Research On The Pattern Segmentation Of Ethnic Minority Costumes Based On Attention Fusion U-Net Model

As an important part of the Chinese nation,ethnic minorities have contributed greatly to the prosperity of the Chinese culture.Among them,ethnic minority costumes are particularly eye-catching.As the most intuitive and vivid cultural carrier in their ethnic culture,ethnic minority costumes carry the unique culture and aesthetic values of ethnic minority for thousands of years,and highlight their enduring cultural values.However,due to the strong impact of modern popular clothing and foreign clothing,ethnic minority costumes inevitably began to decline.Therefore,protecting these ethnic minority costumes cultures that have been passed down for thousands of years has profound practical significance.Due to the large number of ethnic minorities in China and their own unique cultural characteristics,and because ethnic minority costumes are different due to their production process,style and regional differences,most of the existing research methods are based on traditional methods to segment the image of ethnic minority costumes,so as to achieve the purpose of protection and inheritance.In recent years,deep learning technology has made great progress and development,while the research content of combining the deep learning technology with the image segmentation of ethnic minority costumes is still very few.Therefore,according to the characteristics of specific ethnic minority costumes pattern,this paper proposes the corresponding pattern segmentation algorithm of ethnic minority costumes.(1)The composition of Miao costumes pattern is usually complex,and it also shows complexity in textile technology.Aiming at the problem of poor accuracy of Miao costumes pattern segmentation,this paper proposes a Miao costumes pattern segmentation algorithm based on the RSKP-UNet model by integrating attention.Based on the basic framework of the U-Net model,Residual and SKNet attention with attention mechanism are embedded in the encoder of the U-Net model,which can fully capture the image context information while extracting the image feature information in the deep structure,and then integrate ParNet into the decoder to improve the generalization ability of the model.In addition,due to the problem of class imbalance generally existing in Miao costumes pattern,this paper introduces the Lovász-hinge loss function to optimize the network,so as to improve the segmentation accuracy of Miao costumes patterns.The experimental results show that the RSKP-UNet model proposed in this paper is better than the other four contrast segmentation models in terms of indicators.Compared with the benchmark model U-Net,the Dice coefficient is increased by 6.98%,IoU is increased by 11.07%,accuracy rate is increased by 2.89%,recall rate is increased by 6.75%,and accuracy is increased by 3.92%.Therefore,the RSKP-UNet model proposed in this paper can better solve the problem of Miao costumes pattern segmentation,and the segmentation accuracy is high.(2)The composition and color of Yi costumes pattern is extremely complex,and the styles are diverse.In response to the complex characteristics of Yi costumes pattern,this paper proposes a lightweight AES-UNet model for Yi costumes pattern segmentation by integrating multi-scale feature extraction and attention.AES-UNet model is improved on the U-Net model.First,it improves the basic framework of the U-Net model,and then embeds ASPP and ECA attention in its encoder to improve the feature extraction ability of the model and the importance of focusing on channel information.Then,it embeds SA attention in its decoder to highlight the semantic features of Yi costumes pattern and achieve attention operation on Yi costumes pattern.On the basis of Lovász-hine loss function,this paper proposes an improved loss function for Yi costumes pattern segmentation by combining BCE loss function and Lovász-hine loss function,so as to improve the accuracy of Yi costumes pattern segmentation.The experimental results show that the AES UNet segmentation model proposed in this paper is better than the four segmentation models based on depth learning in Yi costume pattern segmentation.Compared with the benchmark model U-Net,the AES UNet model proposed in this paper has improved 4.06%,4.61%,3.61%,2.05% and 3.18% in Dice coefficient,IoU,accuracy,recall and accuracy,respectively.Therefore,AES-UNet segmentation model proposed in this paper can better complete Yi costumes pattern segmentation.