| Due to the long sunshine duration,high accumulated temperature,and high annual solar radiation in Xinjiang.By 2018,the area of cotton planted in Xinjiang has reached 2.644 million hm2,an increase of 189,000 hm2 compared with 2017.Based on plastic film mulch and drip irrigation cotton planting patterns,more than 80% of cotton fields in northern Xinjiang have achieved mechanical harvesting.As the leader of Xinjiang’s economic crops,whether high yield directly determines the economic income of cotton farmers.However,the control of pests and diseases also directly affects the level of cotton production.With the widespread cultivation of BT insect-resistant cotton,aphids have become the main pests that harm cotton yield and quality.The cotton aphid breeds fast,iterative cycle is short,and one generation can be completed in 3-5 days under the right temperature.When the external environment changes,aphids produce migratory individuals through sexual reproduction and migrate over long distances.At the same time,aphids are associated with ants during the growth and reproduction process,and they can migrate short distances through ants.Because aphids produce ’Honey Dew’ during the growth process,on the basis of affecting the yield and quality of cotton,it is also easy to spread the virus,further affecting cotton growth.Aphid pest outbreaks are usually accompanied by cotton disease outbreaks,reduced cotton yields and reduced quality.At the same time,blind application by cotton farmers will also cause economic waste,pesticide residues and environmental pollution.Cotton aphids are the main factors restricting cotton production.Traditional cotton aphid monitoring methods and methods can no longer meet the requirements of modern agriculture and precision agriculture development.How to obtain the aphid quantity information in cotton field efficiently,quickly and accurately is a key part of cotton aphid pest control.Due to the shortcomings of low recognition efficiency,strong subjectivity,and certain plant protection experience in the process of artificially implementing aphid identification and statistics.Therefore,on the basis of setting a yellow stickworm board,using machine vision instead of manual identification and counting of aphids has certain application value.Due to the rapid development of computer vision in the field of agriculture,insect recognition counts based on computer vision and image recognition technology are now frequently reported.The establishment of insect discrimination models mainly depends on the differences in morphological,color or texture features between insects.On this basis,morphological characteristics,as one of the most stable characteristics of insects,have great advantages in the construction of insect discriminant models.Currently in the field of deep learning,with the research of target detection technology is gradually deepening.Commonly used object detection algorithms developed from CNN(Convolutional Neural Network)mainly include Faster RCNN series,YOLO series and SDD series.In the recognition process,the convolutional neural network can automatically extract the target features.In the process of simplifying the manual selection of features,the computer replaces the manual,which eliminates the error caused by subjectivity and reduces the requirements for related field experience.Among them,the models commonly used by Faster RCNN are VGG-16 and Res Net-50.Using computer vision and image processing technology,the color component that is most suitable for segmenting the yellow plate and the background is selected according to the difference in gray values between the yellow plate and the different backgrounds.The precise positioning of the yellow board is achieved by computer image processing.By designing a new yellow plate image processing scheme to improve the ability of yellow plate image processing to adapt to non-uniform illumination,aphids and yellow plates can be segmented under non-uniform illumination.The winged aphid discriminant model is constructed by machine learning to realize the identification and counting of aphids in the yellow board image;the cotton leaf aphid recognition model is constructed by the Faster RCNN model;the trained network is modified to reduce the resolution of the cotton leaf image and reduce the ROI area to solve the problems of too long training model of Faster RCNN model,low recognition accuracy,and inaccurate target calibration.Based on the above research results,an automatic yellow plate cotton aphid monitoring device was designed.Combining the yellow sticky worm board with the automatic image acquisition system,the change of the yellow board is controlled by the remote control system to realize the cotton field aphid trapping and automatic replacement of the yellow board.By controlling the rotation of the camera,the yellow plate image is automatically collected.The main research results are as follows:1 Yellow board positioning and image enhancement based on yellow board imagesIn this study,the optimal color component combination suitable for positioning the yellow plate and suitable for segmenting the target and the yellow plate was selected through the difference in gray values between the yellow plate and the complex background,the insect target and the yellow plate