Research On Image Analysis And Retrieval Methods For Ethnic Minority Clothing

There are many kinds of minorities in our country.Minority clothing are colorful and their visual styles are different.The combination of clothing image parsing and retrieval with national costume culture plays an important role in realizing the digital protection of ethnic costume images and the inheritance of national culture.Because of the lack of semantic labels of minority clothing and complicated local features leading to unsatisfied accuracy of minority clothing parsing and retrieval,a clothing parsing method combining visual style and label constraints is proposed for various minority clothing images.The parsing results are retrieved through multi-tasking deep supervision hash.This method can effectively improve the accuracy of the minority clothing image parsing and retrieval.First,because of the lack of a complete minority clothing image dataset,we constructed a clothing image dataset containing 55 Chinese minorities.The dataset contains clothing images of some minority branches.The existing semantic labels are only used for general fashion clothing,and can't distinguish different minority clothing styles.According to the basic style structure,clothing area,accessories and different visual styles,we have defined the general semantic labels and ethnic semantic labels of minority clothing.At the same time,we set 4 sets of annotation pairs with a total of 8 annotation points.Secondly,considering the different visual styles of minority clothing,and the visual styles of minority clothing are special and diverse,we build a probability model that divides the visual styles of minority clothing.The probabilistic model analyzes the visual style of minority clothing images.Minority clothing images are divided into 7 styles,and semantic labels are further optimized according to the style.The method solves the problem of complicated semantic labels of minority clothing,and can improve the accuracy of subsequent parsing.However,because there are a large number of color patches and many local details such as texture patterns in minority clothing images,the accuracy of the parsing is low.We combine the optimized semantic labels and training images with labeled pairs,add a visual style to the deep fully convolutional neural network Seg Net to fuse local and global features.The visual style side branch network introduces attribute loss,style prediction loss and triple loss to perform preliminary parsing on the input image.The preliminary parsing result is further optimized by the constructed label constraint network to avoid mutual interference of the labels,and the optimized final parsing result is obtained.Finally,the single task model is difficult to learn the various styles of minority clothing.Therefore,we use a multi-task deep supervised hash algorithm to learn the different clothing styles obtained from the parsing results.The clothing features are mapped into binary hash codes.optimize the loss function.At last,through the optimization of the loss function and the similarity calculation,the image retrieval of minority clothing is realized.The experimental results show that the method can accurately realize the image retrieval of minority clothing.