Deep Learning Based Crop And Weed Identification In The Field

Field crop and weed recognition is a key technology for automated weed control equipment.In recent years,with the rapid development of deep learning,target detection techniques applying convolutional neural networks have performed well and gradually become the main method for crop and weed detection,and field weed recognition based on deep learning has become a research hotspot in this field.Using the advantages of deep learning that can extract crop weed features autonomously and strong generalization,we carried out the research of Deep Sort＿Yolo V5 network recognition model based on the fusion of attention mechanism of deep learning,and the main research results achieved are as follows:(1)As a common cereal grain crop,maize has become the main data object of field mechanized weed control trials because of its short growth cycle and high impact of weeds on seedling growth.By reviewing the weed control data of seedling maize at home and abroad,and combining the weed control characteristics of the crop in different growth periods,the three-to five-leaf stage maize was selected as the research object,and the seedling maize data set was collected and produced.In the design of the identification system,weeds were detected indirectly by identifying seedling maize to improve the real-time detection accuracy in order to reduce the problems of high detection complexity,poor detection accuracy and robustness caused by a wide range of weed species and complex environment in the conventional target detection process.(2)An improved attention mechanism fusion Deep Sort＿Yolo V5 network recognition model was proposed to combine the attention mechanism,Deep Sort multi-target tracking,and Yolo V5 recognition network,and introduced an efficient channel attention mechanism after each convolutional layer to improve the weed classification and recognition accuracy.In the optimization of multi-target tracking algorithm based on Yolo V5 fused with Deep Sort,the current mainstream multi-target tracking strategies are studied,the tracking of traditional target tracking algorithms and multi-target tracking algorithms were compared,and finally the framework of multi-target tracking strategy method based on target detection was selected.Using YOLO series detectors,Yolo V3,Yolo V4 and Yolo V5 were compared,and Deep Sort＿Yolo V5 model was selected as the main framework due to the light weight and high efficiency of Yolo V5.(3)The field environment conditions are harsh and complex,and there are many kinds of weed distribution.The traditional weed detection methods required manual design of features,which had the defects of complicated operation,slow detection speed,insufficient recognition accuracy and poor robustness,and were not adapted to the field real-time operation scenarios.Applying the field weeding robot testbed built by the project team,we analyzed and evaluated the feasibility of the proposed improved attention mechanism fused with Deep Sort＿Yolo V5 network recognition model for real-time detection in the field.The experimental equipment was built on the basis of optimized software configuration and hardware equipment,and the performance of the system was tested by laboratory simulation weeding experiment and realtime detection test under natural conditions in the field.The improved recognition model achieved a model size of 67.2 MB.In the experimental test,the time to complete the real-time detection target is 60 ms,and the recognition rate is 96.13%.Experimental results showed that that the improved recognition model meets the recognition requirements of automatic weeding.