Data Augmentation Based On Generative Adversarial Networks

The three elements of AI are data,computing capability and algorithm.Data is the basis which is essential to any research.The rapid development of deep learning algorithm is not only because of the acquisition of large-scale data becomes a reality,but also thanks to high computing capability that makes large data models be well fitted.As the most potential branch in the field of artificial intelligence,deep learning relies on a mass of data.And the size of data sets directly affects the performance of deep learning model.However,it always requires a large amount of real image,text or voice data when training a neural network while these data need to be manually marked with some classified identification information.This takes a lot of time and human cost.This paper put forward using GAN(Generative Adversarial Network)this method to generate the image data extend dataset,especially aim at solving the over fitting problem due to the lack of the dataset when training the neural networks in the study of image classification and recognition.The main topics of this thesis are as follows:1.An improved DCGAN model SDCGAN is proposed,which is based on the activation function SeLU and a_Dropout.This model not only simplifies network parameters,but also alleviates the overfitting phenomenon of DCGAN to some extent.2.The instability of image quality generated during training of existing DCGAN models is studied in depth,this thesis put forward a dynamic label smoothing.This method not only adjusts the intensity of confrontation training of generation network and discrimination network,but also better alleviates the problem of severe fluctuation or even divergence of DCGAN's loss in training to a certain extent,which makes the parameter value of label smoothing better match the characteristics of DCGAN training.3.Deep learning framework tensorflow platform was used to verify the SDCGAN model in this paper.Objective evaluation indexes showed that the model not only generated higher-quality image data,but also improved the recognition accuracy of the classification network trained with generated data.In addition,the simulation results also verify that the dynamic tag smoothing method in this paper can alleviate the loss function curve fluctuation or even divergence in the training of DCGAN network model to a certain extent.