摘要
传统服饰纹样的数字化修复是文化遗产保护的核心任务,但面临真实样本稀缺、损坏类型复杂及修复效率低等挑战。本研究提出一种基于迁移学习的智能修复框架,以ImageNet预训练UNet为基础网络,引入可变形卷积增强曲线特征建模能力,设计跨层注意力机制实现多尺度特征融合,并构建融合边缘损失与结构相似性的复合损失函数。实验结果表明,所提方法(TL-UNet)在噪声、划痕等场景下平均PSNR达29.88 dB,较传统稀疏编码法(SC)与标准UNet分别提升23.5%与9.7%,修复耗时仅0.15s。混淆矩阵显示其对10类纹样元素的分类准确率达96.7%,真实文物测试中FSIM均值保持0.842。
The digital restoration of traditional clothing patterns is the core task of cultural heritage protection,but it faces challenges such as the scarcity of real samples,complex types of damage,and low restoration efficiency.This study proposes an intelligent repair framework based on transfer learning,taking ImageNet pre-trained UNet as the basic network,introducing deformable convolution to enhance the curve feature modeling ability,designing a cross-layer attention mechanism to achieve multi-scale feature fusion,and constructing a composite loss function that integrates edge loss and structural similarity.The experimental results show that the proposed method(TL-UNet)has an average PSNR of 29.88 dB in scenarios such as noise and scratches,which is 23.5%and 9.7%higher than the traditional Sparse coding method(SC)and the standard UNet respectively.The repair time is only 0.15 seconds.The confusion matrix shows that its classification accuracy rate for 10 types of pattern elements reaches 96.7%,and the average value of FSIM remains at 0.842 in the real cultural relics test.
作者
骆王琴
高云兵
潘涛
LUO Wangqin;GAO Yunbing;PAN Tao(Anhui Sanlian University,Hefei 230000,China)
出处
《佳木斯大学学报(自然科学版)》
2025年第7期133-136,共4页
Journal of Jiamusi University:Natural Science Edition
基金
校级项目:服装与服饰设计专业改造提升(22zlgc004)
安徽省高等学校2022年度研究项目(2022AH051976)。
关键词
迁移学习
纹样修复
深度学习
边缘损失
transfer learning
pattern restoration
deep learning
edge loss
