| Author: | Zou, Xingxing |
| Title: | Intelligent clothing aesthetic model |
| Advisors: | Wong, Wai Keung Calvin (ITC) |
| Degree: | Ph.D. |
| Year: | 2021 |
| Subject: | Artificial intelligence Fashion Clothing trade Aesthetics Hong Kong Polytechnic University -- Dissertations |
| Department: | Institute of Textiles and Clothing |
| Pages: | xx, 201 pages : color illustrations |
| Language: | English |
| Abstract: | Over the last decade, deep learning has become a popular technique because of its promising performance in many tasks. State-of-the-art research has been conducted in many fields, including medical, security and education. Recently, the use of artifcial intelligence (AI) in fashion has become an hot topic because of its massive economic potential. However, mainstream approaches suffer from limited generalisability to a specific field for the following reasons: 1. Current AI technologies are all data driven. Their performance is closely concern with the quality of the training set which used for training; 2. They have generally low generalisability to different datasets and poor applicability in different fields; 3. Each specific field has its own characteristics, thus requiring the technologies to be adjusted accordingly. In this study, the author aims to find AI solutions to handle tasks in the fashion domain on the basis of an understanding of clothing aesthetics. There are three progressive levels of tasks in fashion: fashion lover, i.e. which means a model could recognise fine-grained fashion attributes or fashion concepts; personal stylist, i.e. which means a model could mix and match according to the specific requirements, e.g. different people and occasions; design assistant, i.e. which means a model could offer some inspirations for fashion designers. After investigating tasks of the three layers, the author found that basic knowledge learning is the foundation of high level fashion tasks. Thus, the author first solves the problems of basic knowledge learning. Firstly, a fine-grained fashion attribute dataset is constructed with a hierarchical and comprehensive structure of fashion knowledge. With a clear definition of each attribute, the images are manually annotated to construct the FashionAI dataset, which includes both design attributes and key points. The occlusion or uncertain problems that occur frequently in practical scenarios are taken into consideration. In addition, a Fashion Color (FoCo) system is defined for feature extraction of fashion colour to further enhance the recognition ability of colour recognition. The hue-saturation-brightness (HSB) colour system is adopted as the base of FoCo. The HSB colour space is regarded as the set consists of 720 colour units. Those units are further grouped to favour high-order tasks, i.e. colour categories for analyzing fashion style and colour tones for recognizing fine-grained colour. After solving the problems in basic knowledge learning and the ability to recognise basic knowledge in fashion is obtained, three high-order tasks in fashion are addressed on the basis of an understanding of clothing aesthetics. The first task is explainable outfit evaluation. From the perspective of fashion, an outfit is not bad because the attributes are considered as a whole to achieve visual balance. The evaluation judgment is defined by using absolute ratings, i.e. good, normal and bad instead of relative ratings, i.e. the embedding scores. Two approaches are proposed. The first one is based on the decision tree. After recognising the attributes, the model can evaluate an outfit based on mix-and-match principles. This way is most similar to the process of human evaluation. However, it remains a weaknesses of the traditional decision tree, which cannot cover all situations. To solve this issue, the author further proposes an approach that extracts those mix-and-match principles from data. The author also solves two fashion tasks: multidomain fashion product retrieval and automatic fashion illustration generation. The first task addresses the problem of multidomain searching, including domains of sketches, drawings and real products. The author proposes the fAshIon-DA dataset with a feature norm-based approach to enable the sketch to the shop application. The author also introduces a new application i.e. fashion illustration generation with a new approach. The whole framework is based on the classical GAN model with three new parts: 1. the kShapes to provide the position information for the network; 2. the kShapes are adopted to adjust the body-shape between real-photo and the illustration images; 3. the contextual fashionable loss is proposed to further improve the quality of generated images. The presented approach outperforms all the mainstream generative methods in current on the fashion illustration dataset. All in all, this study is based on the assumption that machines could not express their own emotions. Thus, only the physical features i.e. color, material, silhouette etc are taken into consideration. After solving the problems in current basic knowledge learning, the author attempt to solve the problem in a reasoning way based on the clothing aesthetic understanding. Both the datasets and technologies are specifically designed from the perspective of fashion. As far as the author know, this study is the first research to investigate this cross-domain from the perspective of fashion. All proposed frameworks for computer vision tasks are faithfully designed according to the real situation of the fashion industry. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/11321