Research On Clothing Image Classification And Retrieval Based On Deep Learning

With the rapid development of Internet information technology and the advent of the era of clothing e-commerce,the amount of online clothing image data has increased dramatically,and the demand for accurate classification and efficient retrieval of clothing images has become increasingly urgent.Faced with the cumbersomeness and subjectivity of artificial semantic labeling in text-based image retrieval,the "semantic gap" and the one-sidedness of feature description in traditional image feature extraction methods based on content,this paper is based on the deep learning method for clothing images.Classification and retrieval were studied.With the powerful image feature extraction ability of convolutional neural network,it breaks through the limitations of traditional methods,and achieves accurate classification and efficient retrieval of clothing images.The research has social commercial value and academic significance.The main research contents of this paper are as follows:1)A large-scale clothing image data set with 30 categories and 500,000 orders of magnitude was constructed.Comprehensive images of online public clothing image datasets such as DeepFashion,FashionAI,and ModaNet,a large number of pictures accumulated by the ** Provincial Apparel Customization Collaborative Innovation Center project,and clothing images crawled by various e-commerce websites,select appropriate images for research tasks and Classified,get 30 clothing category labels.After the basic pre-processing such as image segmentation according to the border labeling,the clothing image data set of a large-scale attached clothing category label is organized.2)Using the idea of first classification and intra-class retrieval,a deep learning model of clothing image classification retrieval based on improved local sensitive hash algorithm is proposed.It mainly includes three modules: ResNet101 classification module,hash retrieval module,and classification retrieval module:(1)ResNet101 classification module: ResNet101 neural network is used to extract the clothing image features,and the border frame regression is used to extract the clothing subject in combination with the Faster-RCNN network.After removing the interference factors such as background and multi-subject,the clothing image is classified by Softmax classifier.A clothing image classification model for deep learning.(2)Hash retrieval module: Combining the feature extraction in the clothing image classification model,introducing the hash layer after the model,using the loss function of the Softmax classifier to propagate the hash function to perform binary hash coding,using the combined Hamming distance and The sorting result obtained by the similarity measure of Euclidean distance is used as the retrieval result,and a clothing image retrieval model based on deep learning is constructed.(3)Classification retrieval module: Combine the classification model and the retrieval model with the idea of first classification and intra-class retrieval.Firstly,the clothing image data set is batch-classified and the hash index library is constructed.For a retrieval operation,the image to be retrieved is introduced into the model first.The index library under the classification and reaccess category is used to measure the similarity,and the similar images in this category are obtained,and the clothing image classification retrieval model of the first classification and the intra-class retrieval is constructed.3)Analyze the experimental results and design a comparative experiment to verify the performance of the model.The experimental results show that the model has good classification accuracy,retrieval performance and stability.The introduction of the hash layer accelerates the retrieval process.It has a good retrieval effect based on the similarity measure combining Hamming distance and Euclidean distance.The dataset after image segmentation has higher accuracy to the model.Compared with AlexNet,VGG16,GoogleNet,ResNet101 has better feature extraction effect.The idea of classifying and re-classifying the model has modified the clothing category.It has a 17.16% improvement in the retrieval speed and achieves the goal of the clothing field.Sexuality has important research value.