Research And Application Of Intelligent Container Commodity Identification Algorithm Based On Embedded Convolution Neural Network

With the continuous development of artificial intelligence technology,technology based on computer vision has become a research hotspot in various AI application fields.The new retail mode of intelligent container based on computer vision technology has high commercial value.Consumers can purchase goods more easily while container operators can maintain container operation more easily.At the same time,container operators can mine potential data value from big data information of consumers' shopping.However,the current solutions based on computer vision technology are not perfect,and there are still two difficulties to be solved.The first one is that the accuracy of the object detection algorithm is not high enough,and the small error of the detection algorithm may lead to huge commercial losses.The second one is that the hardware cost of deploying the detection algorithm is too high.The amount of calculation in convolutional neural network is huge,which requires powerful computing hardware to provide support.Consequently,it has pushed up the overall production cost of the intelligent container,which is not conducive to commercial promotion.In order to improve the detection accuracy of commodity image and reduce the overall production cost of the intelligent container,this paper designs a lightweight single-stage convolutional neural network,and transplants the algorithm to the embedded development board with built-in NPU hardware accelerator.The main work of this paper is as follows:(1)Combined with the development characteristics of the new retail industry,this paper investigates the development of the unmanned container and the research status of single-stage object detection algorithm,and formulates the research objectives and technical routes of this paper.(2)This paper constructs an intelligent container commodity identification data set with 12 kinds of commodities.The main work includes commodity data collection,preprocessing the original image captured by the camera,and data annotation using annotation tools.(3)A lightweight single-stage convolution neural network(YOLO-embedded)is proposed to improve YOLOv3 detection algorithm.The deep separable convolution network is used as the backbone network of the model to extract commodity features.Then four feature maps of different sizes are designed to construct the feature pyramid to perform the commodity identification.At the same time,it uses loss function of the GIOU boundary box regression to improve the accuracy of commodity bounding box regression.In the experiment of algorithm comparison,YOLO-embedded algorithm gets better results in detection accuracy,speed and model size.(4)The YOLO-embedded algorithm is transplanted to the embedded development board TB-RK3399 Pro D.In the test of 1000 times of code scanning,door opening and shopping,the error occurs twice,and the accuracy reaches 99.8%,which meets the requirements of commercial promotion of intelligent containers.(5)This paper summarizes the related work and significance of this topic,and gives the direction of improvement in the future.