Research And Application Of Clothing Image Retrieval Based On ConvNeXt Network

With the rapid development of the internet and e-commerce,an increasing number of consumers choose to shop online.This trend is particularly evident in the e-clothing industry,where the volume of online transactions has continued to grow year by year.As a result,fashion image data on the web has exploded,and retrieving fashion images that users find interesting has become a challenging problem.Traditional clothing image retrieval methods that rely on labels are unsatisfactory because of time-consuming manual annotation,subjective labeling,incomplete and inaccurate tag information.Clothing retrieval technology based on traditional visual content features also face difficulties in feature selection and semantic gaps,resulting in bottlenecks.Considering the excellent performance of deep learning in the field of image recognition,this paper proposes a clothing image retrieval method based on convolutional neural networks.This article presents a clothing image feature extraction model based on the ConvNeXt network.The ConvNeXt network is built on the foundation of Res Net and incorporates the ideas of the Transformer network,which has achieved satisfactory results in image classification and recognition tasks,making it the basis for the feature extraction of clothing images in this article.Considering that solely utilizing the global branch will have certain limitations,this article adds a local branch and jointly trains the global and local branches during the training process.Finally,this article develops a clothing image retrieval system based on the proposed model.The main contents of this article are summarized as follows:(1)This article introduces a novel clothing image retrieval algorithm based on the ConvNeXt network.The ConvNeXt network has a stronger ability to focus on important areas in clothing images and takes into account the details in clothing pictures.To further enhance the focus on the key areas of clothing images,a GCT-ConvNeXt-Block module incorporating Gated Channel Transformation(GCT)channel attention mechanism is designed,which further improves the model’s accuracy.The article employs cross-entropy loss functions with temperature scaling and label smoothing,as well as a hard negative triplet loss function to enable the model to identify small differences in clothing samples.To validate the effectiveness of the algorithm,the model is tested on the Deep Fashion clothing dataset and compared with other models,such as Res Net,SE-Res Net,Res Ne Xt,and Swin-Transformer,in clothing image retrieval experiments.After detailed experimental comparative analysis,the effectiveness and advancement of the model in clothing image retrieval have been verified.(2)The ConvNeXt network has shown strong performance in capturing image details and specific relationships,but it has limitations when dealing with complex background,different shooting poses,or poorly shot clothing image data,as the information extracted by a single global branch network is insufficient.To further improve the retrieval performance,we adopt a method of training both the joint global branches and local branches together,and use global features for similarity calculation.The local branches use the shortest path among their local features as the distance,which significantly reduces the distance between the same style of clothing when there is a large pose variation.This enables the model to better focus on the overall style and detail information in the image and achieve high-precision retrieval.Experimental verification shows that the addition of local branches leads to better retrieval performance.(3)Based on the model proposed in this article,a clothing image retrieval system was developed using a B/S three-tier architecture,which achieved multiple functionalities,such as clothing image retrieval,user history image retrieval information query,user management,and user login and registration,etc.