Digital Stethoscope Design

Satoshi Takahashi1[0000−0002−1067−6704], Keiko

Yamaguchi2[0000−0002−1129−8012], and Asuka Watanabe3[0000−0002−3539−8253] 3 Kyoritsu Women’s Junior College, 2-6-1 Hitotsubashi, Chiyoda, Tokyo, Japan Abstract. Fashion styles adopted every day are an important aspect of culture, and style trend analysis helps provide a deeper understanding of our societies and cultures. To analyze everyday fashion trends from the humanities perspective, we need a digital archive that includes images of what people wore in their daily lives over an extended period. In fashion research, building digital fashion image archives has attracted significant attention. However, the existing archives are not suitable for retrieving everyday fashion trends. In addition, to interpret how the trends emerge, we need non-fashion data sources relevant to why and how people choose fashion. In this study, we created a new fashion image archive called Chronicle Archive of Tokyo Street Fashion (CAT STREET) based on a review of the limitations in the existing digital fashion archives. CAT STREET includes images showing the clothing people wore in their daily lives during the period 1970–2017, which contain timestamps and street location annotations. We applied machine learning to CAT STREET and found two types of fashion trend patterns. Then, we demonstrated how magazine archives help us interpret how trend patterns emerge.

Figure: System Model & Simulation Flow for Digital Stethoscope Design

These empirical analyses show our approach’s potential to discover new perspectives to promote an understanding of our societies and cultures through fashion embedded in consumers’ daily lives.

Figure: System Model & Simulation Flow for Digital Stethoscope Design

Keywords: Fashion trend · Digital fashion archive · Image processing · Machine learning · Deep learning.

Introduction

Fashion styles adopted in our daily lives are an important aspect of culture. As noted by Lancioni , fashion is ‘a reflection of a society’s goals and aspira- tions’; people choose fashion styles within their embedded social contexts. Hence, the analysis of everyday fashion trends can provide an in-depth understanding of our societies and cultures. In fashion research, the advantages of digital fash- ion archives have continued to attract attention in recent times. These digital ⋆This work was supported by the Japan Society for the Promotion of Science (JSPS) KAKENHI [grant number 20K02153].

2

S. Takahashi et al. archives have been mainly built based on two purposes. One is enhancing the scholarly values of museum collections by making them public to utilize them digitally. Another motivation is predicting what fashion items will come into style in the short term with images retrieved from the internet for business purposes.

However, these archives are not suitable for analyzing everyday fashion trends from the humanities perspective. To identify everyday fashion trends, we need a new digital archive that includes images of the clothes people wore in their daily lives over an extended period. In addition, we need non-fashion data sources showcasing why and how people choose fashion to interpret the trends.

In this study, we create a new fashion image archive called Chronicle Archive of Tokyo Street Fashion (CAT STREET) showcasing fashion images reflect- ing everyday women’s clothing over a long period from 1970 to 2017 in Tokyo.

CAT STREET helps overcome the limitations in the existing digital fashion archives by applying machine learning to identify fashion trend patterns. Then, we demonstrate how magazine archives are suitable in understanding how various trend patterns emerged. The empirical analyses show our approach’s potential in identifying new perspectives through fashion trends in daily life, which can promote understanding of societies and cultures.

Fashion Trend Analysis

Kroeber analyzed fashion trends manually and measured features of women’s full evening toilette, e.g. skirt length and skirt width, which were collected from magazines from 1844 to 1919. After this work, many researchers conducted simi- lar studies . Their work surveyed historical magazines, portraits, and illustrations and focused mainly on formal fashion, rather than clothing worn in everyday life.

The critical issue in the traditional approach is that analyzing fashion trends is labor-intensive. Kroeber is among the representative works in the early stage of quantitative analysis of fashion trends. This type of quantitative analysis requires a considerable amount of human resources to select appropriate images, classify images, and measure features. This labor-intensive approach has long been applied in this field. Furthermore, manually managing large modern fashion image archives is cumbersome. To solve this issue, we utilize machine learning in this study. Machine learning is a computer algorithm that learns procedures based on sample data, e.g. human task results, and imitates the procedures.

