数字赋能:基于视觉Transformer的非遗苗绣纹样数字化提取

Digital empowerment:Digitized extraction of patterns of the intangible cultural heritage Miao embroidery based on visual Transformer

苗绣通常是以线稿描绘的基础图样,搭配丰富色彩的丝线及不同的刺绣技法而形成,出现在一系列如服饰、头饰及配件等纺织品中,但是随着纺织品的损坏和丢失等原因,部分苗绣纹样便也随之流失。目前采用传统手工描绘纹样获取线稿的方式极其不便,因此文章针对苗绣纹样的数字化提取,提出了一种基于两阶段渐进采样视觉Transformer的边缘检测算法,分为全局和局部检测。在两个阶段都引入渐进式采样来定位重要区域,使提取的边缘集中于苗绣纹样主体部分,减少服饰背景等造成的干扰。通过使用多尺度通道注意力特征融合模块,将全局和局部检测的边缘进行加权融合,以获得更清晰的边缘。实验结果表明,该算法与其他算法相比,提取的苗绣纹样获得了更纤细的线条,且丢失的纹样形状线条较少,纹样整体效果与标签图最接近,效果最佳。

Traditional Miao embroidery involves depicting the lines of a basic pattern on paper,cutting the pattern paper tightly onto a cloth backing,and then completing the Miao embroidery using colored threads and a variety of embroidery stitches.Many precious patterns will fade away as the old Miao embroidery breaks down.The digitized extraction of Miao embroidery patterns is not only to reduce the graphic symbols of Miao embroidery into lines,and to reveal the form and function of Miao embroidery’s religionized writing,but also to use the extracted Miao embroidery patterns as digital resources for designers’secondary creation.Miao embroidery patterns depend on all kinds of Miao costumes,and manual extraction is greatly restricted.Thanks to the development of computer digital technology,the digitized extraction method can realize the rapid extraction of Miao embroidery patterns.The digitized collection and design reapplication of Miao embroidery patterns can rescue and protect many Miao embroidery patterns that are on the verge of disappearing.In this paper,for the problems in the process of digitized extraction of Miao embroidery patterns,an edge detection method based on progressive sampling(PS)two-stage visual Transformer is proposed to realize the shape extraction of Miao embroidery patterns.The model is based on visual Transformer and is divided into two stages.PS is introduced in both stages to localize important regions to mitigate the loss of structural information inherent in the simple tokensization scheme in visual Transformer.The extracted edges are made to converge to the main part of the Miao embroidery pattern.In the first stage,a global Transformer encoder is used to obtain the global context on coarse-grained patches.Then in the second stage,local Transformer encoders are used to mine local cues at fine-grained patches.Each Transformer encoder is followed by a bi-directional multi-level aggregation decoder for high resolution features.Finally,the globally and locally detected edges are fused by a multi-scale channel attention feature fusion module to obtain better Miao embroidery pattern extraction.In this paper,PS is introduced in both stages to localize the important regions,so that the extracted edges are focused on the main part of the Miao embroidery pattern,and the interference caused by the background of the dress,etc.is reduced.Clearer edges are obtained by weighted fusion of globally and locally detected edges by using the multi-scale channel attention feature fusion module.The experimental results show that the algorithm obtains slimmer lines in the extracted Miao embroidery patterns compared with other algorithms and loses fewer lines of the pattern shape,and the overall effect of the patterns is closest to the labeled image with the best results.The evaluation of this paper’s algorithm on the edge detection dataset of Miao embroidery patterns is among the best levels,with an ODS score of 0.848,an OIS score of 0.871,and an AP score of 0.910,an improvement of 0.6%,0.8%,and 0.9%compared to the currently higher-ranked EDTER in the three evaluation indexes of ODS,OIS,and AP,respectively.The digitized extraction and design reapplication of Miao embroidery patterns can rescue and protect many dying Miao embroidery patterns,and can also be widely noticed and valued by the society.The digitally extracted Miao embroidery patterns restore the initial form of Miao embroidery patterns,on which designers can boldly practice and create more novel and interesting works.However,the training of this algorithm model for the dataset of Miao embroidery patterns requires a large number of manually drawn labels in the early stage,which is a heavy workload.In the future,the lightweight processing of the algorithm model can be considered to make the extraction of Miao embroidery patterns more efficient and convenient,and it can also be easily applied to any type of textile pattern extraction.