Research And Implementation Of Multi-Label Annotation For Traditional National Costumes Based On Dictionary Learning

Traditional national costume records and inherits a nation's cultural connotation,religious belief,customs and so on.For example,the 'dragon'in the traditional national costume represented 'power' and 'nobility' in the Ming and Qing dynasites,and was the symbol of imperia power.The'butterfly' represents 'happiness' and 'love'.With the progress of the times,the traditional national costume gradually withdrew from the stage of fashion,and its cultural connotation was also diluted.In order to better interpret,protect and inherit the traditional culture of the Chinese nation,it is inevitable to study the multi-label annotation of traditional national costumes.In this thesis,traditional national costumes are taken as the research object.In view of the particularity of traditional national costumes compared to the natural image and the limitation in performance of dictionary learning,two algorithms of quantization space construction with dictionary correlation mining and reconstruction coefficient enhancement are proposed to realize multi-label annotation of traditional costumes images.The feasibility and effectiveness of the two algorithms are verified by experiments on the multi-label data set of court dress images in Ming and Qing dynasties.The main contents of this thesis are as follows:(1)This thesis introduces the theory of dictionary learning for traditional classification.On the one hand,reconstruction error is used to classify;On the other hand,the discrimination ability of reconstruction coefficient is used to complete the classification.Combine these two approaches.(2)This thesis summarizes the multi-label annotation algorithm and the main evaluation indicators.The problem transformation method and algorithm adaptation method in multi-label annotation are introduced in detail.Label-based and sample-based evaluation methods are introduced.(3)For the traditional national costumes,the dictionary learning method is used to construct the quantitative space for different categories.The dragon pattern data set was used for experimental verification.Through the information of feature space,the correlation between categories is extracted and used in the calculation of dictionary correlation coefficient.The relationship between categories was optimized to improve the performance of multi-label annotation,and the multi-label annotation experiment based on reconstruction error was carried out on data set of court dress images in Ming and Qing dynasties to verify the effectiveness of the algorithm improvement.(4)Enhance the linear discrimination ability of reconstruction coefficient.By using SVM and adding the penalty item of error sample,the discrimination ability of reconstruction coefficient was improved.In the test,the linear discrimination ability of the reconstruction coefficient and the reconstruction error were combined to complete the multi-label annotation,which ultimately improved the effect of multi-label annotation.Experiments were completed on the data set to verify the effectiveness of the improved algorithm.(5)Matlab GUI was used to develop the traditional national costume multi-label annotating system.The system can be used to quickly carry out multi-label annotation,model training,testing,feature extraction and other tasks.Based on the reconstruction error and reconstruction coefficient,this thesis improves the dictionary learning algorithm.The task of traditional national costume image quantization space construction and multi-label annotation achieved good results.The results were successfully integrated into the interactive multi-label annotating system.