Research On Fruit Appearance Detection Technology Based On Deep Learning

In recent years,artificial intelligence technology is constantly empowering traditional agriculture.In this paper,we focus on the detection of fruit appearance,using deep convolutional neural network as the main method,and apples,bananas and other fruits are used as the carrier to conduct in-depth research.The core research and work of this paper include:(1)This paper builds image acquisition device,collects and annotates images,and constructs datasets.A multi-fruit classification model based on a convolutional neural network was studied,and 10 types of fruits were classified to provide a solution for the offline unmanned self-selling fruit system.On this basis,this paper realized the classification and recognition of apple surface maturity,and the average accuracy rate of classification of four types of apple surface maturity reached 82%.(2)The three defects commonly detected on apple surfaces include rot,crack and worm disease.An end-to-end apple surface defect detection model is built based on one-stage and two-stage methods,respectively.The experimental results show that SSD?ResNet101 is better than Faster R-CNN?ResNet101 in detecting various defects.This paper also uses the YOLOv3 network to realize the engineering of apple surface defect detection.(3)This paper classifies two types of banana point cloud data.Get the depth information of bananas through the way of photo modeling and the method of using Kinect v2.A 7-layer convolutional neural network is built to complete the detection of two types of bananas of different shapes and sizes,and the agreement rate between the test results and the true value exceeded 91%.This paper also imports banana point cloud data into PointNet.The accuracy of the classification reached 96.25%.