Research On Pest Monitoring Method And System Based On Deep Learning

Insect pests are one of the main reasons that affect the yield and quality of crops around the world.In pest monitoring,accurate and rapid identification of pests is the key part.Traditional pest image collection,identification,and diagnosis rely on expert knowledge to a high degree,the labor cost and time cost are large,and the accuracy of the diagnosis results is not high enough and error-prone,which is difficult to meet the intelligent and precise needs of smart agriculture.Image processing technology based on computer vision has been widely studied in the field of automatic pest detection.However,the current research on pest identification mainly focuses on a single type of pest detection task,and the distribution of these single type of pests is simple and difficult to detect.Low,for pest images with multiple species and complex adhesions,the classification is not accurate,the counting accuracy is not high,and the detection accuracy is low;on the other hand,the pest detection process in the study is basically to manually place pest traps in the field,the images of these trapped pests are collected by manual photography,and then experts manually identify and count the pests in these images.This method is time-consuming and labor-intensive,with a low degree of automation,and the recognition results are not accurate enough.The perception of field pest information lacks timeliness and accuracy,which increases the reliability and feasibility of using computer image processing technology to monitor field pests.Aiming at the above two problems,this paper conducts research from the following aspects:(1)The pest target detection method on the sticky insect board based on the public data set Yellow Sticky Traps is studied.The public dataset Yellow Sticky Traps is a dataset of pests on sticky bug boards produced and published by the 4TU consortium in the Netherlands.The incorrectly labeled data were corrected using the image labeling software Label Img.In order to improve the relative spatial scale of pests and highlight the detailed characteristics of pests,the images in the Yellow Sticky Traps dataset were augmented by scaling,cropping,and background removal,and a total of 117500 images were obtained.On the basis of the residual network Res Net,the single-channel convolution is redesigned into a multi-channel convolution,the downsampling structure is optimized,a deformable convolution kernel is introduced,and a new convolutional neural network model DPe Net is established.For training,DPe Net is trained on the training set using the training strategy of training low-resolution images first and then training all images.On the test set,the AP value of DPe Net reaches 0.941,which has a great advantage over the current mainstream deep learning target detection algorithms.(2)The transfer learning method based on small sample dataset is studied.Fortyfive high-resolution images of sticky board pests under different climatic conditions were manually collected to construct a small-sample dataset Yellow Pest-2022.The dataset contains 5 types of targets including flies,mosquitoes,bees,spiders,and other pests,and is annotated using the image labeling software Label Img.To highlight the semantic features of pests in the dataset,data augmentation methods such as adjusting contrast,exposure,saturation,etc.are applied to the images in the dataset Yellow Pest-2022.The transfer performance of DPe Net on small-sample datasets is studied by using two transfer learning methods to extract network parameters with frozen levy and finetuning without freezing parameters.It is verified that the transfer learning model obtained without freezing parameters has better transfer learning performance,and the AP value on the Yellow Pest-2022 test set reaches 0.940.Compared with the target detection algorithm adopting the same pre-training and transfer learning strategy,the accuracy coefficient R2 of DPe Net for pest count reaches 0.982,which has obvious advantages.The feasibility of transfer learning to enable the model to achieve similar but different pest detection tasks on small datasets is verified with high accuracy.(3)Developed a remote system for field pest monitoring based on convolutional neural network.In terms of hardware design,the system uses Raspberry Pi 4B as the Microcontroller Unit(MCU),uses the IMX179 camera module to obtain field pest images,and uses the temperature and humidity sensor SHT30 and light sensor BH1750 to obtain the field temperature,humidity and light intensity and other environments.information.In terms of information collection and uploading,the system uses the Python program to call third-party libraries such as Open CV and smbus to collect sensor information;build an Apache server on Tencent Cloud,and use the PHP service program to perform data increase and decrease operations on the My SQL database to store and call Field environment information such as temperature and humidity,light intensity;receive image information of pests on the sticky insect board through the POST program and the Flask service.In terms of result access,the processing results such as changes in field temperature and humidity,changes in light intensity,and changes in the number of various types of pests are called through the Apache server and PHP service program,and the processing results are visualized on the terminal using HTML5 web pages.The experiment verifies the cross-platform nature of the system on different terminal platforms,and meets the needs of different terminal devices for data viewing.