Research On Adversarial Generation Network Based On Similarity Evaluation

Generative Adversarial Networks is good at fitting high-dimensional and complex data in the real world and synthesizing samples from learned distribution.It also provides a method to learn the deep representation of training data without additional label information.This kind of generative models receives good results and could be widely used in image synthesis,style transfer,super-resolution,image translation and image Colorization.As the Generative Adversarial Networks is usually composed of two networks,its training process won't stop until the two networks reach the Nash equilibrium.However,the training process is very unstable and very easy to fall in mode collapse,which greatly limits its applications and the efficiency of research.In this manuscript,we explore and study the original Generative Adversarial Networks and its varieties,observe the training process of different network models,and try lots of fantastic tricks.In order to deal with the instability of Generative Adversarial Networks in the training process,this manuscript describes a similarity evaluation module,which is based on the consistency of similarity between latent space and data space,and sets up the loss function to optimize the generator.Thus Similarity Estimation Based GAN(SEBGAN)and relevant training strategy are proposed.In the experiment section,the public available datasets are used to demonstrate the effectiveness of the network model and relevant strategy of training proposed in this paper.The Inception Score,Frechet Inception Distance and other objective evaluation indicators are used to compared our work with other generative models,combined with sample observation and visualization via Uniform Manifold Approximation and Projection.The purpose of stabilizing the training process of Generative Adversarial Networks and effective resistance to mode collapse is achieved.At last,the Similarity Estimation Based Generative Adversarial Networks are discussed and the promising research directions in the future are prospected.