Research On Cotton Top Buds Recognition System Based On Deep Learning

Cotton topping as a key link in cotton field management,and the quality of cotton topping will affect cotton yield.The traditional manual topping method has problems of low work efficiency and high topping cost;mechanical topping has low accuracy and effect,which affects cotton production;due to the abuse of chemicals and the inability to accurately spray and other factors,chemical topping has not been widely promoted.Therefore,the study and improve of low identification accuracy for cotton top buds can improve the topping accuracy by cotton mechanical topping or chemical topping.With the development of agricultural informatization,it has become feasible to achieve accurate identification and positioning of cotton top buds with the deep learning methods.In this study,aiming at realizing the accurate identification of cotton top buds,a cotton top buds identification system is proposed,based on deep learning method.The specific research work and main conclusions are as follows:(1)For the cotton top buds in the natural environment,through picture taking,used the data augmentation method to expand the number of pictures,label Img tool was used for labeling,a dataset of cotton top buds in natural environment in VOC format was prepared;in order to further improve the level of cotton topping,defined the concept of cotton top buds at different maturity stages,according to this concept,through picture taking,used the data augmentation method to expand the number of pictures,label Img tool was used for labeling,a dataset of cotton top buds at different maturity stages in VOC format was prepared.(2)Based on the cotton top buds dataset in the natural environment,the detection performance of YOLOv3,YOLOv4,SSD and Tiny-YOLOv4 algorithms in cotton top buds was compared,The results show that: YOLOv4 algorithm has the best performance;in addition the detection performance of YOLOv3,YOLOv4,SSD and Tiny-YOLOv4 algorithm for cotton top buds under occlusion and backlight conditions was also researched.(3)Based on the cotton top buds dataset at different maturity stages,an improved YOLOv4 algorithm is proposed to solve the problem of low classification and recognition accuracy of the YOLOv4 algorithm.By replacing the Yolo Head module with the Decoupled Head module,the Anchor Free idea is introduced,and the anchor free model is used to instead of the anchor based.The results show that: the m AP value of the improved YOLOv4 algorithm is 90.15%,which is 2.93%,5.95%,13.6%,and 2.17% higher than that of YOLOv4,SSD,Faster-Rcnn,and Tiny-YOLOv4.(4)By investigating cotton plant growth information and the principle of machine vision was uesd,the high ground clearance unmanned vehicle with a ground clearance of1140 mm was determined as the system platform,designed and built the camera mount,combined with the deep learning algorithm,Model deployment was completed,hardware equipments were selected and debugged,and software system such as system environment configuration and camera calibration were completed.(5)The field verification test of cotton top buds identification system was conducted.The influence of different light intensities on the recognition of cotton top buds was explored;the influence of different image resolutions on the cotton top buds recognition system was explored,and the feasibility of the system was tested.Aiming at the problem of low identification accuracy of cotton top buds in cotton topping,this study proposes a cotton top buds identification method based on deep learning,and designs and builds a cotton top buds identification system based on a high-clearance vehicle,which lays a solid foundation for the future research on cotton intelligent topping.