| The fruit industry economy is one of the important components of agricultural economy.Orchard pests always threaten the overall yield and quality of the fruit,affecting the development speed and quality of the fruit industry.Therefore,in the process of fruit planting,accurate and efficient detection of orchard pests is essential for pest management and improving the economic benefits of the fruit industry.The traditional detection methods of orchard pests mostly rely on manual methods or traditional computer vision methods.They have disadvantages such as low detection efficiency,strong subjectivity,poor generalizability,and long detection cycles,which are difficult to meet the needs of intelligent real-time detection.In order to achieve intelligent object detection of orchard pests,the main research contents of this paper are as follows:(1)In order to meet the needs object detection of orchard pests based on deep learning,this paper built an orchard pest color image database(PestImgData)through two methods:web crawler and specimen image collection,and manually annotated all image data,including the category information of the pest and its position coordinate information in the color image.The PestImgData was composed of two parts of image data:PestImgData-1 and PestImgData-2.It contained a total of 24796 color images of orchard pests,covering 7 types of orchard pests.In addition,all orchard pest image data in the PestImgData image database had undergone image normalization,and the image format and size were consistent.In the agricultural field,the available image data of orchard pests was very scarce.The PestImgData image database was not only suitable for the research content in this paper,but also provided data support for object detection tasks in other application scenarios.(2)On the basis of the PestImgData,this paper proposed a 2-stage orchard pest object detection model Faster R-CNN-PEST based on deep learning.Then,the performance of the model was systematically and comprehensively evaluated from multiple perspectives such as recall,precision,mean average precision(mAP),and detection speed.The mAP of the Faster R-CNN-PEST model was 66.37%.It could locate and classify orchard pests in simple scenarios,but the detection effect for overlapping targets and high-density targets was poor,and there were a lot of false detections and missed detections.Based on the Faster R-CNN-PEST model,this paper proposed an improved orchard pest object detection model Res-Faster R-CNN-PEST by improving the network structure and image data of the original model.Compared with the original model,the mAP of the Res-Faster R-CNN-PEST model had increased by 22.56%to 81.34%,and the improved model had improved the detection accuracy of various types of orchard pests to varying degrees.In addition,in the testing process of overlapping targets and high-density targets,the performance of the improved model was both better than the original model.(3)The two-stage orchard pest object detection model was difficult to meet the needs of real-time detection.This paper proposed a one-stage orchard pest object detection model based on YOLOv4:T-YOLOv4-PEST.The mAP of the T-YOLOv4-PEST model was 90.02%,and the detection speed of a single image was 14.44ms,which could meet the needs of real-time detection.On the basis of the T-YOLOv4-PEST model,the following six dimensions of research had been carried out:transfer learning research,activation function research,learning rate research,prior box(anchor box)research,batch normalization research,visual analysis and research on network reasoning,and trained 5 relatively independent new orchard pest object detection models.Through systematic and comprehensive analysis of the overall performance of the above six models,the improvement direction of the model was sought and the optimal parameter configuration of the network structure was determined.Then,based on the optimal parameter configuration of the network model,the T-YOLOv4-PEST model was improved from the following three optimization perspectives:the post-processing non-maximum suppression(NMS)strategy of the improved network model was changed to the DIoU-NMS strategy,the training method was the second fine-tuning training and the training dataset was enhanced.A new orchard pest object detection model FD-YOLOv4-PEST was proposed.At the same time,in order to verify the performance of the improved model,this paper had trained the following five types of orchard pest object detection models based on the common 1-stage object detection algorithm for comparative study:the orchard pest object detection model based on Tiny-YOLOv4,the orchard pest object detection model based on YOLOv3,the orchard pest object detection model based on Tiny-YOLOv3,the orchard pest object detection model based on YOLOv2 and the orchard pest object detection model based on Tiny-YOLOv2.After systematically and comprehensively evaluating the performance indicators of the above 12 models,the F-D-YOLOv4-PEST model was the optimal model no matter in the average precision of each category of pests or the mAP.The mAP of the F-D-YOLOv4-PEST model was 92.86%,and the detection time of a single image was about 12.22ms,which could accurately achieve real-time detection of orchard pests.In addition,the F-D-YOLOv4-PEST model could also accurately locate and classify pest detection targets when the detection target overlap area was too large or the density was too high.(4)Based on the proposed orchard pest object detection model F-D-YOLOv4-PEST,this paper designed and built an orchard pest intelligent detection system.The intelligent detection system for orchard pests proposed in this paper included 4 modules:image acquisition module,preprocessing module,pest target detection module and information display and storage module.In addition,the system could also realize orchard pest object detection in a variety of situations such as single category single target,single category multiple targets and multiple categories and multiple targets. |
