Build a Virtual Dressing Room using Deep Learning
| dc.contributor.advisor | Lê, Đình Huynh | |
| dc.contributor.author | Nguyễn, Trọng Duy | |
| dc.contributor.author | Bùi, Tú Anh | |
| dc.contributor.author | Nguyễn, Huy Hoàng | |
| dc.date.accessioned | 2024-02-23T03:15:13Z | |
| dc.date.available | 2024-02-23T03:15:13Z | |
| dc.date.issued | 2023 | |
| dc.identifier.uri | http://ds.libol.fpt.edu.vn/handle/123456789/3991 | |
| dc.description.abstract | With the rapid development trend of the fashion e-commerce industry, people are purchasing most of their items online and are spending more on them, especially fashion items, since browsing different styles and categories of clothes is easy with just a few mouse clicks. However, with the convenience that online shopping provides, customers tend to worry about how the image of a particular fashion item on the website will fit them. Additionally, retailers tend to automate sales steps with AI technology gradually. Hiring models or photo studios to advertise products takes a lot of time and money for fashion retailers. To solve this problem, 3D virtual fitting technologies have been created, but the difficulty of measuring the depth of clothing and body shape takes more time than 2D images. In this project, we build a virtual fitting system using Deep Learning technology with input from 2D images of people trying on clothes. Our product undergoes two key phases: initially integrating the HR-VTON and GPVTON virtual try-on methods with DressCode and VITON HD datasets, aiming to present high-fidelity visual representations of garments. The meticulous preprocessing of user-provided data involves six structured steps followed by established methodologies. Addressing challenges in the Openpose step, including finger keypoint loss, requires a comprehensive reiteration of the Human Parse for the DressCode dataset. Experiments with low-resolution images prompted reassessment, substituting resolution enhancement with StableSR after SRGAN phrase generation. Efforts to balance realism and fidelity lead to meticulous crafting of user interaction aspects, integrating interface and backend functionalities for a virtual dressing room with super high-resolution image upscaling. These measures aim to enhance user experiences within the project's scope. The preprocessing phase refined Densepose and Agnostic steps for improved detail integration into the Try-on model. Challenges in the HR-VITON method, generating fixed-resolution and low-sharpness images, lead to the use of StableSR for resolution augmentation, surpassing input image resolution. The Virtual Dress Room offers "Virtual Dressing" and "Upscaling Resolution" features, allowing users to virtually dress models and flexibly adjust image resolution based on preferences. | en_US |
| dc.language.iso | en | en_US |
| dc.publisher | FPTU Hà Nội | en_US |
| dc.subject | Trí tuệ nhân tạo | en_US |
| dc.subject | Artificial Intelligence | en_US |
| dc.subject | Deep Learning | en_US |
| dc.subject | Virtual Try-on | en_US |
| dc.subject | Image Super Resolution | en_US |
| dc.title | Build a Virtual Dressing Room using Deep Learning | en_US |
| dc.title.alternative | Xây dựng phòng thử đồ ảo sử dụng công nghệ học sâu | en_US |
| dc.type | Thesis | en_US |
Các tập tin trong tài liệu này
Tài liệu này xuất hiện trong Bộ sưu tập
Bộ sưu tập thuộc về Trung tâm Thông tin - Thư viện - Trường Đại học FPT
Địa chỉ: Phòng 207 - Tầng 1 - Km 28 - Khu công nghệ cao Hòa Lạc - Thạch Hòa - Thạch Thất - Hà Nội
Điện thoại: 844.66805912 - FAX: - Email: thuvien_fu_hoalac@fpt.edu.vn