Research On Detection Method Of Storage Pests Based On Deep Learning

China is a large country,a large population and a large grain producing country.In the process of grain storage,the annual grain loss is about 0.2% to 0.5% of the total grain storage,in which the damage caused by the stored grain insect damage accounts for about half of the total loss.Therefore,to reduce the losses caused by pests,pest prevention and control has become the key technical problems of food security in China,and the detection and recognition of grain pests has become the key problem of pest control.Based on the study of stored grain pests at home and abroad,grain pest detection methods include sampling,lure,acoustic,near infrared,image recognition and image recognition method.Because of its high recognition rate,simple operation,low cost and other characteristics,image recognition method has become a research hotspot and main technical means in the field of pest control.Most of the traditional image recognition feature extraction is carried out manually,which has many limitations and shortcomings.At the same time,image recognition and classification based on deep learning has become a hot research topic at home and abroad.Deep learning through the bionic methods,use the artificial neural network to simulate human visual system,and automatically learn the characteristics of an image in an unsupervised manner.Therefore,depth learning can significantly improve the accuracy of image recognition.Aiming at the problem of stored grain insect detection,this paper explores the detection and recognition method of grain pests based on deep learning.The main research work is as follows:1.Compared with other methods such as artificial neural network,BP neural network and so on,this paper systematically studies the methods of deep learning,such as sparse automatic coder,restricted Boltzmann machine,deep belief network,convolutional neural network and so on.By analyzing the convolution of the neural network model,structure,algorithm and application evolution,this paper provides the support for the detection and recognition of the grain insect image based on deep learning.2.Are beetles(Curculionidae,Trogossitidae)and butterfly(Butterflies and Danaus plexippus)data were collected.The corresponding pest database is established.A 5-layer convolution neural network model(1 input layer,2 convolution layers,2 full connection layers)is established,which uses Sigmoid as the activation function and uses the mean square error(MES)as the loss function.By the experimental results,this paper analyzes the problems and shortcomings of the detection and recognition method of the grain pests based on the 5-layer convolution neural network model.3.For the training of small sample set model does not have the question of generalization ability,this paper proposes a method for constructing the sample set of grain pests based on image warping technology.Three kinds of image enhancement techniques,such as image scale transformation,image rotation and image elastic distortion,are used to construct the training set of the insect pest image.Experimental results show that the convolution neural network model with image warping technique has better generalization ability and better detection and recognition effect.4.For the training of shallow convolutional neural network model does not have the characteristics of complex expression ability,a neural network model based on depth convolution is proposed in this paper.A 7-layer convolution neural network model(1 input layer,2 convolution layers,2 pool layers,2 full connection layers)is designed,which uses the ReLU as an activation function take softmax + cross-entropy as the loss function,uses the deep learning Caffe framework.The experimental results show that the proposed method can significantly improve the acquisition ability of complex features without increasing the training cost.The detection rate of the beetle was as high as 95%,and the recognition rate of butterflies was increased by 20%.