An Intelligent Analysis And Prediction Framework For Womenswear Silhouette Based On Data Mining Technology

The modernized Chinese industry has been experiencing some major changes recently.One of changes is to advocate the innovation behind the Chinese brands.In the Chinese fashion industry,question is raised on how to transform the “made in China” situation to “designed in China”,which helps to boost the competitiveness to the local brands over the world.For the consumers purchasing the clothing,the clothing silhouette is also an important consideration in addition to the color,texture and material.Therefore,the clothing silhouette,which is the start of the fashion design,is an important task to fulfill for an innovative design.However,the systematic silhouette studies in the literature were rather limited.For example,only finite amount of research was done to regarding to the standardization definition,intelligent identification,trend prediction and related research to the clothing silhouette.Meanwhile,the biannual fashion week around the world release massive clothing production and catwalk pictures,which include different kinds of edge cutting information.However,the fashion designers could not fully get the inspiration from these pictures due to the complex nature of such information and the difficulty to process the information.In addition,with the advancement the modern technologies like the big data,the fashion designers could utilize the computer power as a tool to facilitate their designs.Consequently,the tools and methodologies to efficiently and intelligently extract the useful information out of massive amount of data become rather desirable.With advanced interdisciplinary technologies,this dissertation focuses on various womenswear silhouette related tasks,which first identifies,extracts,and measures the silhouette from the static catwalk images.Then,under the framework of data mining,the dissertation proposes a womenswear silhouette trend prediction system based on the BP neural network.The proposed research serves as a quick,convenient,and accurate silhouette trend estimation tool for the fashion brands and designers for women,offering new ideas for the future original designs.The presented work contains three major parts.The first part of the dissertation highlights the commonly used methods viable for the body segmentation,face detection,skin detection,graph cut and so on.The human body is characterized by the Maximum A Posteriori probability model,and the foreground models from the original figures are extracted with the graph cuts algorithm.The details of these algorithms and principles are explained herein.Then batch processing is adopted due to a large sample of data by using the developed computer technology and mathematical models.With the segmented model images,the next step includes the definition of the key points of the generalized silhouettes,the measurement of the corresponding dimensions by Hough transform method,and the process to archive the data.The created database according to such procedures serves as the input to numerically establish the definition of womenswear silhouette,which are usually determined by the subjective judgments of the fashion designers.Furthermore,the intelligent measurement criterions focusing on various special cases are presented.The measurement procedure covers a wide range of scenarios including those when the shoulders(the boob tube top)and knees(the skirts,shorts,and the bottomless designs)are not included in the design.Besides,a meaningful concept of “drastic change point” is proposed,which is proved important to numerically handle a wider range of womenswear clothing..Therefore,based on the proposed measurement procedures,the collected data for the numerical evaluation of womenswear silhouette becomes more reliable.The outcome of the measurement could be used in the subsequent sections.The second part of the study proposes an innovative,numerical,and standardized womenswear silhouette classification and definition following the commonly used alphabetical convension.Such definition is a huge contribution in the field of fashion design as very limited amount of standardized silhouette definition could be found in the literature.The numerical definitions of the five basic silhouettes(i.e.,A-line,H-line,T-line,O-line,X-line)as well as S-line(i.e.,one of the most commonly observed silhouettes)are provided.The proposed method could automatically classify the six aforementioned womenswear silhouettes considering the different slopes collected from four main clothing parts: the shoulders,the waist,the hip,and the knees.In addition,six different clothing situations are discovered according to different designs,which are: 1.continuous with bottom designs,2.continuous and bottomless,3.with bottom designs and drastic change between the shoulder and the waist,4.with bottom designs and drastic change between the waist and the hip,5.with bottom designs and drastic change between the hip and the knees,6.bottomless and with a drastic change point elsewhere.Following the provided steps,womenswear silhouettes for all the cases from the measurement database are identified and categorized to the corresponding ensembles systematically.It could be concluded that the descriptive definition of the silhouette is successfully transformed into numerical definition.Last but not the least,the validation study shows that the presented identification procedure has an accuracy over 93%.The research has a practical significance,reducing the likelihood of misjudgment and improving the efficiency while identifying the womenswear silhouettes.Finally,the study summarizes the fashion release channels,the aspects influencing the silhouette trend,and the general fashion releasing process.In addition,the brief concepts and principles of the data mining theory,time series data mining theory,and BP neural network model are highlighted,which are the tools used to explore the future trends of the womenswear silhouettes.Due to the lack of silhouette trend prediction research,an innovative womenswear silhouette trend prediction methodology is proposed in the study.The proposed method is stemmed from the data mining framework that could process tremendous amount of data from existing and future studies.The outputs from the former sections of the dissertation,i.e.the womenswear silhouette dimension database and the intelligent silhouette identification system,are feed into the established BP neural network model for the trend estimation.Besides,the measurement,the identification,and the classification procedures are fully validated by the image data collected from 15 world famous brands over 11 consecutive seasons.Through the analysis and data mining of the collected womenswear silhouette data,the study reveals a list of reliable silhouette trends.The corresponding accuracy of the prediction system is relatively high,which is illustrated by the error histograms and output/target comparison charts.Multiple innovative aspects are presented in this dissertation,which contributes to the embryo research to numerically define and evaluate clothing silhouettes.The major contributions of the work are summarized as:(1)Realization of the intelligent extraction and dimension measurement of the womenswear silhouette.The dissertation proposes the intelligent body segmentation procedures to process the static catwalk images and the automated measurement of the key dimensions of the womenswear silhouettes based on several mathematical models.With such efforts,the database of the womenswear silhouette dimensions could be established.It is further shown that the database is accurate in the measurement of the key silhouette dimensions,considering all of the observed design variabilities and establishing relevant rules of measurements.(2)Establish the numerical standard of the classification of the womenswear silhouetteThe dissertation establishes the numerical definition of the six womenswear silhouettes(A-line,H-line,T-line,O-line,X-line and S-line)according to their shape features.Meanwhile,the results are calculated by the angles between the main supporting parts of the human body(the shoulder,the waist,the hip and the knee).The proposed silhouette definition helps the designers to significantly reduce the mistakes caused by the subjective judgments.In addition,such definition also paves the necessary foundations for the future numerical and theoretical studies of the silhouette.(3)Realization of the intelligent identification of the womenswear silhouetteBased on the proposed numerical definition of the womenswear silhouette,the dissertation further divides the womenswear into six sub-conditions based on different silhouette designs.Corresponding rules are set up to realize the identification of the womenswear silhouettes.The identification procedure helps the designers to automatically determine the womenswear silhouette with high efficiency and accuracy.and provides the basic technical support for the analysis in the trend prediction of the silhouette at the same time.(4)The intelligent trend prediction of the womenswear silhouetteThe dissertation predicts the trend of the womenswear silhouette under the data mining framework.According to the time-series characteristics of the collected data,the BP neural network model is applied as the mining technology for the system.The dissertation verifies and predicts the trend of the womenswear silhouette by the categorized silhouette information obtained from the catwalk images from famous womenswear brands.It is shown that the proposed system is accurate and efficient,which provides reliable references for the future designs of the womenswear silhouettes.