Research On Grocery Product Detection And Recognition Technology By Template Matching And One-shot Deep Learning

Supermarket is one of the most important shopping places in modern society.With the rapid development of the retail industry in recent years and the impact of the "New Retail" concept,traditional offline supermarkets are gradually becoming ever more digitalized,there even have been unmanned supermarkets in some cities.An important part of digital transformation is shelf management.As the carrier of products,shelves are the core of supermarkets.From customary manual manager's manual shelf management to automated intelligent shelf management,this change can significantly improve management efficiency and customer experience.Due to the demand of obtaining the types and quantities of products on the shelves intelligently and quickly,this paper studies the technology of product detection of the shelf images,then designed and implemented a product positioning and recognition method which combines template matching and one-shot deep learning.This method contains two phases:product positioning and product classification.By recurring pattern algorithm,product template images and shelf images are matched by feature points,and all the single product candidate regions in the shelf images are found according to the matching relationship.Using deep learning classification networks VGG and ResNet to classify the images of single product candidate region,then the classification results of the various region are combined through post-processing to generate the product detection results of the entire shelf image.In order to increase the influence of the detail region from a product image on the classification results of classification network,the attention maps of the input images are added to the structure of the classification network,and the effect of the attention map is verified through comparative experiments.The entire product positioning and recognition process in this paper was ran on a public shelf product dataset that achieved 92.71%precision and 88.12%recall on 73 shelf images which contain 183 categories of products.