融合MultiDeepPPL的苗族服饰分割研究

Research on the segmentation of Miao costumes integrating MultiDeepPPL

针对基于深度学习的苗族服饰图像语义分割模型存在提取过程中服饰掩码拟合质量不高的问题,文章提出一种基于渐进式注意力学习的多尺度深度学习金字塔模型(Multi-scale Deep learning pyramidal network of progressive attentional learning,MultiDeepPPL)分割苗族服饰。首先,设计了一种密集跨级连接网络,充分利用多尺度方式提取特征的特性,融合不同尺度特征;然后,嵌入了一种渐进式注意力学习金字塔结构,从不同的特征图中迁移相似性与跨尺度相似性,并采用空域注意力和3D卷积对前述特征进行融合。实验结果表明,所提模型在苗族服饰数据集上平均交并比(Mean Intersection over Union,MIoU)达到0.873,类别平均像素准确率(Pixel Accuracy,MPA)达到0.943,Dice相似系数达到0.912,召回率(Recall)达到0.8951。上述评估指标结果表明,文章所提方法明显优于当前其他语义分割算法,为少数民族文化的研究提供了一种有效可行的方法。

There are a wide variety of costume patterns in China,which are the carriers of national culture.Ethnic minorities reflect their religious culture and totem culture in costume patterns and architectural decoration,not only for self-beatification but also for conveying special cultural significance.Ethnic minorities have experienced regional migration and cultural integration in the process of evolution,resulting in great changes and differences in costume patterns in different periods and regions,as well as complex structures and various categories of the same-style ethnic costumes with different details and attributes.Miao costumes,characterized by complex structure,bright colors,diverse textures and patterns,various styles and rich ornaments,were selected as the example to conduct further studies.Since it is difficult to segment the local details of Miao costumes and distinguish the high-level visual semantic attributes from low-level feature semantic attributes using current semantic segmentation models based on deep learning,a new model MultiDeepPPL was proposed in this paper to improve the above-mentioned shortcomings of deep learning,enhance the efficiency of Miao costume pattern segmentation and provide a new perspective for minority culture research and dissemination.To address the problem of low quality of clothing mask fitting in the extraction process of semantic segmentation model of Miao costumes based on deep learning,a new multi-scale deep learning pyramidal network of progressive attentional learning(MultiDeepPPL)was proposed for the segmentation of Miao costumes.Firstly,a dense cross-level connection network was designed adopting a two-level structure.In the first-level structure,the low-scale and medium-scale feature images were fused in the encoder;in the second-level structure,the fused results were fused as a whole with high-scale feature images.Then,a pyramid structure of progressive attentional learning was embedded,which could input a pair of cross-scale feature layers of the same size into the pyramid structure to progressively learn self-similarity and cross-scale feature information.After the feature information was extracted,the features were assigned with pixel-level weight,and all the feature information was aggregated by 3D convolution.The model can fully extract the feature information of Miao costumes of different scales,strengthen its feature extraction ability,learn self-similarity and cross-scale similarity,and aggregate the features.The experimental results have shown that the Mean Intersection over Union(MIoU)and Mean Pixel Accuracy(MPA)of the proposed model reached 0.873 and 0.943 on the Miao costumes data set,respectively.Dicesimilarity coefficient reached 0.912,and Recall reached 0.8951.The results of the above evaluation indicators show that the proposed method is obviously superior to other existing semantic segmentation algorithms,and it provides an effective and feasible method for the study of ethnic culture.The MultiDeepPPL model proposed in this paper has achieved good segmentation effects in the segmentation of Miao costumes,indicating that this model can help researchers automatically and accurately segment Miao costumes patterns.However,other ethnic minority datasets(such as Zhuang costumes and Bai costumes,etc.)are relatively insufficient,resulting in low model training fitting degree and poor segmentation accuracy.It is planned to train an efficient model based on unsupervised object segmentation with a small amount of data in the subsequent research.