Research On Deep Learning Of Commodity Detection In Unmanned Retail Environment

Online e-commerce has developed rapidly under the popularity of the Internet and logistics.However,after a period of rapid development,traditional e-commerce has been facing problems such as sluggish online growth in recent years and has entered a bottleneck period.Therefore,it has to combine online e-commerce with offline sales under the background of ”new retail” proposed by Alibaba to seek breakthroughs.Smart container is the promising development direction of major e-commerce companies.After trying to use non-visual solutions,with the significant breakthrough of deep learning and convolutional neural network in the field of computer vision,deep neural network-based visual solutions have become the research focus of smart container solutions.This paper has done some research on the static vision solution of smart containers.The main work is as follows:(1)Firstly,the most commonly used static vision solution in the market – the method of using the object detection model based on the deep neural network to conduct full-supervised commodity detection on the container –has been tested and analyzed.In this paper,Faster R-CNN,a two-stage object detection algorithm which does best in the performance of object detection,as well as YOLOv3,a single-stage target detection algorithm,were used to conduct experiments respectively.The experiments were based on the data sets collected by real container scenes,and the results were evaluated by the specific method for the ac-curacy of smart container commodity detection proposed by the author.It is found in the experiment that although this method can complete highprecision automatic inventory and settlement by completely detecting the goods in the container at two moments before and after consumption,it has problems of high cost for labeling data,updating models,as well as deployment.(2)In view of the problems found in the above experiments,this paper proposes a novel pairwise image difference detection algorithm called DiffNet,which combines the characteristics of similarity learning of siamese network with the object detection algorithm.It detects the difference between the image-pairs in the container before and after the consumption directly and finds the commodity which has different location.Then,a commodity classification model would be used to identify the classes of different commodities.In this way,the smart container completes automatic settlement.The static solution of smart container with DiffNet as the core only labels the bounding boxes of different goods between a pair of input images,so the labeling cost is much lower than the full supervision object detection.Only the classification model needs to be updated when there are new products,so the updating cost is low.The deployment of DiffNet and commodity classification model has lower hardware requirements and lower deployment cost than that of full-surveillance target detection,and the accuracy of this algorithm is similar to the full-surveillance object detection method,so that it has strong practical value.In summary,the paper firstly implements the mainstream solution in computer vision for the commodity detection in the unmanned retail environment.Aiming at the problem of high labeling cost in its application,a novel deep neural-based algorithm for deep differential recognition was put forward,which achieved more ideal effect in application.