Design And Implementation Of Embedded Convolutional Neural Network Bread Intelligent Retail System

In recent years,in the wave of "Internet+",cutting-edge technologies such as artificial intelligence are moving from academic research to commercial implementation.The deep integration of artificial intelligence and the real economy has become one of the important national strategies,which has also brought new opportunity.At present,many retail methods use manual identification of bar code cashier checkout,which will cause the problem of excessive cashier labor cost and long checkout time.This paper uses artificial intelligence technology to achieve machine detection and recognition of bread and the machine completes the cash register checkout.Intelligent bread recognition and settlement is the key technology of this paper.Through research on the recognition accuracy,detection speed and use cost,we design and implement an embedded convolutional neural network bread intelligent retail system to promote the intelligent development of the retail industry.This article divides the system architecture into several modules:image acquisition,bread detection and recognition,and pricing payment,and focuses on the in-depth study of the related algorithms of the bread detection and recognition module.The bread data set is established by collecting and labeling bread mold images.Based on the theory of convolutional neural network and target detection and recognition technology,the method of bread detection and recognition is analyzed.This paper first designed a regional candidate two-stage bread recognition model.The mAP reached 0.99 to verify the feasibility of the bread recognition task.Later,the end-to-end singlestage bread recognition model was designed to increase the detection speed by 4 times while the mAP maintained at 0.99.Use jetson TX2 as an embedded experimental platform.In order to realize the commercial implementation of the system,this article focuses on the cost issue,using the cheaper C10 device with Android system as the embedded platform,the device has poor performance and the model capacity needs to be limited to about 10MB,facing this problem,This paper introduces a lightweight convolutional neural network and improves the algorithm,uses image processing to extract candidate regions,and uses a lightweight convolutional neural network to implement classification and recognition methods to reduce the model capacity by nearly 80 times to 3MB and maintain mAP at 0.98.Aiming at the problem of image processing affected by lighting,this paper further improves the algorithm,and uses a light-weight convolutional neural network after network pruning combined with ncnn framework to achieve the approximate effect and solve the lighting problem.Finally,this article designs an Android app to transplant the algorithm and implement the entire embedded bread intelligent retail system.After testing,the system has a fast response speed,accurate recognition,and can be quickly used in bread stores.Further research in the future can focus on the training optimization of the bread data set,shorten the training time and simplify the training process.