Design And Implementation Of Embedded Beverage Recognition System

Nowadays,the traditional beverage sales channels become saturated and the promotion cost is expensive.As a new channel,the automatic beverage vending machine has been one of the most popular new retail models for beverage manufacturers and distributors due to its low channel cost and beverage brand effect.The core function of the beverage vending machine is the beverage identification system,which can provide necessary information for the delivery of goods and the management of storage and logistics through the real-time detection of beverage types and inventory in the vending machine.With the development of deep learning and computer vision technology,the beverage recognition system based on the embedded terminal is gradually put into application.The system can intelligently analyze the image collected by the camera and perceive the beverage type and inventory in the vending machine,which has the characteristics of small size,high accuracy,fast speed and low cost,leading to a very large market potential.Limited by the hardware performance of the embedded terminal,how to select the appropriate detection and recognition model,while meeting the system accuracy and real-time requirements,and how to compress and optimize the model is the difficulty of deep learning technology application in the embedded beverage recognition system.In view of these problems,the thesis mainly does the following work:(1)Based on the comprehensive comparison of the performance of typical object detection models,YOLO v2-Tiny neural network is selected as the core network of the system.Through the USB camera installed on the top of the vending machine,more than 8,000 beverage images of common 41 kinds are collected under different light intensity and angles,and then they are made into the beverage recognition dataset of the VOC fomat,on which the detection model is trained.Considering the efficiency of model iteration in practical engineering applications,this thesis studies the method of accelerating model training by transfer learning.In order to further improve the recognition accuracy,the thesis also explores the two-stage scheme where sub classes are recognized after the broad classes are detected.Finally,darkflow is used to transform the weight of the obtained YOLO v2-Tiny detection model,and then the tensorflow quantitative tool is employed to compress and optimize it.The volume of the model can be compressed by 75% to meet the requirement of running in the embedded end with limited resources.(2)The embedded beverage recognition system is designed and implemented on the rk3399 embedded platform,where the cross-platform deployment is achieved with the tensorflow framework,and the system realizes the functions of picture detection,threshold adjustment,diary recording,etc.In consideration of the demand of practical applications,the algorithm and model are encapsulated and encrypted to ensure the security of the system,which enables it to be put into the market conveniently and quickly.The embedded beverage recognition system developed in the RK3399 embedded platform can achieve the detection speed of 0.53 fps and the accuracy of 96%,which can meet the needs of practical applications.