Research On Miao Costume Pattern Segmentation By Full Convolutional Network Based On Attention Mechanism

Ethnic costumes are an important part of the ethnic culture,not only in terms of their variety and exquisite style,but also their incomparable artistic value.However,with the invasion of foreign cultures,it has become an indisputable fact that ethnic costumes are gradually declining.Hence,the efficient and accurate segmentation of ethnic costume patterns is of great importance to the protection,development and inheritance of ethnic costume culture.At the same time,the textures,colors and styles of ethnic costume patterns are analysed from a computer visual perspective,which in turn leads to the digital conservation of ethnic costume patterns,which has a certain role in promoting the development of ethnic costume culture.However,due to the following problems:(1)Miao costume patterns have the characteristics of bright colors,rich styles and complex textures;(2)the pooling and up-sampling operations in the full convolutional network result in the loss of some feature information in the image;(3)the full convolutional network does not capture image contextual information well,resulting in poor segmentation performance when the full convolutional network model is applied to segmentation of Miao costume patterns.Therefore,in this paper,aiming at the Miao costume pattern dataset,attention mechanism is integrated into the full convolutional network to improve the segmentation performance of the model.The main work and innovation of this paper are as follows:(1)A CBAM attention module-based segmentation model for Miao costume patterns is proposed to address the problems of large colour differences,style and texture diversity in Miao costume patterns,as well as the loss of some feature information in images due to pooling and upsampling operations in full convolutional networks,resulting in low efficiency and accuracy of image segmentation.The model enables the model to better correlate features of interest from the local level to the global level by extracting image features.At the same time,data enhancement strategy is adopted to increase training data to improve model generalization ability.Finally,the experimental results show that the model increases the Io U by 14.79%and 18.21% and the Dice coefficient by 11.03% and 13.95% compared to the U-Net and FCN models using only less than 1/2 of the training parameters.(2)Aiming at the insufficient ability of the full convolutional network to capture global context information,which leads to low segmentation performance,a Miao costume pattern segmentation model based on the dual attention mechanism and the full convolutional network is proposed.First of all,this paper uses full convolutional network architecture to build the backbone network.Secondly,through the effective combination of channel and spatial attention,a dual attention mechanism is constructed,and the dual attention mechanism is integrated between the convolutional layers,so that the network can better obtain the global context information of the image,focus on important features,inhibit unnecessary features,and thus effectively improve the performance of image semantic segmentation.Finally,the experimental results show that the model outperforms the full convolutional network in the five evaluation metrics of accuracy,intersection ratio,Dice coefficient,sensitivity and precision rate.In summary,this paper investigates and improves the full convolutional network from the above two aspects.Compared with representative methods in recent years,the performance of model segmentation is improved,and its robustness,subjective visual effects and objective evaluation metrics are significantly improved.The experimental results show that the model proposed in this paper can be used as a potential segmentation tool in the field of ethnic minority clothing patterns.