Research On Embedded System Of Liquor Selection Based On Computer Vision

During the production of liquor,liquor selection is a key process.And in this process,the workers can distinguish liquor of different quality and store them separately.The traditional liquor selection technique has passed down from generation to generation,which adopts the method of "watching the liquor flowers then selecting liquor".The bubbles generated by the impact of liquor on the bowl are called the liquor flowers.By observing the visual characteristics of the liquor flowers,liquor is distinguished and selected by the workers.This method has shortcomings such as different selection standards,strong dependence on experience,and lack of successors,which is not conducive to the stability of product quality.Replacing manual work with automatic liquor selection system is the trend of development.And the process of liquor selection belongs to visual classification tasks,which is very suitable for computer automatic implementation.Moreover,embedded devices have advantages such as the low power consumption,high cost performance and easy maintenance.So using embedded devices to implement an automatic liquor selection system can further reduce costs.However,embedded devices also have disadvantages such as low computing performance,small memory bandwidth and on-chip cache.Therefore,running a program on an embedded device can not be as easy as on a high-performance computer.In order to highlight the advantages of embedded devices,make up for the disadvantages,and finally implement a visual liquor selection system which is suitable for embedded devices,this thesis has carried out in-depth research.In this thesis,a lighter model and a more efficient computing strategy are adopted in the design,making full use of the embedded device’s resources,while maintaining high accuracy of liquor selection,and finally achieving the effect of maximizing strengths and minimizing weaknesses.It can accurately,reliably,efficiently and cost-effectively complete tasks of liquor selection on embedded devices.The work of this thesis specifically includes:1)An embedded automatic liquor selection system is designed,and the system has two modes: the normal mode and the testing mode.In the normal mode,the system automatically and stably implement the task of liquor selection according to computer programs.Once the programs fails,the system would switch to the testing mode.In the testing mode,workers replace computer programs of liquor selection with manual work,which has low efficiency but make production normal again.At the same time,the updated program can be quickly deployed,verified and tested on hardware in the testing mode.2)A special method for deploying programs is proposed,which is implemented based on Webassembly.The entire liquor selection program is compiled into an efficient Webassembly module,which is loaded and called by the web page.The web page and Webassembly module are saved on the upper computer,and any lower computer can access the web page through a browser to obtain the program.The testing mode of the system uses this special deployment method,by modifying the program saved on the upper computer,it can realize rapid changes to any lower computer,so that the updated program can be quickly tested and verified,while the time in the testing phase can be reduced.3)A vision-based liquor selection program is implemented.Compared with relevant literature,the pre-processing method in this thesis is improved.Meanwhile,a lightweight liquor selection neural network model called Wine Res6 is designed,and the size of the model using the Caffe framework is only 655.4KB.At the same time,a state machine-based post-processing algorithm is designed,which can avoid fluctuations that only rely on the results of the neural network model,and effectively improve the accuracy of liquor selection.Finally,the accuracy rate of 97.2% is achieved on the test set.4)An inference engine called ESNN is implemented,which is suitable for embedded devices to run the neural network model.ESNN can efficiently run common neural networks.And it is implemented in pure C++,with strong cross-platform capabilities.Meanwhile,its functions are compact and complete.It also does not depend on any third-party libraries.The liquor selection program in this thesis uses ESNN to run the neural network.The implementation of ESNN draws on the design ideas of other inference engines,frameworks like NCNN and Caffe,but has some unique innovative optimizations.