Research On Image Retrieval Technology Of Ethnic Costume Based On Improved High-Resolution Net

The traditional costumes of ethnic minorities in China have complex patterns and special composition,which poses a great challenge to the retrieval task of ethnic image.On the one hand,the performance of current graphic content based ethnic minority clothing image retrieval algorithms is limited.On the other hand,due to the rich colors and diverse styles of ethnic minority clothing images,clothing images contain rich semantic features.When using traditional retrieval algorithms for fashion clothing or daily clothing to perform ethnic minority clothing retrieval tasks,local semantic feature clues of clothing are often ignored.Ordinary models based on convolutional neural networks can extract deep semantic features from clothing images,and ordinary clothing image retrieval algorithms perform well in large public datasets.However,they ignore the fine grained features of various components and accessories in ethnic minority clothing images,and the detection effect is not ideal.Therefore,this paper proposes a minority clothing image retrieval algorithm based on the improved HRNet network,and implements a minority clothing image retrieval system based on the improved HRNet network.The main work is carried out around the following four points:(1)A database of ethnic minority clothing images has been established.2096 ethnic minority clothing images are divided into 50 categories by clothing style.The key points of all clothing images are labeled according to 7 key points,including headwear,left shoulder,right shoulder,left wrist,right wrist,left trouser leg,and right trouser leg.All labeling information(including border boxes,body masks,and key point annotations)is saved in the data set in the form of a file,and a national clothing image resource library containing annotations such as national clothing images,ethnic clothing categories,and key points is established.(2)Extract image features from the image database using feature descriptors Hog,Daisy,and an improved network based on the Vision Transformer model,respectively.Save all annotation information(including bounding boxes,body masks,and keypoint annotations)in the form of a file in a national clothing image resource library containing annotations such as ethnic clothing images,ethnic clothing categories,and keypoints.This article extracted the HOG and Daisy features of all dataset images and saved them.Then,the dataset image images were input into one set of improved Vision Transformers to extract another set of features.The Vision Transformer model was trained using the method of ethnic minority classification.After the model is trained,a set of output features are extracted from the last hidden layer and saved.These features will be used for image reordering during the retrieval process.(3)A minority clothing image retrieval algorithm based on improved HRNet has been proposed.The improved HRNet is used to identify clothing key points,and the improved HRNet is used as the backbone network of the entire network to extract features.The output of the backbone network is directly used as a set of feature vectors.Then input the clothing key point information and feature map into an improved Swin Transformer encoder for serialization coding,and obtain another set of feature vectors.Splice the two sets of outputs to obtain retrieval features.The Euclidean distance between retrieval features is used as the main basis for clothing image retrieval,supplemented by other features of the input image to reorder the retrieval results.Using the Top-k accuracy as an evaluation indicator on the ethnic minority clothing dataset,the results show that the proposed method performs better than traditional retrieval methods such as Deepfashion and DARN,with a Top-20 accuracy of up to 99.0% on the dataset.(4)The design and implementation of an ethnic minority clothing image retrieval system based on improved HRNet network.Users can retrieve similar images through this system by inputting a clothing image.The system uses Py Qt5 to achieve front-end visualization,integrating display and interaction functions,loading retrieval model weights,feature processing functions,key point detection functions,and image retrieval functions.The system achieves the effect of providing users with a user-friendly and interactive image retrieval system for ethnic clothing.