Research And Application Of Weed Detection Algorithm In Cotton Field Based On YOLOv3

Modern agriculture is gradually becoming intelligent and precise,especially in the field of weed automatic identification in the seedling stage of crops.In the growth and development period of cotton seedling,weeds encroach on the growth space of cotton seedlings,which will cause harm to the healthy growth of cotton seedlings.Without determining the exact location of weeds,large-scale pesticide spraying will not only cause waste of pesticides,but also pollute water quality,air and soil.At present,the recognition and classification and location of weeds are mainly accomplished by the feature extraction of weeds by deep convolution neural network.In the real-world scene,because of the different shapes of weeds,complex background,the influence of mutual occlusion between leaves and the real-time requirements of the algorithm in practical application,the target detection algorithm is still difficult to be applied in the actual scene.Therefore,in view of the characteristics of mutual shelter,similar shape and texture and soil on the surface of weeds in Xinjiang cotton field,this paper proposes a better optimized YOLOv3 network model based on the weed detection target in cotton field,and designs and completes the real-time detection system of miscellaneous grass with the core of the network model.In view of the practical application scene of weed detection in cotton field,this paper takes weeds near cotton seedling as the detection target in the process of moving,and adopts an improved feature extraction method,which is to apply the improved auto focus layer to the existing vgg16 and darknet53 feature extraction network model based on deep learning.By improving the feature extraction network,the improved YOLOv3 network model with high recognition accuracy and good robustness is trained.The four weed identification models were constructed by using Faster R-CNN network which takes VGG16 and VGG16 as feature extraction network and YOLOv3 network which takes Dark Net53 and improved Dark Net53 as feature extraction network respectively.Four models are tested on the marked PASCAL VOC test set,and the four network models are compared with the same data sets.The test results show that the network has a great advantage over several common network models and feature extractors.The system design of weed real-time detection and the design of the following autonomous spraying system are completed on NVIDIA Jetson AGX Xavier NX.Based on the optimized YOLOv3 target detection algorithm,the platform of cotton seedling and weed real-time detection system includes three modules: image acquisition module,real-time detection module and autonomous spray module.After the embedded hardware platform is built,the experiment is carried out.The experiment was divided into three groups: cotton seedling and weed image detection,cotton seedling and weed video detection and real-time detection.The experiment shows that NVIDIA Jetson AGX Xavier NX can meet the real-time requirements.In the actual scene,the weed detection is completed in real time,and the detection accuracy and detection time are improved.After weed detection,the signal can be fed back to the follow-up autonomous spraying system,so that the valve of spray head is opened to complete the independent spraying.The experiment shows that the embedded hardware platform can meet the needs of weed real-time detection and self spraying.