under different color components.The sub-adaptive threshold segmentation realizes the precise positioningof the yellow board,and the complex background(sky,clouds,cotton,and soil)around the yellow board is eliminated by a mask.An image enhancement method based on gray features is proposed.The original image is Gaussian filtered under the G color component.The filtering result is different from the original image.The difference result is quotient with the original image to enhance the color characteristics of the aphid.Segmentation of insect targets and yellow boards under non-uniform lighting conditions.The test results show that the combination of H and G color components in each color component of each color space can better achieve the positioning of the yellow plate,the segmentation of the target insect and the background;and the image under two non-uniform illumination conditions,strong and weak,through gray enhancement processing.The segmentation result is stable,the target morphological characteristics are kept good,and the ability to remove interference is strong in the process of gray-enhanced image processing.2 Identification and counting of cotton aphids based on yellow boardEliminate ladybugs,leaf hoppers,and noise on the yellow board by setting the area threshold.Aphid recognition models were constructed based on the differences in morphological characteristics between aphids and thrips to achieve recognition between aphids and other insects.Based on the BP neural network recognition model,10 yellow plate images were randomly selected for recognition and counting.The test results show that the ladybug and leafhopper on the yellow board are eliminated by setting the area threshold to 150.Remove the noise interference on the yellow board by setting the area threshold 20.SVM and BP neural networks were used to construct aphid discrimination models.The recognition accuracy of different built-in nuclei constructed by SVM is not much different.BP neural network with stronger generalization ability was used for verification.It was found that the ability of BP neural network to recognize aphids was slightly stronger than SVM.The morphological features that are independent of length and related are selected as input vectors of the BP neural network to reconstruct the recognition model.It is found that different morphological features have different degrees of impact on the recognition of aphids.Only the eccentricity,the long-short axis ratio,the complexity and The recognition model was constructed with the roundness-like morphological feature parameters,and the model accuracy reached 89.5%.In the process of aphid recognition and counting,the recognition rate of aphids was 93.33%.In the actual application process,the identified objects are no longer classic aphids and thrips,which leads to an increase in the rate of miss identification.3 Identification and calibration of aphid species in cotton leaves based on deep learningIn this study,a faster RCNN algorithm was used to construct an aphid species identification model,and the initial data model was used to determine the batch size and learning rate parameter hyperparameters.By introducing the trained network and modifying it,reducing the resolution of cotton leaf images and reducing the ROI to solve the process of Faster RCNN network detection of cotton leaf aphids caused by fewer samples and high resolution of the image itself The model training time is too long,the recognition accuracy is low,and the target calibration is not accurate.The experimental results show that the Faster RCNN model has the best recognition effect when the minimum batch processing is set to 1 and the learning rate is set to 0.001.By introducing a trained network and modifying it and reducing the image resolution,the model recognition accuracy increased from 9.1% to 71%,and the training time decreased from 32 hours to 4 hours.4 A cotton aphid monitoring device capable of automatically replacing yellow boardsCombined with the remote control APP,a monitoring device that can automatically replace the yellow aphid cotton aphid is designed.The device comprises a cotton aphid trapping area,an image acquisition area,a weather sensor assembly,a control box and a main body support.The cotton aphid was trapped by yellow sticky paper,and the yellow plate image was obtained through the image acquisition system.The yellow plate was positioned and the aphid color characteristics were enhanced through image processing.The aphid recognition model was established by setting the area threshold and morphological characteristics,and finally real-time monitoring of the field The purpose of the number of aphids on the yellow plate image.In this research,image acquisition devices,automatic control technology,image processing methods,and recognition models based on machine learning were used to realize the automatic acquisition of aphid images,the positioning of yellow boards,the removal of ladybug leafhoppers and noise,and the discrimination between aphids and thrips.Through the automatic replacement of the yellow board,the occurrence of cotton aphid adhesion and overlap is reduced,and the problems that the aphid image has a complex background during the processing process are bound to be difficult. |
PDF Full Text Request