基於圖像之語意導向與細節保留的 2D 虛擬試穿網路

Semantic-Guided and Detail-Preserving Image-Based 2D Virtual Try-On Networks

2D 虛擬試穿是近年來熱門的研究項目，只要輸入欲穿上的衣服圖片，便能更換指 定人物圖像的服裝。本研究提出了基於圖像之語意導向與細節保留的2D 虛擬試穿網路 (SD-VTON)，將基於較新穎的ACGPN 網路架構去做改善，透過ACGPN 三大架構： 語意生成、衣服變形、內容合成，我們發現語意生成是造成錯誤的最大原因，因此將神 經網路修改為UNet++來改善人物解析，欲減少epoch 來產生同等或更佳的生成結果， 修改神經網路訓練深度或進行其他參數的調整，在200 epochs 的測試結果中，有86.1% 的SSIM 分析值優於ACGPN，且於140 epochs 就有與其相近的試穿結果，這意味著我 們將可以省去不少訓練的時間。在實用性方面，SD-VTON 在整個架構最初加入剪取衣 服的步驟，利用LIP-JPPNet 生成語意模板，開發一個品質更好的虛擬試穿系統。

 

2D virtual try-on has become a hot topic in recent years. It can change what a person image wearing by inputting a desired clothes image. In this study, we propose a visual try-on network, namely Semantic-guided and Detail-preserving Image-based 2D Virtual Try-On Networks(SD-VTON), which improve the architecture of a novel network, ACGPN. Considering the three major modules of ACGPN, including the Semantic Generation Module(SGM), the Clothes Warping Module (CWM), and the Content Fusion Module(CFM), we find out that SGM is the main reason causing the problems mentioned above. Consequently, we substitute UNet++ for the original neural network to improve human parsing, and hoping it makes the same or even better result with fewer epochs. In the test result of 200 epochs, there are 86.1% SSIM scores higher than ACGPN’s. Moreover, the result of 140 epochs is extremely close to that of 200 epochs, which means that we can save more time on training. In terms of practicality, SD-VTON add a Clothes Tailoring Module at the head of the overall architecture, generating semantic segmentation with JPPNet, and develop a virtual try-on system with higher quality.