請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/49921
標題: | 從街頭看時尚:應用深度學習瞭解流行趨勢 From Street Photos to Fashion Trends: A Deep Learning Approach for Fashion Understanding |
作者: | Fu-Hsien Huang 黃富纖 |
指導教授: | 盧信銘(Hsin-Min Lu) |
共同指導教授: | 許耀文(Yao-Wen Hsu) |
關鍵字: | 服裝辨識,多標籤分類,含噪標籤,電腦視覺,深度學習,多任務學習,流行趨勢分析, Fashion Recognition,Multi-Label Classification,Noisy Labels,Computer Vision,Deep Learning,Multi-Task Learning,Fashion Trend Analysis, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 現今的社群媒體上充滿了大量服飾穿搭的照片,為了解街頭時尚(Street Fashion)提供了新的管道。為了更了解亞洲街頭時尚,我們建立了Multi-task Fashion Attributes Recognition Network (FARNet)深度學習模型進行多標籤分類(Multi-label Classification),此模型利用社群網站使用者所提供的含噪標籤(Noisy Labels)及穿搭圖片,辨識圖片中的服裝顏色、類別及花色,並同時處理含噪標籤、將其修正為正確標籤。另外,為提供FARNet模型訓練資料,本研究建立一個全新的資料集,命名為RichWear。此資料集包含322,198張以日本為主的街頭時尚照片,同時含有大量的文字標籤、使用者資訊、服裝品牌,以及使用者加入的Hashtags;資料集中所有圖片的服裝屬性均有含噪標籤,另有一個4,368張圖片的子集,除了含噪標籤外,也加上人工驗證標籤(Human Verified Labels),供模型訓練使用。實驗結果顯示,我們的方法在測試資料集上的預測表現,顯著優於其他基準線方法。至於街頭流行趨勢分析方面,我們利用FARNet模型預測出來的服裝標籤探索服裝流行趨勢,並且對RichWear資料集的圖片以高斯混合模型(Gaussian Mixture Model, GMM)進行集群分析(Cluster Analysis),藉此,我們成功找出了日本與其他亞洲地區的街頭流行趨勢,並且發現時尚風格(Fashion Style)會隨著季節而變化,具有時間上的動態性。 There has been an increasing interest in using deep learning and computer vision techniques for fashion recognition. However, most existing methods predict multiple clothing attributes of fashion images individually, rather than simultaneously. In addition, few studies focus the fashion trend analysis on street styles of Asian areas. In this work, we create a new dataset named RichWear, which contains 322,198 street fashion images with massive labels, users’ information, clothing brands, and user-created hashtags. The dataset focuses on street styles of Japan and other Asian areas. In addition to noisy labels, there is also a subset with expert-verified clothing attributes. To improve fashion recognition, we propose a multi-task Fashion Attributes Recognition Network (FARNet) for multi-label classification. Instead of predicting each attribute individually, FARNet simultaneously predicts three types of clothing attributes; meanwhile, it addresses the noisy labels and correct the noisy labels at the same time. Experimental results show that our methods significantly outperform the compared baselines. For street fashion analysis, we use the predicted labels to find clothing trends, and then perform clustering on the images of RichWear to gather visually correlated images. We successfully find street fashion trends as well as discovering style dynamics in Japan and other Asian areas. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/49921 |
DOI: | 10.6342/NTU202003000 |
全文授權: | 有償授權 |
顯示於系所單位: | 統計碩士學位學程 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-1108202019023500.pdf 目前未授權公開取用 | 4.75 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。