Research On Commodity Recognition Algorithm Based On Embedded Platform

In recent years,with the rapid development of mobile payment,artificial intelligence,edge computing and other technologies,the traditional retail model has undergone tremendous changes.Under the background of "new retail" proposed by Alibaba,the intelligent vending cabinets based on computer vision have received extensive attention.Smart vending cabinets need to quickly and accurately identify commodity specified by customers,and then obtain commodity information and perform settlement.Recently,most commodity identification systems are deployed on servers.With the increase in the number of smart vending equipment,the pressure on the servers have increased,making the real-time detection of commodity unreachable,which in turn affects the customer’s shopping experience.This thesis design and implement a lightweight commodity recognition algorithm and deploy it on embedded platform.The main works are as follows:In view of the requirements of the accuracy and recognition speed of the merchandise recognition algorithm of the sales container,this thesis uses the YOLOv3 algorithm.Firstly,8512 commodity images were collected in 9 categories,and the commodity images were labeled as PASCALVOC data format using a labeling tool for model training and testing.Secondly,a lightweight network model DS＿YOLOV3 is proposed by using deep separable convolution and inverted residual structure in Mobile Net V2 model to reconstruct YOLOV3 model.Finally,aiming at the problem that the YOLOV3 algorithm uses the k-means clustering algorithm to generate the prior box is not strong in robustness,the k-means++ algorithm is used to cluster the commodity data set to get the prior box.The experimental results show that the m AP of the DS＿YOLOv3 algorithm is 94.69%,and the detection speed is 20.34F/s,which meets the requirements of smart vending cabinets in terms of detection speed.Aiming at the problem of low positioning accuracy of the frame regression loss function of the DS＿YOLOv3 algorithm,this thesis chooses CIo U as the frame regression loss function to improve the accuracy of target positioning.In addition,due to the local information loss caused by DS＿YOLOV3 algorithm in preprocessing and multi-scale detection,the SPP module is used to extract multi-scale features under different receptive fields,and the local features and global features are intergrated to obtain richer semantic information.The experimental results show that for the improved DS＿YOLOv3 algorithm,m AP reaches 97.93%,and the detection speed is19.51F/s.This research has improved the detection speed,the commodity recognition accuracy of the smart vending cabinet without affecting the detection speed.In this thesis,the YOLOV3 algorithm is improved and a commodity recognition algorithm suitable for embedded platform deployment is proposed.Experiments are carried out on the commodity data set.The results show that the newly proposed algorithm has been improved in detection speed and accuracy,which meets the needs of smart vending cabinets.