Implementation Of Granary Pest Monitoring System Based On Deep Convolutional Neural Network

In the process of grain storage,pests are the biggest threat to grain.The grain loss caused by pests in China's granaries reaches millions of tons every year,causing huge economic losses to the country.Therefore,the monitoring of pests in the granary is very necessary.At present,the vast majority of granaries carry out fumigation several times a year to reduce the losses caused by pests.In order to reduce the number of unnecessary fumigation,it is necessary to monitor granary pests in a timely and accurate manner.The core of monitoring is intelligent identification and counting the pests in the granary.According to the 6 types of granary pest images,the thesis uses deep convolutional neural network(CNN)technology to design and implement a granary pest monitoring system.The main research work of the thesis is as follows:(1)The L-SSD pest detection network is proposed,which realizes the discrimination and positioning of the 6 common granary pests under the background of grain fragments.Based on the SSD-V network,the training speed is improved by reducing the scale of some convolutional layers,optimizing the loss function to increase the sample utilization rate of difficult cases,increasing the number of anchor points in the network and optimizing the confidence threshold to improve the detection of pests.The test results show that the mAP(Mean Average Precision)value of the LSSD network for the detection of granary pests has increased by 1.5% to 10.8% compared with other object detection networks,and compared with different types of SSD networks,it has increased by 1.8% to 3%.(2)Aiming at the problem of the decrease of the accuracy of counting caused by the missed inspection of pests,the pest counting method of area segmentation is designed.This method first divides the image into grids and divides them into sub-graphs.The sub-graphs are binarized and thresholded.Then,the mathematical morphology method is used to eliminate the background interference of the sub-graphs.Afterwards,the watermarked pests in the image are marked watershed.The algorithm performs sticky segmentation,and finally counts the number of regions(the number of pests)segmented by all subgraphs.Experiments show that the average accuracy of this method is 90.87%.(3)A granary pest monitoring system is designed.The system mainly includes functions such as reading images,pest detection,pest counting and pest warning.The system is deployed and tested in the simulation of the granary.The test results show that the system functions properly,the average time to parse the pest image is 3.53 seconds,and the average mAP value for the pest detection of the granary is 82.11%.