Research On Imbalanced Image Classification Method Based On Generative Adversarial Networks

In the fields of machine learning and data mining,datasets are often imbalanced,which negatively affects the accuracy of classification algorithms.Creating data for the minority classes to rebalance the dataset is a general and effective solution to address this problem.For low-dimensional data,there are already a variety of classical algorithms that can perform efficient data augmentation.For high-dimensional image data,image transformation methods can be used to oversample the minority class images.To some extent,such methods can alleviate the negative impact of class imbalance problem,but the effect is very limited.As a powerful generative model,Generative Adversarial Networks(GANs)can capture the true distribution of data and directly generate high-quality synthetic data.To improve the accuracy of deep learning classifiers,we augment imbalanced image datasets by using GAN models.The main contributions are as follows:(1)We propose Discriminative Information Optimization Conditional Capsule GAN(DIO-CCaps GAN).The discriminative ability is improved by using the capsule network in the discriminator of the GANs and by adding the conditional constraint information to the capsule discriminator.In addition,when the training data is imbalanced,the images generated by the GANs are biased towards the majority classes.To address this problem,we optimize the output of the discriminator to make it more sensitive to the class information,which improves ability of the generator to synthesize images that match the minority classes.(2)We propose Generator Information Enhancement GAN(GIE-GAN).The generator of the GIE-GAN,which is initialized by using the pre-trained Variational Autoencoder,can get common knowledge of all classes of data before adversarial training,so that the stability of training can be enhanced.In addition,during the training process,an independent classifier is added to the discrimination part of the GIE-GAN to provide the generator with additional classification information,which effectively alleviates the problem that the generated data is biased towards majority classes.The experimental results show that on a variety of imbalanced image datasets,the proposed methods can generate high-quality diverse image samples.Compared with other considered algorithms,the performance of the deep learning classification algorithms significantly improves on the datasets augmented by our methods.