Research On Commodity Recognition Algorithm Of Self-service Vending Machine Based On Deep Learning

Huge market demand and convenient shopping methods have made vending machines a popular research direction in the retail industry.However,at present,these commodity detection technologies of vending machines have their own shortcomings.RFID tags are easy to tear.The gravity-based commodity detection method has limited detection types.The vision-based commodity recognition method has low accuracy.The cost of multi-sensor fusion recognition is high.In recent years,the rapid development of object detection based on deep learning has brought a new method for the commodity detection in vending machines,but the deep learning model is complicated,parameters are redundant,and the calculation is large.This article takes the vending machine as the application background,and mainly studies the commodity detection algorithm,including:(1)By analyzing the functional and performance requirements of the vending machine system,a commodity detection method based on static visual detection is determined,and the workflow of the vending machine is formulated;(2)The internal structure of a Z-type vending machine is designed to solve the problem of blocked products;(3)Select the camera according to the field of view,and use Open CV to correct the distorted image in combination with camera calibration method.(4)Make a commodity data set.Using the data augmentation method,it solves the problem of time-consuming and labor-intensive manual labeling of a large number of pictures and the problem of over-fitting the model training caused by insufficient data;(5)Use the K-means++ method to reset the Anchor boxes of the product data set.GIo U is used to calculate the box loss during the training.The training loss decreases faster,the accuracy rate and recall rate increase,and finally the model fitting is accelerated to a certain extent;(6)The double pruning method is used to compress the YOLOv3 network,which greatly reduces model parameters and calculations,speeds up model inference,saves computing resources,and solves the problem of heavy calculation load caused by parameter redundancy in deep learning models.The model does exploratory research on terminal deployments with insufficient computing resources.