Data Augmentation Model Based On Deep Generative Model

In recent years,the performance of Convolutional Neural Network(CNN)in image classification is superior.Many studies have proved that CNN can achieve high accuracy rate trained by large-scale labeled dataset.However,in some special fields,due to various restrictions,the number of samples in the training dataset is much smaller than the scale of the training dataset used in the normal case,which eventually leads to the overfitting of the training dataset.At present,data augmentation algorithms have been widely used in many fields.The data augmentation algorithm performs on the training set,improving the quantity and quality of the training dataset,and improves the performance of the network trained by the augmented dataset.This paper studies the fog weather situation map as a small dataset scene,and proposes two data augmentation algorithms based on improved generative models.1.Concern the problem that a novel data augmentation method called GMM-CGAN was proposed,which was combined Gaussian Mixture Model(GMM)and CGAN(Conditional Generative Adversarial Networks).Reparameterize the latent generative space as a mixture Gaussian model and learn the model's parameter with training of GAN;sample vector z from mixture model and fed it into CGAN;the outputs by trained CGAN obeyed the distribution of the sample.2.In order to deal with the difficulty of classes generated by learning with multi-class training datasets,the multiple latent space convolutional variational auto-encoder(MLSCVAE)is proposed.The whole latent space of the training dataset is divided into different categories according to classes of dataset,and then the latent space of the training dataset is learned and generated separately.The training dataset of lacking samples scene is constructed by extracting part of data from three datasets,i.e.fog weather situation map,MNIST and CIFAR 10.Using self-designed TestNet convolution neural network to test augmented datasets,GMM-CGAN and MLSCVAE models are evaluated based on the accuracy and over-fitting rate.The experimental results show that the accuracy of network classification trained by GMM-CGAN and MLSCVAE three augmented datasets is 87.5%,86.9%,84.3% and 86.3%,87.7%,85.4%,and the over-fitting rate is 1.25,1.13,1.28 and 1.15,1.18,1.25.It is proved that theGMM-CGAN and MLSCVAE models are significantly improved compared with the previous methods.