Research On Image Classification Based On Laplace Mixture Loss Function

In recent years,under the background of big data,massive data has flooded into the Internet.It is very important to identify relevant information with the help of computer automatic classification.Image classification is one of the popular research directions in the field of computer vision,and it is also an important basis for applications such as automatic driving and face recognition.Improving the accuracy of image classification is also a hot topic in current research.This paper focuses on improving the classification accuracy of images from the aspect of loss function.Gaussian distribution is the most common distribution.In reality,many data follow Gaussian distribution,but there are still many data that follow the heavy-tailed distribution.Laplace distribution is a distribution with heavy-tailed properties.In this paper,through visualizing the features extracted by the convolutional neural network,we find that the features extracted by the convolutional neural network may also follow the Laplace distribution.So starting from the probability distribution of the features extracted by the deep neural network,we shaped the feature space into Laplace mixture distribution and applied it to the loss function and proposed Laplace Mixture(LM)loss function.This loss function includes two parts,one is the classification loss function,the classification loss function refers to the cross entropy between the posterior probability obtained after shaping the depth feature space into a Laplace Mixture distribution and the class label;the other part is the likelihood regularization term.In order to further reduce the distance between the feature and the class center and make the feature better represented as the learned Laplace distribution,this paper maximizes the logarithmic likelihood function value of the feature and at the same time scales to get the likelihood Regularization term.The classification loss part and the likelihood regularization term together form the two parts of the Laplace Mixture loss function.Furthermore,the symmetric feature space constructed by the LM loss function can be theoretically proved to be robust to adversarial attacks.Finally,this paper applies the Laplace Mixture loss function to the image classification of the deep neural network.The experimental results of the MNIST and CIFAR-10 datasets running on VGG11 and VGG13 show that the LM loss function has better performance than the Gaussian Mixture loss function.Therefore the LM loss function has better classification performance with a structure such as convolutional layers plus fully connection layer of the convolutional neural network.It is also proved that the LM loss function can promote high classification performance and accurate modeling of feature distribution.