重庆工商大学学报（自然科学版）2026，Vol.43Issue(4)：10-17,8.DOI:10.16055/j.issn.1672-058X.2026.0004.002
基于轻量级和笔触的超高清图像艺术风格迁移
Artistic Style Transfer for Ultra-High-Resolution Images Based on a Lightweight Model with Large Brushstrokes
摘要
Abstract
Objective With the continuous advancement of modern technology,image resolution and user demands continue to increase.However,limited GPU memory frequently causes out-of-memory errors during ultra-high-resolution(UHR)style transfer tasks.Furthermore,existing UHR artistic transfer methods often neglect brushstroke characteristics,predominantly relying on small brushstrokes for style transfer.To address these challenges,this paper presents a lightweight framework integrating patch cutting,termed the Patch Cutting and Lightweight Model(PCLM).Methods To effectively reduce memory consumption,the proposed model employed a dual optimization strategy.First,UHR images were segmented into patch collections,a process that broke down a large-scale task into smaller subtasks to fundamentally address the memory bottleneck.Simultaneously,knowledge distillation was utilized to compress model parameters.In addition,a patch relation loss(PRL)was proposed to control both brushstroke size and patch consistency during style transfer.Results Extensive qualitative and quantitative experiments demonstrated that PCLM successfully realized UHR style transfer with large brushstrokes under constrained memory conditions.While maintaining a minimal memory footprint,the model achieved the following evaluation metrics on the test set:0.495 928(SSIM),11.610 026 dB(PSNR),0.546 720(LPIPS),and 3.373 454 GB(GPU memory usage).Conclusion Theoretical analysis and empirical results confirm that PCLM effectively resolves the memory consumption challenge and generates high-quality UHR artistic paintings featuring large brushstrokes.This study provides a valuable reference for subsequent UHR image processing tasks.关键词
切割补丁/超高清/模型压缩/大笔触Key words
patch cutting/ultra-high-resolution/model compression/large brushstroke分类
信息技术与安全科学引用本文复制引用
杨天,丁一峰,尹淑婷,李长庚..基于轻量级和笔触的超高清图像艺术风格迁移[J].重庆工商大学学报（自然科学版）,2026,43(4):10-17,8.基金项目
安徽省自然科学基金项目资助(2023085MF220). （2023085MF220）
