Research On Efficient Retrieval Methods For Large-scale Commodity Images

With the advent of the era of big data and the popularity of mobile Internet,the ecommerce industry has developed rapidly.Product image retrieval,as an important technical support for e-commerce,has received more attention.How to improve the speed and accuracy of product image retrieval to meet the needs of users needs to be solved urgently.Traditional image retrieval systems are mainly text-based image retrieval(TBIR).However,text data has great limitations,it is difficult to accurately describe image information,and it depends on whether the search keywords entered by the user are accurate.In contrast,Content-based Image Retrieval(CBIR)can overcome these problems,because the image itself can provide more information.In CBIR,the extraction of image features is the most critical step,which directly affects the performance of the entire retrieval system.This article deeply researches the related research of traditional methods and deep learning methods in the field of commodity image retrieval,and summarizes the research status of large-scale commodity image retrieval.Compared with deep learning methods,traditional methods have limitations such as difficulty in describing global information,high computational complexity,difficulty in describing specific objects,and inability to consider the spatial location of objects.Since Alex Net won the championship in the ILSVRC2012 competition,deep learning has attracted much attention and quickly dominated.This paper conducts research based on convolutional neural networks.The research content is a retrieval method for large-scale commodity images.The main technical difficulties of large-scale product image retrieval are as follows: 1)Huge data scale.2)The amount of data varies greatly between different types.3)The dimensions of the products being photographed vary widely.4)Image sizes vary.Aiming at the difficulty of large-scale commodity image retrieval,this paper designs a method based on convolutional neural network to achieve fast and accurate retrieval.A SHN(SPP-HASH-NET)model combining pyramid pooling strategy and hash learning is proposed as the feature extraction part of the article image retrieval method in this paper.In order to improve the robustness of the model to the deformation of commodity images,a pyramid pooling strategy is adopted to achieve multi-scale feature fusion.In order to achieve efficient product image retrieval,generate a hash code to calculate the similarity.In order to learn a more distinctively hash codes,a deep highly interrelated loss function is designed to constrain the hash codes.The experimental results prove that the method can retain the complete original image information,solve the negative impact brought by the change of image scale,and realize fast product image retrieval through hash coding.In the commodity image retrieval experiment,the training time of the model is shorter than other models,and the mean average precision of the model reaches 91.52%,and the retrieval performance is better than the current mainstream methods.