems application, 2006, pp.431–436.

31.

Kawabata

Niwa

Yamashita

. Recent developments in the evaluation technology of fiber and textiles: Toward the engineered design of textile performance. J Appl Polym Sci 2002; 83: 687–702.

32.

Gersak

. Development of the system for qualitative prediction of garments appearance quality. Int J Clothing Sci Technol 2002; 14: 169–180.

33.

Pavlinic

Gersak

. Predicting garment appearance quality. Open Textil J 2009; 34(2): 29–38.

34.

Klatzky

Lederman

. Multisensory texture perception. Multisensory Object Perception in the Primate Brain, New York: Springer, 2010. Chapter 12.

35.

James

Kim

. Dorsal and ventral cortical pathways for visuo-haptic shape integration revealed using fMRI. Multisensory object perception in the primate brain, New York: Springer, 2010. Chapter 13.

36.

Bernard W and Juan CA. ‘Cognitive’ memory. In: proceedings of international joint conference on neural networks, Montreal, Canada, 31 July–4 August 2005, pp.3296–3299.

37.

Newell

. Visuo-haptic perception of objects and scenes. Multisensory object perception in the primate brain, New York: Springer, 2010. Chapter 14.

38.

Pawlak

. Rough set theory and its applications in data analysis. Cybern Syst 1998; 29: 661–688.

39.

Dubois

Prade

Yager

. Fuzzy information engineering: A guide tour of applications, New York: Wiley, 1996.

40.

Takagi

Sugeno

. Fuzzy identification of systems and its applications to modeling and control. IEEE Trans Syst Man Cybern 1985; 15: 116–132.

41.

Osborn

. Applied imagination, Rev. ed. New York: Scribner, 1957.

42.

Osborn AF. Applied imagination: principles and procedures of creative problem solving. 3rd rev. ed. Scribner: New York, 1963.

43.

Civille

Dus

. Development of terminology to describe the handfeel properties of paper and fabrics. J Sens Stud 1990; 5: 19–32.

44.

Guidelines for the subjective evaluation of fabric hand. AATCC technical manual, 2007, pp.382–385.

45.

Meilgaard

Civille

Carr

. Sensory evaluation techniques, Boca Raton, FL: CRC Press, 1991.

46.

Weisgerg

. Applied linear regression, 3rd ed. New York: John Wiley & Sons, 2005.

47.

Jolliffe

. Principal component analysis, 2nd ed. New York: Springer, Inc, 2002.

48.

Wolfgang

Léopold

. Applied multivariate statistical analysis, 2nd ed. Berlin, Heidelberg: Springer Press, 2007, pp. 233–272.

49.

Xue Z, Zeng X, Koehl L, et al. Study on relation between visual and tactile perception of textile products using a fuzzy inclusion measure. In: Proceedings of 9th Pangborn sensory science symposium, Toronto, Canada. 2011.

50.

Xue Z, Zeng X and Koehl L. Study on interactive mechanism between visual features and tactile properties of textile products. In: proceedings of international conference on Kansei engineering and emotion research, Penghu, Taiwan, 22–25 May 2012.

51.

Pawlak

. Rough sets. Int J Comput Inf Sci 1982; 11: 341–356.

52.

Siegel S. Nonparametric statistics. Topsever Y., trans. Ankara: Ankara University Publishing House, 1977.

53.

Wang

Ruan

Kerre

. Mathematics of fuzziness-basic issues, Berlin: Springer, 2009. Chapter 6.

54.

Wang

Mao

. Computer image processing and recognition technology, Beijing: Higher Education Press, 2001.

Extracting fabric hand information from visual representations of flared skirts

Abstract

Keywords

Experiments

Experimental preparation

Sample skirts

Creation of visual representations

Sensory terms

Evaluation scale

Sensory experiments

Experiment 1 (fabric hand evaluation)

Experiment 2 (visual feature evaluation)

Mathematical methodology

Classification consistency (CCons*)

Problematic

Classical classification consistency (CCons)

Modified classification consistency (CCons*)

Ranking consistency (RCons)

General consistency measure

Results and discussion

Results from Experiment 1

Results about Experiment 2

Selection of RFs

Discussion on mechanical properties

Discussion on constructional properties

Discussion on fabric color and luster

General conclusion

Footnotes

Funding

References

Classification consistency (*CCons**)

Modified classification consistency (*CCons**)