Classification And Recognition Of Crop Seedlings And Weeds Based On Attention Mechanism

The production of crops is accompanied by the growth of weeds,which affects the yield,quality and economic benefits of crops.In the seedling stage of crops,weeds and crops have similar similarities in morphology and color,which brings difficulties for automatic identification and weeding of machines.In the preparation of weeding,the specific identification of crop and weed species is a guide to the development of pesticide spraying schemes.Aiming at this problem,this dissertation proposes a method for classification and identification of crop seedlings and weeds with attention mechanism,which effectively solves the problem of crop seedlings and weeds identification,and serves as a reference for precise weeding work in the later stage.The specific work is as follows:(1)Research on image classification methods for crop seedlings and weedsThrough the analysis of some mainstream algorithms for image classification,such as KNN,SVM,BPNN,CNN,etc.,the datasets are preprocessed for the characteristics of weed and crop seedling datasets,and 12 algorithms are used for these algorithms.The crop or weed data set is trained,and the training results are compared and analyzed.From the perspective of average classification accuracy,the method most suitable for the data set of this dissertation is the ResNet-50 model in CNN,and the classification accuracy rate reaches 91.89%..(2)Combining the attention mechanism and the cross-layer bilinear pooling algorithm to improve the CNN modelAccording to the characteristics and deficiencies of the ResNet-50 model,the ResNet-50 model is improved accordingly.The fine-grained classification algorithm under weak supervision is added to the model.The cross-layer bilinear pooling enhances the features and makes the network stronger.Ability to express.In the new network,the attention mechanism was introduced and the data set was trained.The result of classification accuracy was improved by nearly 4 percentage points based on the ResNet-50 model,reaching 95.62%.The accuracy of each type of crop or weed is derived in the new network.In this dissertation,based on the characteristics of weeds and crops in the early stage of growth,different image classification methods were used to train the weeds and crop seedling datasets.From the comparative analysis of training results,the classification methods and models most suitable for the datasets were obtained.Then introduce the attention mechanism and fine-grained classification in the model to extract more features in the image,and predict the actual effect of the combined model of ResNet50+attention mechanism+ cross-layer bilinear pooling,and finally realize 12 kinds of weeds.The accuracy of the respective categories of data with crop seedlings.It has certain guiding significance for precision weeding in agriculture.