In recent years, machine learning has contributed to the development of digital humanities information processing . In the field of fashion, modern fashion styles are complex and diverse. Applying machine learning to fashion image archives may be expected to be of benefit in quantifying fashion trends more precisely and efficiently.

A Novel Approach to Analyze Fashion Digital Archive from Humanities

Digital Fashion Archives

In recent years, digital fashion archives have been proactively built in the fields of museology and computer science. However, the construction of these archives was prompted by different research objectives. Many museums and research in- stitutions have digitized their fashion collections not only to conserve these historically valuable collections but also to enhance the schol- arly values by making them available to the public. Vogue digitized their mag- azines from 1892 to date, the Japanese Fashion Archive collected representative Japanese fashion items from each era, and the Boston Museum of Fine Arts archived historical fashion items from all over the world.

On the other hand, in computer science, the motivation to build digital fashion archives is generally short-sighted and primarily to predict the next fashion trends that can be used by vendors to plan production or by online stores to improve recommendation engines. Several studies have created their own fashion image databases to fit the fashion business issues they focused on . Table 1 presents the best-known public databases from fashion studies in computer science.

Unfortunately, the fashion databases proposed in previous studies have sev- eral limitations in terms of the approach to everyday fashion trends. First, the level of detail of location annotation in existing databases is insufficient to an- alyze everyday fashion trends. People belong to a social community, and some social communities have their own distinct fashion styles and their territory. For instance, ‘Shibuya Fashion’ is a fashion style for young ladies that originated from a famous fashion mall in Shibuya . Young ladies dressed in ‘Shibuya Fashion’ frequent Shibuya, one of the most famous fashion-conscious streets in Japan. Hence, we need fashion image data with location annotations at the street level to focus on what people wear in their daily lives.

Second, most databases consist of recent fashion images from the last decade because they were obtained from the internet. The periods covered by these databases might not be sufficiently long to determine fashion trends over longer periods. By examining how fashion changes over extended periods, sociologists and anthropologists have found that fashion has decadal-to-centennial trends and cyclic patterns . An example is the hemline index, which is a well-known hypothesis that describes the cyclic pattern in which skirt lengths decrease when economic conditions improve . This pattern was determined based on obser- vations of skirt lengths in the 1920s and 30s.

Furthermore, we need data on how and what consumers wear to express their fashion styles. However, fashion photographs on the internet are one of two types: one displays clothes that consumers themselves choose to wear, and the other consists of photographs taken by professional photographers to promote a clothing line. As previous studies built databases by collecting fashion images from the internet, existing databases do not reflect only the fashion styles chosen by consumers. This is the third limitation of existing fashion data archives.

Taken together, no existing database has everyday fashion images that span over a long period with both timestamp and location information. In this study,

4

S. Takahashi et al. Table 1. Popular fashion databases used in computer science.

Date Unit

we define everyday fashion trends as trends of fashion styles that consumers adopt in their daily lives. Everyday fashion exists on the streets , changes over time, and trends and cyclic patterns can be observed over extended periods.

Hence, we create a new fashion database, CAT STREET, is an attempt to solve these limitations and analyze everyday fashion trends.

Street-Fashion

We created CAT STREET via the following steps. We collected street-fashion photographs once or twice a month in fashion-conscious streets such as Hara- juku and Shibuya in Tokyo from 1980 to date. In addition, we used fashion photographs from a third-party organization taken in the 1970s at monthly in- tervals in the same fashion-conscious streets. Next, by using images from the two data sources, we constructed a primary image database with timestamps from 1970 to 2017 for each image. The photographs from the third-party organization did not have location information; hence, we could only annotate the fashion images taken since 1980 with street tags.

Fashion styles conventionally differ between men and women. To focus on women’s fashion trends in CAT STREET, two researchers manually categorized the images by the subjects’ gender, reciprocally validating one another’s cat- egorizations, and we selected only images of women from the primary image database. Some images from the 1970s were in monochrome; therefore, we gray- scaled all images to align the color tone across all images over the entire period.

