Study On Cotton Seedling Recognition Based On Deep Learning

The planting mode of modern agriculture is gradually mechanized,large-scale and intensive.Although the labor cost and cycle time are saved to a certain extent,the strain of field seedlings caused by mechanical operation and the inability of entropy sowing still reduce the emergence rate of cotton seedlings.As the largest high-quality cotton base in China,the planting area in Xinjiang accounted for 85% of the country in 2021.In order to control the high yield and excellent quality of cotton,increase agricultural income and improve the management of fine cotton field,scientific seedling identification and positioning is imperative.Therefore,the key technology of real-time monitoring of cotton field is accurate and rapid identification and positioning of seedlings.The main research contents and results of this thesis are as follows:(1)Taking the experimental field covered with plastic film as the research area,an improved yolov5 algorithm was proposed to identify and locate Cotton Seedlings under a simple background,which solved the problems of high cost and slow aging caused by traditional manual methods.the data was enhanced by mosaic.In this thesis,senet attention mechanism was introduced into yolov5 s,which is defined as yolov5 sSE algorithm.The improved yoov5 s embeded the attention senet mechanism into the CBL module in the backbone to help the model pay more attention to the target of interest,extract deeper semantic features,train five different depth network models of yolov5 s,yolov5m,yolov5 x,yolov5l and yolov5s-SE,and compare the network performance indicators.The experimental results showed that the overall performance of yolov5s-SE was the best,with an accuracy of 87.95% and a regression rate of 87.57%,It could quickly identify and accurately locate seedlings.(2)The experimental field without plastic film was defined as a complex background.In view of the problems of weed occlusion and blurred image in the growth process of open field cotton seedlings,this thesis combined convolution neural network to realize the recognition and location of seedlings under complex background.Input cotton seedling image in the network,and four feature extraction networks including VGG16,Resnet50,Mobilenetv2 and Resnet50 fused with FPN were used to build a cotton seedling recognition and location model based on improved faster r-cnn.According to the size characteristics of cotton seedlings,the anchor frame technology was designed to meet the size of cotton seedlings,and ROI align was used to reduce the coordinate deviation caused by quantization operation and improve the segmentation effect of the model.The experimental results showed that the average recognition time of the improved faster r-cnn model for a single image was 0.289 s and the average accuracy is 89.19%.In complex background,it could accurately recognize small-scale targets.Through feature fusion,the detector can have better recognition performance for blurred and occluded targets.(3)In order to compare the fitting performance of the two improved methods under natural conditions and find the optimal algorithm suitable for the actual situation,the cotton seedlings under simple background and complex background are fused in the data set.Based on the method of migration learning,nine convolution networks,yolov5 s,yolov5m,yolov5 x,yolov5l,yolov5s-SE,VGG16,Resnet50,Mobilenetv2 and faster r-cnn of Resnet50 with FPN,were compared and analyzed.The results showed that,The accuracy rate of yolov5s-SE was 90.83%,the recall rate was 92.16%,and the reasoning speed was124 fps.While maintaining the small target detection accuracy,it still improved the reasoning speed,which was more in line with the positioning and recognition of Cotton Seedlings under natural conditions.In this thesis,cotton seedlings in 2 ～ 4 leaf stage are taken as the research object,combined with UAV remote sensing technology and a variety of deep learning methods to realize the identification and positioning of cotton seedlings,so as to provide technical support for subsequent cotton field management and fine plant protection.