基于Transformer两阶段策略的古代服饰线图提取  

Ancient clothing line drawing extraction based on Transformer two-stage strategy

作　　者：周蓬勃[1] 冯龙 武浩东 寇宇帆 ZHOU Pengbo;FENG Long;WU Haodong;KOU Yufan(School of Art and Media,Beijing Normal University,Beijing 100032,China;National and Local Joint Engineering Research Center for Cultural Heritage Digitization,Northwest University,Xi’an 710127,China)

机构地区：[1]北京师范大学艺术与传媒学院,北京100032 [2]西北大学文化遗产数字化国家地方联合工程研究中心,陕西西安710127

出　　处：《西北大学学报(自然科学版)》2025年第1期75-84,共10页Journal of Northwest University(Natural Science Edition)

基　　金：国家自然科学基金(62271393);国博文旅部重点实验室开放课题(1222000812,CRRT2021K01)。

摘　　要：古代服饰线图提取旨在精确获取轮廓与形状信息,以助于再创作和传统服饰保护。但现有方法增加网络以提高泛化性,导致参数量大增。为此,提出了基于Transformer的两阶段边缘检测方法,旨在解决图像局部信息丢失以及模型参数量大的问题。第一阶段将图像分割成16×16粗粒度补丁,利用编码器进行全局自注意力计算以捕获补丁间依赖;第二阶段采用8×8细粒度无重叠滑动窗口覆盖图像,通过局部编码器计算窗口内注意力有效捕捉细微边缘且降低成本。设计了轻量特征融合模块,支持全局与局部特征的高效整合。实验结果表明,该方法在古代服饰和公共数据集上边缘轮廓信息提取效果优于现有方法,ODS指标平均提升15.9%。虽然OIS和AP未超过Informative Drawing,但在模型体量和耗时方面具有明显优势。The extraction of ancient costume line drawings aims to precisely obtain contour and shape information to aid in re-creation and traditional preservation.However,existing methods increase network depth to improve generalization,leading to a significant increase in the number of model parameters.Therefore,this paper proposes a two-stage edge detection method based on Transformer,aiming to solve the problems of local information loss in images and large model parameter sizes.The first stage divides the image into 16×16 coarse-grained patches and uses an encoder to perform global self-attention calculations to capture dependencies between patches;the second stage covers the image with an 8×8 fine-grained non-overlapping sliding window and calculates the attention within the window through a local encoder to effectively capture subtle edges and reduce costs.A lightweight feature fusion module is designed to support efficient integration of global and local features.Experimental results show that this method outperforms existing methods in extracting edge contour information on ancient costume and public datasets,with an average improvement of 15.9%in the ODS metric.Although OIS and AP does not surpass Informative Drawing,this method shows obvious advantages in model size and time consumption.

关 键 词：边缘检测 TRANSFORMER 轻量特征融合模块 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]