Street-fashion photographs generally contain a large amount of noise, which hinders the precise detection of combinations of clothes worn by the subjects of photographs. To remove the noise, we performed image pre-processing as the final step in the following manner. We identified human bodies in the photographs A Novel Approach to Analyze Fashion Digital Archive from Humanities

5

Fig. 1. Data overview of CAT STREET. Harajuku and Shibuya are the most famous streets in Japan and Others include images taken in other fashion-conscious streets such as Ginza and Daikanyama.

using OpenPose , an machine learning algorithm for object recognition, and removed as much of the background image as possible. Subsequently, we trimmed the subject’s head to focus on clothing items based on the head position, which was detected using OpenPose. Fig. 1 shows an overview of the data contained in CAT STREET. The total number of images in the database was 14,688.

At the end of the database creation process, we checked whether CAT STREET met the requirements for a database to capture everyday fashion trends. First, CAT STREET comprises photographs captured on fashion-conscious streets in Tokyo. It reflects the fashion styles women wore in their real lives and does not include commercial fashion images produced for business purposes. Second, CAT STREET has necessary and sufficient annotations to track fashion trends: monthly timestamps from 1970 to 2017 and street-level location tags. Images in the 1970s do not have street tags; however, they were taken in the same streets as the photographs taken with street tags since 1980. Hence, we could use all the images in CAT STREET to analyze the overall trends of everyday fashion in fashion-conscious streets as a representative case in Japan.

Retrieving Everyday Fashion Trends

In this section, we estimated a fashion style clustering model to identify the fashion styles adopted in fashion images. Then, we applied the fashion clustering model to CAT STREET and retrieved everyday fashion trends indicating the

6

S. Takahashi et al. extent to which people adopted the fashion styles in each year. Finally, we verified the robustness of clustering model’s result.

Fashion Style Clustering Model

To build a fashion style clustering model, we selected FashionStyle14 as the training dataset. It consists of fourteen fashion style classes and their tags, as shown in the first row of Fig. 2. Each class consists of approximately 1,000 im- ages, and the database consists of a total of 13,126 images. The fashion styles of FashionStyle14 were selected by an expert as being representative of mod- ern fashion trends in 2017. By applying the fashion clustering model to CAT STREET, we measured the share of each modern style in each year and re- garded it as the style’s trend.

We trained four deep learning network structures as options for our fash- ion clustering model: InceptionResNetV2 , Xception , ResNet50 , and VGG19 . We set weights trained on ImageNet as the initial weights and fine-tuned them on FashionStyle14, which we gray-scaled to align the color tone of CAT STREET, using the stochastic gradient descent algorithm at a learning rate of of 104. For fine-tuning, we applied k-fold cross-validation with k set as five.

The F1-scores are presented in Table 2. InceptionResNetV2 yielded the high- est F1-scores among the deep learning network structures. Its accuracy was 0.787, which is higher than the benchmark accuracy of 0.72 established by ResNet50 trained on FashionStyle14 in the study by Takagi et al. . Therefore, we con- cluded that InceptionResNetV2 could classify gray-scaled color images to fashion styles and adopted the deep learning network structure InceptionResNetV2 as the fashion style clustering model in this study. We applied this fashion style clustering model to the images in CAT STREET, and Fig. 2 shows sample im- ages classified into each fashion style.

Verification

The fashion style clustering model consisted of five models as we performed five- fold cross-validations when the deep learning network structure was trained, and each model estimated the style prevalence share for each image. To verify the clustering model’s robustness in terms of reproducing style shares, we evaluated the time-series correlations among the five models. Most fashion styles had high Table 2. F1-scores of fine-tuned network architectures. The highest scores are under- lined and in boldface.

Inceptionresnetv2 Xception Resnet50 Vgg19

Macro Avg.

0.690

Weighted Avg.

0.689

A Novel Approach to Analyze Fashion Digital Archive from Humanities

7

Fig. 2. Overview of FashionStyle14 and sample images in CAT STREET classified into different fashion styles using the fashion style clustering model. ‘Conserv.’ is an abbreviation for Conservative. The A&F is originally labeled as Ethnic . We changed this original label to ‘A&F’ because this category contains multiple fashion styles such as Asian, African, south-American, and Folklore styles.

time-series correlation coefficients of over 0.8, and the unbiased standard errors were small. Some fashion styles, such as those designated Dressy and Feminine styles, exhibited low correlations because these styles originally had low style shares. These results indicate that our fashion style clustering model is a robust instrument for reproducing the time-series patterns of style shares. There were no images for 1997 and 2009; we replaced the corresponding zeros with the averages of the adjacent values for the analysis in the next section.

