Research On Few-Shot Retail Product Recognition System

With the development of economy,people's consumption level is increasing,all kinds of physical products circulating in the market have greatly improved people's quality of life,but the wide variety of retail products have also brought great difficulties to people's choices.How to focus on the use of information technology,transform the circulation and sales methods of retail product,reshape the industry ecology,and integrate online and offline experiences are the problems that need to be solved in recent years.This paper uses online and offline supermarkets as the application background.The intelligent supermarket product recognition system constructed achieves the goal of partially integrating retail product information,bringing users a more convenient query method.This has the positive significance to the circulation and promotion of product.Generally,product recognition systems combine computer vision and information retrieval technology.According to the application field,it can be divided into two categories: product settlement scenarios and similar product search scenarios.This paper studies the application direction of similar product search,and focuses on improving related intelligent algorithms.Few-shot retail product recognition system constructed in this paper is mainly composed of four parts: product feature extraction based on convolutional neural network,product database,product retrieval model and answer push.The key technologies involved include data set image preprocessing,Convolutional neural network,Few-shot learning.Among them,the product retrieval is the core of the system,the main function is to calculate the similarity between the product image entered by the user and the database product information,and how to push the most similar class information to the user,and how to maintain a small number of new class samples.Excellent classification accuracy is the focus of this paper.Therefore,the main work of this paper is to propose a new metric-based Few-shot learning method for supermarket product classification models,and to improve its convolutional neural network,and at the same time complete a small sample-based supermarket product class detection system construction and performance testing.The metric-based Few-shot learning model designed in this paper aims at the problem that the traditional feature connection semantic representation is not sufficient,innovatively proposes a feature connection algorithm based on deep local descriptors,and combines the convolutional neural network for metric learning.Through the comparative experiments of 5-way 1-shot and 5-way 5-shot training methods,the optimal classification accuracy on the RPC retail commodity data set is achieved.In response to the overfitting problem of convolutional neural networks,this paper uses fine-tuned ResNet34,ResNet50,Inception-v3 models to improve the network width and depth of the feature extraction module and metric learning module,and selects the best through combined comparison experiments.Network structure,and image enhancement to the training data set.The improved Few-shot learning classification model further exerts the measurement function of the convolutional neural network,and the classification accuracy is improved again,reaching 99.10%.In the system construction,the above model is split into two parts,which are used as the system's product feature extraction model and similarity calculation model.Finally,through the system performance test experiment,the feasibility of the improved algorithm in this paper is verified.