Research On Image Semantic Understanding Of Traditional National Costume Pattern Based On Deep Learning

Traditional national costume is the wisdom crystallization and artistic beauty of human labor civilization,which contains the long-standing cultural heritage and historical accumulation of the Chinese nation.Using deep learning technology to realize the semantic understanding of costume pattern image is conducive to the inheritance,development and innovation of traditional costume.This paper takes the traditional national costume pattern image as the research object,analyzes it from two aspects:the dominant information(ontology)and the deep semantic information(hidden meaning).And uses the deep learning technology to construct the image double-layer multi label annotation model from two different angles to realize the semantic interpretation of the pattern image,designs and realizes the costume pattern image semantics according to the needs Understand the system.The main work of this paper includes:(1)Construct the data set of traditional national costume pattern image.Firstly,the standard of costume digital collection metadata is designed,and then the books are digitized collected by scanner.Secondly,the images are labeled from the two levels of ontology and implicit meaning.Finally,the data set is divided into training set and test set according to the proportion of 8:2,which lays a solid data foundation for the subsequent practice.(2)Based on fusion learning,a double-layer multi label annotation model is constructed.The relationship between the ontology information and hidden meaning information of traditional costume pattern image is regarded as progressive.That is to say,the middle layer of convolutional neural network is used to extract the ontology layer features of pattern image,and the high layer is used to extract the implicit layer features of pattern image.At the same time,based on the dependency between ontology information and implicit information,an image double-layer multi-label annotation model based on fusion learning is proposed.In the traditional costume pattern image data set,the mAP of ontology annotation and implicit annotation are 0.88 and 0.82 respectively.In addition,compared with the variant structure of FL-DMAM model,the advantages of FL-DMAM model in this paper are fully proved.(3)Based on multi-task learning,an image double-layer multi-label annotation model is constructed.The relationship between the ontology information and the implicit information of the traditional national costume pattern image is regarded as a juxtaposition.Resnext-50 is used as the main network of shared features,and then a branch structure is constructed for each task by using the attention mechanism.A multi task learning based image double-layer multi-label annotation model is proposed,and add ELASTIC structure can further improve the performance of the model.The MTL-DMAM model is better than other similar models in the experiment of traditional costume pattern image data set.(4)Based on the above two image semantic understanding models,the traditional costume pattern image semantic understanding system is designed and implemented for the system business needs and the actual needs of users.It supports registration or login,pattern image semantic understanding,personal information management and other practical functions,and serves as a platform for users to independently complete the semantic understanding of clothing pattern image.