Analyzing Fashion Trend Patterns

There are two representative fashion-conscious streets in Tokyo: Harajuku and Shibuya. Geographically, Harajuku and Shibuya are very close and only a single transit station away from each other; however, they have different cultures. In particular, the everyday fashion trends in these streets are famously compared to each other. Some qualitative studies pointed out triggers that form the cultural and fashion modes in each street using an observational method . However, they simply reported the triggers with respect to the street and style and did not compare the functioning of these triggers on the mode formations in every- day fashion trends from a macro perspective. This section sheds light on this research gap and interprets how the fashion trends emerged with non-fashion data sources.

First, we compared two everyday fashion trends in fashion-conscious streets with CAT STREET to investigate how fashion styles ‘boom’ or suddenly increase in prevalence, and classified the resultant patterns. Fourteen styles were classi-

8

S. Takahashi et al. fied into two groups: a group comprising styles that emerged on both streets simultaneously and a group with differing timings for trends observed in the streets. We selected two fashion styles from each group as examples, including A&F and Retro from the first group and Fairy and Gal from the second group; Figs. 3(a), 4(a), 5(a), and 6(a) show their average style shares, respectively.

As a non-fashion data source, we focused our attention on magazines because fashion brands have built good partnerships with the magazine industry for a long time. Fashion brands regard magazines as essential media to build a bridge between themselves and consumers, and magazines play a role as a dispatcher of fashion in the market. We quantified the ‘magazine’ trends to capture how magazines or media sent out information to consumers. For this purpose, we used the digital magazine archive of Oya Soichi Library . Oya Soichi Library has archived Japanese magazines since the late 1980s and built the digital archive.

The digital archive houses about 1,500 magazines, and one can search for article headlines of about 4.5 million articles. The archive covers a range of age groups and mass-circulation magazines that people can easily acquire at small bookstores, convenience stores, and kiosks.

By searching for headlines including fashion style words and specific topics from articles with tags related to fashion, we could quantify the ‘magazine’ trends that indicate how many articles dealt with the styles and specific topics to spread the information to a mass audience (Figs. 3(b), 4(b), 5(b), 7, and 8). The ‘magazine’ trends labeled as Articles in the figures were normalized to the range of 0 to 1 for comparison.

Finally, to infer why and how the everyday fashion trends emerged, we com- pared the fashion style shares extracted from CAT STREET to the ‘magazine’ trends in the Oya Soichi Library database in chronological order.

Simultaneous Emergence Of Fashion Styles

We selected two styles, A&F and Retro, as examples in the group of styles that simultaneously emerged on both streets. The A&F style is inspired by native costumes . Fig. 3(a) shows the A&F style’s upward trend in the late 1990s in Harajuku. Simultaneously, the A&F style gradually became accepted in Shibuya in the late 1990s and reached the same share level in the mid-2000s. The Retro style, another example in the first group, is an abbreviation of retrospective style . According to this style’s definition, an overall downward trend is plausible in both streets, as shown in Fig. 4(a). Fashion revival is one of the relevant fashion trend phenomena of the Retro style, and there are a wide variety of substyles representing fashion revivals under the Retro style, such as the 60s look, 70s look, and 80s look. In Fig. 4(a), slight peaks can be observed in the early 1980s and late 1990s, suggesting that the 80s look, which was in vogue in the early 1980s, was revived in the late 1990s.

To quantify how magazines spread information about these styles, we plotted the ‘magazine’ trends by searching for articles with headlines that included terms related to the style names. Figs. 3(b) and 4(b) indicate two relationship patterns between style shares and ‘magazine’ trends; Fig. 3(b) shows a synchronization

Related Journal Articles & DOI Links

Selected peer-reviewed publications relevant to 12 Lead ECG Acquisition. Click the DOI to access the full paper (may require institutional access).

Why Choose Us?

Bangalore guidance for robotics, Spectre and autonomous systems projects.