Desgin Of An Insect Monitoring System With Deep Learning For The Tomato Culviation

Tomato is an important economical crop in China,occupying an addressed position in the trade of fruits and vegetables.Insects are one of the most important factors affecting the quality and yield of the tomato production.Fast and early monitoring of insects should be conducive to the efficiency improvement of the insect control.It could also contribute to the tomato production promoting and farmers’ income increasing.Traditional pest identification methods still have some problems such as strong subjectivity,high reliance on expert knowledge,low efficiency,and high application costs.Thus,it is urgent to develop a real-time,accurate and rapid insect monitoring system to improve the efficiency of pest monitoring and early warning for tomato growers.The main contents of this thesis include:(1)The construction of an insect image dataset for tomatoes.The deep learning model had strong dependence on the dataset.But currently there were still few datasets focusing on tomato pests.It was difficult to train a recognition model with high detection performance.Therefore,in this thesis,a total of 3985 original images of insects in tomato fields were collected through two methods,including field shooting and open source dataset screening.In order to enhance the generalization of the model,data augmentation methods such as random cropping,color distortion,random rotation,and random erasure were applied in this thesis to expand the original image pools.Finally,an dataset with 18359 images of insects in tomato fields was constructed(including cutworm,red spider mite,beet armyworm,flea beetle,aphid,whitefly,corn borer,meadow moth,bollworn,armyworm).(2)Selection of the baseline detection model for agricultural insects.The performance of the single-stage and double-stage detection models on the COCO dataset was summarized in this thesis.The YOLO v5 series was selected due to three aspects of accuracy,speed,and deployment.At the same time,the performance of the YOLO v5 series was tested on the open source agricultural pest dataset Pest24.The test results showed that the number of parameters using YOLO v5-m was half of that when using YOLO v5-l.The reasoning speed was 86%,while the recognition accuracy was only 3.3% which were respectively faster and lower than cases using YOLO v5-l.Considering the influence of parameter quantity on model deployment,YOLO v5-m was finally selected as the baseline model of this thesis.(3)Construction of the tomato insect detection model S-YOLO v5-m.There were several problems of the baseline model such as weak feature extraction ability for tomato pests,low attention to the target area,and background noise in the feature due to the small size of the insects and the similarity to the background color.The SPD-Conv module,Ghost module,CBAM attention mechanism and AEM module were introduced to construct the S-YOLO v5-m model.Compared with the ablation test results the pre-improved YOLO v5-m model,the m AP of S-YOLO v5-m model had increased of4.73%,while the number of parameters reduced of 31%,and the inference speed of a single image was shorten with 1.3ms.It could be seen from the visualization results that S-YOLO v5-m still had certain advantages in the detection of small target pests,with a higher degree of attention to small targets and richer feature information.At the same time,S-YOLO v5-m had faster convergence and stronger learning ability.In addition,several common detection models were also compared in this thesis.The overall performance of S-YOLO v5-m was still higher than other common models.(4)Development of an initiative tomato pest monitoring system.In order to solve the problems of poor migration and difficulty of passive capture of some tomato pests,an initiative pest monitoring device was designed in this thesis.It could obtain the image data at any position in a certain area through the sliding module,pan-tilt module and camera.By deploying the intelligent computing platform of the S-YOLO v5-m model,the realtime monitoring of insects was realized.Meanwhile,visual interfaces were developed using the Py Qt language,including model training,user self-inspection,and intelligent monitoring.With this system,the users will be able to conduct reasonable insect management during tomato cultivation according to the data reports generated by the monitoring system.The monitoring system developed in this thesis was tested to be working well in tomato fields.