| Cotton is one of the world’s important cash crops and plays an important role in the economic development of both the world and China.Aphis gossypii Glover is one of the main pests of cotton seedlings,with a short growth cycle,long harm time,easy to break out into disasters,and very difficult to prevent and cure,seriously affecting the quality and yield of cotton.At present,judging the harm grade of Aphis gossypii Glover is mainly based on manual surveys,time-consuming and inefficient.How to quickly and effectively identify and monitor cotton seedling aphids is of great importance for their precise prevention and cure.This study is based on deep learning object detection technology to achieve rapid and accurate monitoring of Aphis gossypii Glover harm,and the main study contents and conclusions are as follows:1.The Aphis gossypii Glover harm detection data set was constructed based on expert experience and national grading standards,which contains a total of 16,950 images.Three object detection algorithms,Faster R-CNN(Region-based Convolutional Neural Network),SSD(Single Shot Multi Box Detector),and YOLOv5(You Only Look Once version 5),were selected to construct the Aphis gossypii Glover harm detection model and conduct training tests.The results showed that the YOLOv5 model had the highest m AP(mean Average Precision)value and FPS(Frame Per Second)value with 95.6% and 63.99,respectively,under the same training parameters.Based on the comparison of the three models,it was learned from the experimental comparison that when the image resolution was 640×640,the YOLOv5 s model had the best performance.2.To further enhance the detection effect of the Aphis gossypii Glover harm detection model,this study chose to improve the YOLOv5 s model at an image resolution of 640×640.Through comparison,it was found that the Aphid YOLOv5 s model with CA(Coordinate Attention)attention mechanism and adaptive spatial feature fusion added to the YOLOv5 s model at the same time had the best performance,with the m AP value of 97.8%,which was 1.8% higher than the original YOLOv5 s model.The FPS value was 76.63,which was only a decrease of 0.42.After comparing with other object detection models,we found that the m AP values of Aphid YOLOv5 s were 0.3%,3.0%,10.4%,and 36.3% higher than those of YOLOv8s(You Only Look Once version 8s),YOLOv4(You Only Look Once version 4),Faster R-CNN and SSD,respectively.The FPS values were 8.88,65.42,66.19,and 68.99 higher than the other models,respectively.3.In order to achieve real-time detection of Aphis gossypii Glover harm in the field,this study developed the Aphis gossypii Glover harm real-time detection software based on the Android platform.After testing and verifying,the detection accuracies of ‘Zhongmiansuo49’ and ‘Xinluzhong66’ were 91.9% and 89.2%,respectively;the detection times of single Aphis gossypii Glover harm images were 0.290 and 0.286 seconds,indicating that the practicality of Aphid YOLOv5 s Aphis gossypii Glover harm real-time detection software is better and can be used for the detection of different cotton varieties,providing field investigators with a more convenient and faster means. |
