Research On Lingwu Long Jujube Detection And Maturity Classification System Based On Deep Learning

Lingwu long jujube is a very important economic fruit in Ningxia Hui Autonomous Region,but at present it is mainly picked by hand,which has the problems of high labor intensity and low picking efficiency.In order to realize the automatic and intelligent picking of Lingwu long jujube,it is necessary to design an intelligent picking robot for Lingwu long jujube,and the vision system with high recognition rate is the core part of the picking robot.In this paper,we propose a light-weight target detection method for the detection of Lingwu long jujube images without using a pre-trained model to achieve high detection accuracy.Secondly,in order to guide the robot to pick Lingwu long jujube with specific maturity and further realize the commercialization of Lingwu long jujube later processing,a network model meeting the requirements of Lingwu long jujube maturity classification task is proposed.Finally,a system with the functions of detection and ripeness classification of Lingwu long jujubes is designed to facilitate users.The main research contents of this paper are as follows:(1)Research on the target detection problem of Lingwu long jujubes.Firstly,a target detection dataset of Lingwu long jujubes is constructed.Then,to address the problem that the traditional SSD model loaded with a pre-trained model cannot change the network structure and cannot be used when the memory resources of the device are limited.The network structure of the SSD model is improved,and a light-weight target detection method is proposed to achieve high detection accuracy of the Lingwu long jujube images without using the pre-trained model.The improvements include 1)replacing the original backbone network ResNet50 with an improved DenseNet network;2)replacing the first three extra layers in the SSD model with an Inception module;3)adding a multi-level fusion structure.After introducing the experimental equipment and evaluation metrics,experiments are conducted on the target detection dataset of Lingwu long long jujubes.Finally,the comparison test results show that all the proposed improvements are effective.The mAP of improved SSD model is 96.60%,the detection speed is 28.05 frames/s,and the number of parameters is 1.99×106,which is 2.02 percentage points and 0.05 percentage points higher than the mAP of the SSD model and the SSD model(pre-trained),respectively.The number of network structure parameters of improved SSD is 11.14 × 106 less than that of the SSD model,which can achieve better detection results without using the pre-trained model,can meet the requirements of lightweight networks,and can be deployed on devices with limited memory resources.(2)Research on the maturity classification of Lingwu long jujube.Firstly,we constructed the Lingwu long jujube maturity classification dataset.After introducing the experimental equipment and evaluation indexes,we tested the performance of seven networks,which includes ResNet101,MobileNetV2,ShuffleNetV2,VGGNet,DenseNet,GhostNet and EfficientNetV2 on the maturity classification dataset.After comprehensive comparison,ShuffleNetV2 was selected as the base network model for the Lingwu date maturity classification task.We also follow the principle of increasing the correct rate as much as possible while ensuring the classification speed and number of parameters.The improvements for ShuffleNetV2 network include 1)replacing the 3 × 3 convolution in the right branch of the ShuffleNetV2 module with a 5 × 5 convolution;2)adding a SRM attention module to the ShuffleNetV2 module;3)changing the number of ShuffleNetV2 module group number from[4,8,4]to[3,6,3].Finally,the effectiveness of the above improvement strategies is experimentally demonstrated.The classification accuracy of improved network ShuffleNetV2-5 × 5-SRM-[3,6,3]is 90.56%,the speed is 93.03 frames/s,and the network structure parameters is 0.33×106.Compared with ShuffleNetV2,the classification accuracy is improved by 4.86 percentage points despite sacrificing some speed and parameters.The parameters are reduced 0.02 × 106,which meets the improvement requirements.(3)Design of the Lingwu long jujube detection and maturity classification system.We first analyzed the user requirements.Then designed the overall framework of the Lingwu long jujube detection and maturity classification system.Finally the system is tested and the interface is demonstrated.The test results show that,after loading the improved SSD model and ShuffleNetV2-5 ×5-SRM-[3,6,3]training parameters,the system is able to complete the target detection and maturity classification of Lingwu long jujube well,which can lay the foundation for the subsequent practical application of the system.The detection and maturity classification system designed in this paper can solve the two problems of detection and ripeness classification of Lingwu long jujube.This system is simple and easy to operate,which can also provide visual technical support for the development of intelligent picking robot of Lingwu long jujube.