Research On Target Recognition Method Of Weeding Robot Based On Lightweight Network Model

During the seedling stage of crops,weeds are important hazards that affect the normal growth of crops,especially competing with crops for soil nutrients,resulting in poor crop development and yield reduction.Most of the current weeding methods are manual weeding and chemical weeding,which have their own drawbacks,such as low weeding efficiency and chemical pollution.Therefore,the study of robotic weeding technology has the practical significance of solving manpower and reducing environmental pollution.In this thesis,the problem of identifying corn seedlings and weeds is investigated with robotic corn seedling field weeding.The current network model based on deep learning has more parameters,which leads to its poor real-time performance on weeding robots and makes it difficult to meet the actual weeding work requirements.Moreover,the target detection network is poor in recognizing corn seedlings and weeds of the same color system and has the problem of missing detection.To this end,this thesis proposes a target recognition method for weeding robots based on a light-weight network model,and deploys and validates it in practice on an embedded weeding robot.The main research work is summarized as follows:(1)A corn field weed dataset was constructed for the recognition scenario of the weeding robot in the field for the training and experiment of the model.The dataset collected images of growing corn seedlings and weeds in three directions: overhead view,40° oblique view and 75° oblique view,as well as in sunny,rainy and cloudy day environments,to improve the adaptability of the corn weeding robot in the field environment.At the same time,the collected images were screened,cropped and data tagged to provide data support for subsequent research.(2)A real-time target detection algorithm GBC-YOLOv5 s based on a lightweight network model is proposed,which well balances speed and accuracy while significantly reducing computational cost.First,an improved Ghost Net network architecture is used to refine the weed features generated by the backbone network.Second,a feature fusion strategy based on the convolutional attention module CBAM mechanism is proposed to improve the detection speed and detection accuracy.Finally,experiments are conducted on the dataset with the proposed model size of 3.3 MB,and the detection time of the input image reaches15.6 ms with an average accuracy of 96.3%.Comparing the experiments with the existing Faster RCNN,YOLOv4-tiny and Yolov5 s target detection models,the model size is reduced by 98.95%,90.3% and 77.39%,the m AP is improved by 5.1%,3.9% and 2.1%,and the detection elapsed time is reduced by 142.9ms,36.9ms and 85.7ms,respectively.(3)An embedded weeding robot recognition system was developed,and the effectiveness of the system was experimentally verified.The system was used to conduct comparison experiments between YOLOv5 s and the GBC-YOLOv5 s target detection model proposed in this thesis in a variety of complex environments in the field,and the results showed that the average FPS before the improvement was 11 and the average FPS after the improvement was 53.5,indicating that the improved detection speed was significantly improved compared with the speed before the improvement,and the improvement of the missed detection phenomenon in a variety of environments was obvious,which provides a weeding It provides technical support and practical basis for the development of robot technology.