A Data Augmentation Method For Image Class Imbalance Problem Using Generative Adversarial Networks

Class imbalance problem is a common problem affecting the learning progress in machine learning field.An effective way to solve class imbalance problem is to conduct data augmentation on minority classes.For low-dimensional data,there have been many classic techniques to do data augmentation in an effective way.But those techniques may not be good for high-dimensional data like images.Oversampling methods with image transformation are widely used to do data augmentation on image data sets.Those methods can reduce the negative influence caused by class imbalance problem but often have their limits.Generative adversarial nets(GANs)have got a wide range of attention and researches in recent years as a class of neural network models.Given a class of real samples,GANs are able to generate synthetic samples which are similar to the real ones.Inspired by the characteristics,we used GANs to do data augmentation on image data in order to reduce the influence caused by class imbalance problem in classification tasks.We also analyzed the underlying noise in the synthetic samples and proposed our solutions.The main contributions of this dissertation are as follows:1)Designed the CycleGAN suitable for image augmentation tasks by combining the design philosophy of CycleGAN with residual blocks.2)Augmented minority images with CycleGAN,and thus the abilities of the classifiers were improved in the image class imbalance problem.Comparative experiments were also carried out to reveal how the numbers of synthetic samples influence the classification results.3)Analyzed the underlying noise problem existed in the synthetic samples from CycleGAN.The possible influences to the classifiers caused by the underlying noise were also revealed.4)Proposed a method to train images while keep the classifier unchanged,which can reduce the underlying noise in the synthetic samples.This method improved the classification accuracy and stabilized the results when the experiments were repeated.