Research On Fast Recognition Algorithm Of Container Goods Based On Deep Learning And Its Application

With the rapid growth of the national economy and the continuous improvement of residents'living standards,people are paying more and more attention to the experience and feelings in the consumption process.Based on advanced technologies such as Artificial Intelligence,Big Data and Mobile Payment,the "Smart+ New Retail" model that integrates online and offline has emerged at the historic moment and is gradually becoming the key to competition in the new domestic retail industry.Intelligent vending cabinets based on Computer Vision have therefore become a hot spot in the industry.In order to bring users a convenient,fast and comfortable shopping experience,intelligent vending containers must consider the issues of product identification accuracy and identification speed.This is both the focus and the difficulty of the research on container product intelligent identification algorithms.In response to this problem,this paper analyzes and compares network frameworks such as Faster R-CNN and YOLOv3,and finally selects YOLOv3 with a relatively balanced recognition speed and accuracy,and more stable and reliable performance as the main framework.Under this framework,we first designed single-door cabinets and double-door cabinets with different placement structures for different sales scenarios,and designed data set construction methods and training strategies for these two different imaging environments.In order to further improve the accuracy of recognition,by analyzing the particularity of the target of the container goods,the recognition network is improved in a targeted manner:1)Aiming at the characteristics of generally small target size,a multi-scale detection improvement strategy is proposed.By integrating more shallow feature information,the feature pyramid's representation ability is enhanced,the detection accuracy of small targets is improved,and the missed detection rate is reduced;2)According to the feature that the target size does not change much,the size of the real labeling frame is re-stated through the clustering algorithm,so that the a priori frame is more in line with the target size of the commodity than the original preset frame,and the detection accuracy of the target commodity on the network is improved;3)Aiming at the problem that the MSE loss function has a small tolerance for the change of the commodity target scale,the GIOU loss function is used as the loss evaluation standard for target frame coordinate regression to further improve the detection accuracy.In order to further improve the recognition speed of smart vending products,this article also innovatively proposes a mosaic image-based product rapid recognition model and integrates it into the improved YOLOv3 framework.As we all know,the current target recognition algorithms are mostly using full-color images for recognition and classification,and full-color images are obtained by demosaicing the original Bayer format images captured by the camera.In order to make the image have a better visual effect,it usually undergoes interpolation,denoising,enhancement and other processing processes.These operations will cause a certain loss to the original information in the mosaic image,and will add too much redundant information.The innovation of the algorithm in this paper is to skip the demosaicing process and directly send the original image into the deep neural network for end-to-end recognition and classification.This not only avoids complex image restoration algorithms and reduces computational burden,but also enables faster acquisition of image classification and identification information.Based on the above reasons,this paper designs a preprocessing network based on hole convolution,phase convolution and sub-pixel convolution on the mosaic image data set collected in the intelligent vending cabinet scene.And through comparative experiments,it is verified that the strategy has higher detection accuracy and faster detection speed than the traditional full-color image target recognition algorithm.In addition,based on the stability and reliability of the system,based on the proposed network architecture,this paper designs and implements an intelligent vending machine system,and realizes the functions of scanning the code to open the door,freely select the goods,and close the door automatically.At present,it has been tested online and is in good condition,which greatly improves the user's shopping experience.During the Coronavirus pandemic,the use of intelligent vending cabinets can effectively keep physical distancing by avoiding unnecessary travel and shopping,stay away from large groups of people,thereby reducing the risk of transmission,and helping the prevention and control of the pandemic.