查詢結果分析
來源資料
頁籤選單縮合
題名 | 使用人體姿態遷移技術、臉部姿勢轉換技術以及影像修復技術之舞蹈影片生成系統=Dance Video Generation System Using Human Pose Transfer, Facial Posture Transfer and Image Inpainting Techniques |
---|---|
作者姓名(中文) | 鄭旭詠; 余執彰; | 書刊名 | 前瞻科技與管理 |
卷期 | 12:2 2024.05[民113.05] |
頁次 | 頁29-49 |
分類號 | 312.83 |
關鍵詞 | 人工智慧; 生成對抗網路; 生成模型; 姿勢轉移; 影像修復; Artificial intelligence; Generative adversarial network; Generative model; Pose transfer; Image inpainting; |
語文 | 中文(Chinese) |
DOI引用網址 | 10.6193/JATM.202405_12(2).0003 |
中文摘要 | 本篇論文利用生成對抗網路建立了一個舞蹈影片生成系統,可將單張影像和目標舞蹈影片輸入,使照片中的人物跳舞。系統主要採用人體姿態遷移技術以及臉部姿勢轉換技術,讓電腦生成符合目標姿態的人物影像,並修復由背景切割導致的空洞。此外,系統還採用多尺度區域提取器捕捉身體特徵,並將區域風格損失納入損失函數。在臉部姿勢轉移方面,採用漸進式人臉角度轉換架構,並加入遮罩鑑別器以提高生成影像品質。最後,針對修復過程中的顏色偏移問題,系統使用基於CIEDE2000色差公式的感知顏色損失函數進行處理,使修復結果更符合人類視覺感知。 |
英文摘要 | This paper builds a dance video generation system using a generative adversarial network, which can input a single image and a target dance video to make the people in the photo dance. The system mainly adopts human posture transfer technique and facial posture transfer technique to allow the computer to generate an image of a person that matches the target poses. Also, image inpainting technique repairs the holes in the picture caused by the characters being cut out from the background. In addition, the system uses a multi-scale region extractor to capture body features and incorporates region style loss into the loss function. For face pose transfer, a progressive face angle transformation framework is adopted and a mask discriminator is added to improve the quality of the generated image. Finally, for the color shift problem during the inpainting process, the system uses the perceptual color loss function based on the CIEDE2000 color difference formula to handle the problem, so that the inpainting results match human visual perception better. |
本系統之摘要資訊系依該期刊論文摘要之資訊為主。