Research And Implementation Of Semantic Segmentation System Based On Goal-driven Traditional National Costumes Patterns

With the rapid development of computer technology,the extraction of patterns in costume images,processing and digital storage are of great significance to the protection and research of traditional patterns.Therefore,this paper uses the current significant performance of deep learning technology to analyze the feature extraction and semantic segmentation modules of traditional costume images,improve the existing semantic segmentation model,and construct a supervised semantic segmentation model based on the attention mechanism.Semantic segmentation of traditional patterns,and a traditional costume semantic segmentation system based on the proposed model design.The main work of this paper includes:(1)Construct the data set of traditional nation costume Image.First according to the digital image acquisition metadata standard,by scanning and cameras to traditional costume image acquisition equipment such as digital collection,and the associated with the data set of preprocessing,and then according to the requirements of semantic segmentation tasks for acquisition of image pixel level annotations.Finally,the data set is divides into training set and test set according to the proportion of 7:3,which lays a data foundation for the subsequent practice.(2)Based on global context learning,a semantic segmentation model is constructed.Aiming at the pattern of traditional costume images that are easy to be confused,small objects are easy to lose,and edge details caused by unique textures,this paper analyzes the advantages of self-attention mechanism in obtaining long-distance context dependence,and adds the Expectation-Maximization attention module as the semantics of the model The segmentation module proposes a semantic segmentation model based on global context learning,and uses the idea of residuals to optimize the iterative process of the Expectation-Maximization attention mechanism and reduce the loss of details in the iteration.Experiments have proved that the improved Expectation-Maximization attention module can better establish the relationship between all positions of the feature map,and effectively improve the segmentation quality of traditional patterns.(3)Based on salient fusion learning,a semantic segmentation model Is constructed.Aiming at the characteristics of traditional costume with fine texture,complex colors,and large target scale changes,this paper studies the dynamic scaling strategy of the ELASTIC structure and the characteristics of the convolutional attention module to learn salient features.By introducing a set of parallel volumes in the feature extraction stage The integrated attention module,combined with the improved Expectation-Maximization attention module,proposes a semantic segmentation model based on saliency fusion learning.Experiments prove that the convolutional attention module effectively suppresses invalid features,makes the features of the object region more prominent,and effectively improves the ability of the improved Expectation-Maximization attention module to segment traditional patterns.(4)Designed and implemented a traditional costume pattern Image segmentation system.Based on the proposed semantic segmentation model,the system can segment traditional patterns from traditional costume with high quality.At the same time,it also provides the extended function of retraining,supports custom model training,and finally provides a pattern material database module,which is the pattern material.The expansion of the database provides technical support.