|
| 基于注意力特征统计匹配和尺度风格捕捉的动漫游戏视频风格稳定迁移方法 |
| A Stable Style Transfer Method for Anime and Game Videos Based on Attention Feature Statistical Matching and Scale Style Capture |
| 投稿时间：2025-12-24 修订日期：2026-04-26 |
| DOI： |
| 中文关键词: 风格迁移 注意力机制 特征统计匹配 尺度风格捕捉 动漫游戏 |
| 英文关键词: Style transfer Attention mechanism Feature statistical matching Scale style capture Animation and games |
| 基金项目:江苏省概念验证中心项目，全固态高性能轻量化微片皮秒激光器（POCC2024M08）；安徽省质量工程一流核心课程项目，三维设计与制作（2023hxkc174） |
|
| 摘要点击次数: 9 |
| 全文下载次数: 0 |
| 中文摘要: |
| 动漫与游戏在数字娱乐中的广泛应用,动漫游戏的视频风格一致性已成为当前研究的热点。然而,现有风格迁移方法多以图像为研究对象,应用于视频时易产生风格闪烁、风格表达不足等局限。为此,研究提出一种基于注意力特征统计匹配与尺度风格捕捉的动漫游戏视频风格稳定迁移方法。以注意力特征统计匹配为基础,引入多尺度风格捕捉机制,增强风格表达的同时提升视频迁移稳定性。实验结果表明,该方法在视频稳定性验证中,时间感知图像块相似度(Temporal Learned Perceptual Image Patch Similarity,tLPIPS)误差降至0.07,Warping Error在高速运动场景下控制在0.09。在动漫与游戏风格表达实验中,多尺度融合使弗雷歇距离(Fréchet Inception Distance,FID)显著降低至42.80,基于边缘的结构相似性指数(Edge-based Structural Similarity Index Measure,Edge-SSIM)达到0.85,并在高频线条风格中取得0.92的高保真度。综上,研究方法能够在动漫与游戏视频场景中实现兼具风格表现力与时序稳定性的风格迁移效果。 |
| 英文摘要: |
| The widespread application of animation and games in digital entertainment has made the consistency of video style in animation and games a hot research topic. However, existing style transfer methods mostly focus on images, which can lead to limitations such as style flickering and insufficient style expression when applied to videos. To address this, this study proposes a stable style transfer method for animation and game videos based on attention feature statistical matching and scale style capture. Based on attention feature statistical matching, a multi-scale style capture mechanism is introduced to enhance style expression and improve video transfer stability. Experimental results show that in video stability verification, the Temporal Learned Perceptual Image Patch Similarity (tLPIPS) error is reduced to 0.07, and the Warping Error is controlled at 0.09 in high-speed motion scenes. In experiments on style expression in animation and games, multi-scale fusion significantly reduced the Fréchet Inception Distance (FID) to 42.80, achieved an Edge-based Structural Similarity Index Measure (Edge-SSIM) of 0.85, and obtained a high fidelity of 0.92 in high-frequency line styles. In summary, the research method can achieve style transfer effects that combine stylistic expressiveness and temporal stability in animation and game video scenes. |
|
View Fulltext
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
