E-commerce Image Object Recognition Based On Transfer Learning

Today,as the Internet becomes more and more involved in personal life,e-commerce shopping has become an indispensable part.Data companies will obtain e-commerce images from social networks and identify them.During the process of identification,it was discovered that products that were originally accumulated by merchants,when they were updated,found that the image data of new packaging on social networks was very small.To reach a certain time,there was enough data to identify new packaging products.This creates the drawback that data is not timely,and the precious data with annotations that data companies have accumulated at the beginning is wasted.How to use the large amount of annotated data accumulated at the beginning to label a small amount of data in new areas to identify a large number of new areas of untagged data? Based on the above requirements,this paper proposes an e-commerce image recognition based on migration learning.Migration learning is a method that is more effective as a data processing tool to process massive amounts of complex data.This article starts from the necessity of transfer learning and systematically studies transfer learning from the three levels of theory,method and practice.Specifically,based on the deep neural network,this paper studies the application and effect of migration learning in e-commerce image object recognition to improve the accuracy of data classification tasks in the target domain and the selection of deep neural networks under the migration learning model.The problem is to compare the classification effect of sample data in different spaces.The work of this article starts from the following two aspects:(1)Analyze existing migration learning methods,summarize existing migration learning methods,and provide theoretical support for migration learning methods and applications.Detailed summary and classification of existing migration learning methods;in-depth study of traditional deep learning techniques such as shared weights,model pre-training and semi-supervised learning,finding their relevance to transfer learning,and using existing deep learning techniques to guide migrationLearn algorithm design.(2)Applying supervised learning techniques in migration learning,establishing a migration learning approach based on deep neural network architecture,enriching sample-based migration learning theories and methods.Through the recognition performance of the model to compare the migration learning effect of data in different fields.In the model selection,the validity of the selection method is verified from the empirical formula,and the migration performance of different networks is compared longitudinally,and the reasons for the success and failure of the migration are analyzed,thereby providing guidance for migrating knowledge from data to a new scenario.Experiments on the e-commerce image dataset demonstrate the feasibility of the method.