一种融合注意力机制的苗族服饰图案分割方法  

A Miao clothing pattern segmentation method based on attention mechanism

作　　者：万林江 黄成泉 张博源 王琴 周丽华 WAN Linjiang;HUANG Chengquan;ZHANG Boyuan;WANG Qin;ZHOU Lihua(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang,Guizhou 550025,China;Engineering Training Center,Guizhou Minzu University,Guiyang,Guizhou 550025,China)

机构地区：[1]贵州民族大学数据科学与信息工程学院,贵州贵阳550025 [2]贵州民族大学工程技术人才实践训练中心,贵州贵阳550025

出　　处：《毛纺科技》2022年第12期95-101,共7页Wool Textile Journal

基　　金：国家自然科学基金项目(62062024);贵州省省级科技计划项目(黔科合基础-ZK[2021]一般342)。

摘　　要：针对苗族服饰因缺少像素级标注的数据库、元素多元化、纹饰图案不规则等引起的目标区域特征提取难度大的问题,提出了一种融合了注意力机制的SegNet分割模型(SE-SegNet)。在改进的SegNet模型中融入通道注意力SE模块,关注更多的细节特征,旨在于加强对目标特征的提取,实现苗服饰图案的自动分割。实验结果表明,该模型在苗族服饰数据集中,像素准确率为92.69%,交并比值为85.27%,相似系数为92.05%。与其他模型相比,该模型分割结果更精细,在苗族服饰图案分割的效果得到显著提升。服饰图案分割效果的提升对苗族服饰文化的保护和发展具有重要意义。Aiming at the difficulty of feature extraction in target region caused by lack of pixel-level annotated database,mul-tiple elements and irregular decorative patterns of Miao clothing,a SegNet segmentation model(SE-SegNet)integrating attention mechanism was proposed in this paper.Specifically,the channel attention SE module was integrated into the improved SegNet model to focus on more detailed features,in order to strengthen the extraction of target features and realize the automatic segmentation of Miao clothing patterns.The experimental results show that the pixel accuracy of the proposed model is 92.69%,the intersection ratio is 85.27%,and the Dice similarity coefficient is 92.05%in the dataset of Miao nationality clothing.Compared with other models,the segmentation results of this model are finer,and the effect of the segmentation of Miao clothing patterns has been significantly im-proved.The improvement of costume pattern segmentation effect is of great significance to the protection and development of Miao clothing culture.

关 键 词：苗族服饰 SegNet模型 SE-SegNet SE模块 图案分割 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]