Research And Application Of Corn Seedlings And Weeds Detection Algorithm Based On Deep Learning

Accurately identifying harmful weeds in agriculture and efficiently using chemical pesticides at the minimum dose to achieve increased crop yield without causing adverse effects on the environment is an inevitable trend in future agricultural development.As one of the main sources of food in China,the healthy and abundant growth of maize seedlings plays an important role in food security and industrial production.This article focuses on the identification of grain crops such as corn and associated weeds,with associated weeds as the main research object.Based on the actual category characteristics and pesticide characteristics of weeds in the field,by improving the accuracy of target detection algorithms for identifying single family weeds and reducing the file size of exported models to adapt to mobile hardware devices with poor performance,Site specific weed management(SSWM)is achieved,Provide reference for achieving precise weed removal research in the future.The main research work of this paper is as follows:First,the data set is missing.Collect data under different environmental conditions during the best weeding period of corn from three to five leaves,and use data enhancement to increase the diversity of the image and improve the robustness of the target algorithm.Finally,more than 10000 images are obtained,which are labeled with Labelimg software and stored in YOLO and VOC formats;The classification of families and genera can effectively reduce the possibility of recognition errors of the same kind but with slightly different phenotypic characteristics.Second,study the corn weed detection algorithm model based on the single-stage target detection algorithm.Analyze the lightweight feature extraction network and the dataset,backbone,Neck,and head of the YOLO series to provide theoretical support for subsequent algorithm improvements.Using the corn weed test data set and the field weed data set taken in complex situations,the original YOLO-Fastestv2 and the lightweight YOLO-v4tiny,YOLO-v4 Mobile Netv3,YOLO-v5s were tested Through comparative experiments,the lightweight and real-time YOLO-Fastestv2 model was established as the application algorithm.Third,in view of the relatively low recognition accuracy of the YOLO-Fastestv2model for weeds and the lack of recognition of small target weeds,an improved algorithm for YOLO-Fastestv2 is proposed.Use K-Means++and Moscia enhanced methods to preprocess the input data to ensure the data stability of the model;use Soft Pooling to replace the original Max Pooling,and replace the convolution improvement with a step size of 2 in the downsampling stage of the backbone with SPD volume In order to facilitate the model to extract features;add Sim AM non-parameter attention mechanism to the FPN feature fusion part,so that the model can ensure the improvement of recognition performance without significantly increasing the calculation burden;use Focal Loss and SIOU Loss to improve the original loss function,Make the algorithm converge more smoothly and have good performance on objects that are difficult to recognize;use the Nadam optimizer to constrain the change of the learning rate,so that the fitting can be closer to the optimal point more quickly.The improved YOLO-Fastestv2 network proposed in this paper,while achieving the recognition effect of YOLO-v5s,the model file is only increased by 0.87M compared with the original YOLO-Fastestv2.The results show that the detection precision and recall of the improved YOLO-Fastestv2 model are 93.97%and 90.02%,respectively,which are 1.76%and 19.69%higher than before the improvement.The F1 value is 10.28%,5.55%and 0.29%higher than the YOLO-v4tiny,YOLO-v4Mobile Netv3,and YOLO-v5s models,respectively,and about 12.16%higher than the original model.The m AP value is 5.57%,3.56%and 0.03%higher than that of the YOLO-v4tiny,YOLO-v4Mobile Netv3,and YOLO-v5s models,respectively,and about6.36%higher than the original model.The output model size is only 22.8%,47.9%,36.4%of the three network models.In terms of detection speed,YOLO-Fastestv2 pursues the fastest speed and has fewer parameters.After the improvement,the recognition accuracy has improved,but the corresponding model file size has also increased,and the detection speed has also slowed down to 200 frames/s.,45 frames/s slower than the original.However,compared with several other lightweight YOLO target detection models,it is still79frames/s,65frames/s,and 75 frames/s higher.Experiments show that the m AP of this method is 94.28%.In the case of data with complex background,high light,small target,rain,shading,occlusion,and the same subject but not in the training data set,different algorithm models were tested,and the improved YOLO-Fastestv2 has good generalization effect.Fourth,apply research on the corn weed algorithm model for the classification of families and genus.Choose selective chemical herbicides as the spraying liquid,industrial digital cameras as the information collection equipment,NVIDIA AGX as the processor for precise identification,STM32 as the controller for variable spraying,and pressure nozzles with conical nozzles as the spraying mechanism,the precision spraying operation experiment on field weeds.The recognition accuracy of corn was 96.5%,and the recognition accuracy of broad-leaved weeds and gramineous weeds were 95%and 85.7%,respectively,which verified the recognition performance of the model for corn weeds.In the spraying effect of this experiment,the single-category weed pesticide droplet coverage density is 45/cm~2,and the multi-category small weed pesticide droplet coverage density is87/cm~2,and the multi-category quantity is large.Too much droplet coverage density is difficult to verify,and it can actually meet the control requirements of herbicide variable spraying.The research provides solutions to alleviate the problems of pesticide residues,pesticide drift,excessive use of pesticides,and the reduction of agricultural management personnel.