Recognition Of Common Pests In Agriculture And Forestry Based On Deep Learning

China is a largely agricultural country.In the process of cultivating crops,it encounters different kinds of pests every year,which causes the crops to decrease in yield and quality.When the disaster is serious,it will even lead to a large area of crops.Accurate and effective classification and identification of insects are an important prerequisite for timely pest control and avoiding huge economic losses of crops.Insects are the most diverse species of animals in the natural environment and are difficult to identify.Traditional insect classification and identification work rely mainly on insect expert to identify.They identify according to professional knowledge and research experience or reference literature.But even with professional knowledge and rich experience,it is difficult to avoid category confusion.Therefore,the development of a rapid and effective classification and identification system for pests will contribute to the prevention and control of crop pests,thereby promoting agricultural development and reducing economic losses.In order to achieve rapid recognition of the common pests in agriculture and forestry,an automatic pest identification method based on deep learning is proposed in this paper.The main work and research results of this paper are shown as follows:(1)This article collected RGB images of 27 adults and 5 larvae.In order to avoid the serious problem of uneven label distribution when the neural network divides the training batch,the data were enhanced by image processing methods such as cropping,distortion,color adjustment,and background change.The constructed image dataset CPAF has a total of 32,000 insect images,of which 22,400 are used as training sets and 9600 are used as validation sets.(2)VGGNet was optimized based on the CPAF data set.After a series of experimental studies,it was found that ReLU activation function had the best effect.The model RDIII can achieve 99.02% recognition accuracy after training 10,000 steps,and the Adam method is the better choice for the optimizer.The suppression of the overfitting problem is most effective when the Dropout probability is set to 0.5.(3)According to the research results of VGGNet model optimization,an effective deep neural network model CPAFNet is designed inspired by the GoogLeNet.The model has better performance in both recognition effect and training rate.After 6 thousand steps,the recognition accuracy rate reached 99.06%.Under the same number of training steps,CPAFNet training consumes less time and the recognition accuracy is preferable to model RDIII.(4)In the study of the hidden layer feature extraction visualization of VGGNet,DRIII,and CPAFNet models,the outstanding performance of CPAFNet in color and texture feature extraction further proves the feasibility and effectiveness of the model.The results of the model optimization research and the CPAFNet depth model proposed for the CPAF dataset have a good practical significance for the intelligent identification of agricultural and forestry pests.