Design And Implementation Of A Product Identification System For Smart Retail

With the rapid development of China's economy,the income level of residents continues to increase,and the variety of retailers is extremely rich.Although ecommerce has brought great convenience to residents' consumption,offline retail sales(such as department stores,supermarkets,convenience markets,convenience stores,etc.)are still the most important consumption places for Chinese residents.However,in the offline retail scene,problems such as low settlement efficiency,high labor cost,and poor settlement experience often occur,especially in areas with high consumption peaks and high density of residents.Although self-service scanning code settlement technology has been widely used,there are still problems such as complicated operation and inefficiency.Therefore,designing and developing a computer vision-based automatic recognition and settlement system for bulk products has important research and application value.In order to solve the above problems,this paper designs a series of product detection and recognition models for offline retail scenes based on deep learning technology and implements a batch product automatic recognition and settlement system.Firstly,the system captures images of multiple products placed on the shelf through a binocular RGB camera.Secondly,uses the product detection module to locate the position of the product in the image.Thirdly,use the fine-grained product identification module based on metric learning to construct the product feature database.Finally,recognize the product category by feature retrieval and matching.When the product category is obtained,the settlement is completed.The main contributions of this paper are as follows:Firstly,in order to accurately and quickly locate the products in the image,we designed a product detection module which based on PVANet.The module adopts an efficient network structure,which can improve the detection speed while maintaining the detection accuracy.In addition,the Inception module is introduced to learn the effective product features and the C.Re LU structure is introduced to improve the detection speed.By training and testing the model on the product image dataset collected in the real scene,the precision and recall of our detection model was 99.8% and 99.4% respectively and the detection speed was 14 ms/image.Secondly,in order to obtain the categories of products in the image accurately,we design a product recognition module which based on HDC network.The module adopts a cascade network structure which can accelerate the network training speed and enhance the learning ability of the network to the difficult samples.In order to improve the recognition performance of network,we use A-softmax loss function to train the network.By training and testing the model on the product image dataset collected in the real scene,the mean?recall of our model was 99.52%.Finally,in order to train the algorithm model and test the overall performance of the system,we collected and labeled a large-scale product dataset in the real retail scenario.The categories of products in dataset is 500 and the number of pictures is about 100,000 pictures.The bounding box and category of the product are marked on each image.The overall system performance reached 98.21%,and the settlement efficiency was 0.5 seconds/transaction.