Pest Image Identification Method Based On Bayesian Convolutional Neural Network

The prevention and control of garden pests are one of the most important tasks in ecological environment protection.With the development of graining for green and urban greening policies,people’s requirements for the living environment are constantly improving,and the prevention and control of garden pests are particularly important.There are a variety of garden pests,and the cost of manual identification of the pests is quite high and difficult to implement.With the deployment of computer vision technology,many methods have been applied to identify garden pests images,and have achieved good results in the task of pests identification and classification,including some traditional image feature extraction methods,such as bag-of-words(bag-of-words,BOW)and SIFT(Scale Invariant Feature Transform)feature detection e.t.In recent years,with the development of deep learning,the Convolutional Neural Network(CNN)has gradually come into our sights and is widely used in various classification tasks.Due to CNN can extract the abstract features in the image by itself,and extract representative and abstract features by training,so CNN has a strong expressive ability and achieves better classification result.The weight parameters in the traditional CNN are shown as a point estimation,e.i all the weights are any certain values,and they are updated by methods such as backpropagation e.t.And prevent overfitting by dropout and regularization techniques during training,it has achieved good results in practical applications.However,from the perspective of probability learning,traditional CNN also has a little drawback.For example,a small number of datasets will lead to overfitting,and CNN will be overconfident for data in categories that are not trained in the training set.The classification results,the uncertainty in the training datasets cannot be well evaluated correctly,and lack of the generalization ability.This paper proposes a Bayesian Convolutional Neural Network(BCNN)based on the Bayesian method to classify several common garden pests images.Each weight parameter in traditional CNN is a form of point estimation,e.i the weight is a certainty value.But in BCNN,each weight is initialized in the form of a Gaussian distribution and approximated to the real posterior probability distribution and trained the network by variational method,and update the parameters of each distribution.In order to reduce the parameter quantity of the model and the training cost,the paper also proposed a method to prune the model parameters.Since each weighs parameter in BCNN is initialized in the form of Gaussian distribution instead of point estimation.The resulting model is equivalent to the average of multiple models,thus improving the generalization ability of the network also avoids the overfitting.Finally,we compare different method on the same pests dataset,and the experimental results show the effectiveness of our method.