| 摘要: | 網路購物的蓬勃發展帶動實體商店逐漸轉型，積極開發網路銷售平台，為了在電子銷售通路上提升其銷售額，因此，欲提供更多商品相關資訊給消費者，激發客戶對商品的期待感，其中，服飾類別為網路購物銷售主要產品之一，因此，各種服飾品牌紛紛推出虛擬試衣，消費者不用親自試穿衣服，也能有對該衣服之參考依據。
電腦視覺技術為虛擬試衣重要之一環，生成對抗網路(GAN)已被廣泛使用於此類領域，在本論文中，我們採取三個譜歸一化生成對抗網路(SNGAN)及Openpose人體節點偵測去生成盡可能自然的試衣成果。三個譜歸一化生成對抗網路分別用以生成目標人物手臂、目標人物著目標衣服之變形衣服與最後的試衣結果，因非傳統3D建模後再試衣，因此，本論文將以視覺試衣稱之，整體架構由三模塊構成：語意生成模塊、衣服變形模塊、內容融合模塊，透過空間變形網路保留衣服細節，用以改善視覺試衣，使生成結果更加貼近真實自然的試衣成果，並與其他亦為視覺試衣之實驗進行比較。
本論文僅針對上半身試衣，並對衣服、手部另做評比，搭配不同的損失函數，比較生成對抗網路之訓練結果。未來，本論文所提及之模型亦可應用於其他各式類型實驗，例：髮型設計、下半身試衣、衣物設計模擬等。
關鍵詞：生成對抗網路、譜歸一化生成對抗網路、虛擬試衣、視覺試衣、人體骨幹偵測
;The vigorous development of online shopping drives the transformation of physical stores. The physical stores actively develop online sales platforms in order to increase their sales on electronic sales channels. Therefore, we want to provide more product information to consumers. With the information, we can stimulate customers’ desire for the products. Then customers buy it. Clothing is one of the main products sold in online shopping. Therefore, various clothing stores have launched virtual try-on system. Consumers do not need to try on the clothes in person, and they can have a reference for the clothes.
Computer vision technology is an important part of virtual try-on system. Generative Adversarial Network has been widely used in this field. In this paper, we adopt three Spectral Normalized Generative Adversarial Networks and Openpose for human body keypoints detection. The three Spectral Normalized Generative Adversarial Networks was used respectively to generate the arms of target person, the warping clothes, and the final try-on result. We will call it visual try-on because of it only depends on image. The proposed method is composed of three modules, including semantic generation module, clothing warping module, and content fusion module. The clothes details are retained through the spatial transformer network. We hope to make the generated results as close as to the reality. This paper only try-on the upper clothing. We compare clothes and hands separately, with different loss functions. In the future, the methods mentioned in this paper can also be applied to various other types of experiments, such as hair styling, lower clothing try-on, clothing design simulation.
Keywords: Generative Adversarial Network(GAN), spectral normalized Generative Adversarial Networks(SNGAN), virtual try-on, visual try-on, Openpose |
