Research On Key Technologies Of Fine-grained Image Retrieval For Ethnic Costumes

With the development of e-commerce,the potential value of the clothing market is gradually discovered.A series of retrieval tasks have appeared in the field of computer vision such as clothing retrieval,clothing identification and clothing recommendation.Our country has a large number of ethnic minorities.There are a variety of national costumes,complex costume structures,rich semantic attributes,and national characteristics.If we can combine costume image retrieval with national costume culture,it is great significance for national costumes and the inheritance of national minority cultures to the digital protection.National clothing with different clothing styles,accessories and patterns leads to low accuracy of fine-grained retrieval of national clothing images.To address the above problem,a global-local feature extraction model for fine-grained clothing image retrieval is proposed to extract the global and local features of the image.And the preliminary retrieval results are obtained by applying a similarity measure to the global features.Then the similarity for re-ranking was measured by the local features of the top-50 returned by the preliminary retrieval results and the query image.The final retrieval results were output by the result of re-ranking.The method can effectively improve the accuracy of national clothing image retrieval.Firstly,national clothes have complex costume structures,rich semantic attributes,national characteristics and more fine-grained semantic attributes.The dataset with labeled is also lack.Firstly,based on the existing national clothing image dataset,this paper filters and expands the dataset for the experimental requirements of this article.According to the area where the fine-grained attribute of each national clothing image is located,the fine-grained semantic attribute of the national clothing image is customized.Secondly,the labeled images were used to train the detection model.The detection results were segmented,and different feature extraction branches were input according to the classification results.And different feature extraction branches define different loss functions to extract global and local features from the input image.It solved the problem and this model can effectively extract the global and local features of the national clothing image.Secondly,because there are a lot of fine-grained semantic attributes,the accuracy of retrieval is low.We combine fusion features and re-ranking method to retrieval national clothing image.The preliminary retrieval results are obtained by applying a similarity measure to the global features.Then the similarity for re-ranking was measured by the local features of the top-50 returned by the preliminary retrieval results and the query image..The final retrieval results were output by the result of re-ranking.Experimental results on the national clothing image dataset showed that the method we proposed can improve the accuracy of national clothing image fine-grained retrieval.Finally,based on the two methods above,we combine the application scenarios and user needs of fine-grained retrieval of national clothing images,a national images retrieval system framework is designed,and the system prototype design is implemented.The system has a simple interface,complete functions,and accurate retrieval results,which can better reflect the effectiveness and practicability of the method in this article.