Application Of Furniture Image Retrieval Based On Convolutional Neural Network

With the arrival of the big data era,the number of furniture images on the Intern et has gradually increased.If the semantic attributes of the image are manually annotated and searched by this semantic feature,it will take a lot of time and labor costs when facing a large amount of furniture image data.Because of the subjectivity of people,neither the annotator nor a person who wants to search for an image can accurately and consistently describe all information of the image,which also brings corresponding challenges to the search.Therefore,finding a highly efficient and convenien t furniture image retrieval method is a significant issue.In recent years,image features extracted through CNN can not only effectively represent images,but can also be successfully applied to many tasks such as image classification,image retrieval,an d so on,which makes experts and scholars more and more research on deep learning.This paper studies the furniture retrieval technology based on the convolutional neural network and completes the following work:1.Faced two problems in which the sample repetition rate is high,and many categories contain irrelevant images,we proposed two algorithms that can remove duplicates between classes and unrelated samples in the class.The Embedding represented by the corresponding image is extracted by the deep conv olutional neural network,and the distance between Embeddings is calculated in the European space,so that the duplicate and irrelevant images in furniture database can rapidly and accurately removed.We evaluated the performance of the deduplication and d e-correlated algorithms by using the two criteria “time” and “accuracy”.We separately showed some samples after deduplication and removing irrelevance samples.After the operation of different data preprocessing,the Recall is tested based on Goog LeNetplu s model,which proves the effectiveness of the two screening algorithms.The two methods overcome the complex problem in the traditional screening algorithm.2.For the large amount of furniture image features and categories,we improved the GoogLe Net network model by adding two fully connected layers in its structure,expanded the capacity of the network model,and named it Goog Le Netplus.To effectively construct sample pairs and improve the retrieval training efficiency,the loss function of Lifted Structured Feature Embedding(LSFE)was chosen as the retrieval model.The effects of four different models,Siamese Network,Triplet Network,GoogLe Net,and GoogLe Netplus based on LSFE on furniture image retrieval are analyzed and compared.Experiments show that the improved model Goog Le Netplus is superior to the other three models in retrieval ability.The analysis explored some factors influencing the accuracy of furniture image retrieval and the loss function convergence of the four models during training.We presented some successful and failed query instances and analyzed the corresponding reasons.Finally,based on the parameters of the trained convolutional neural network model,a mobile image retrieval client prototype for furniture is built.