Research And Application Of Generative Adversarial Networks Based On L2 Norm Constraint

Deep learning has brought another development craze to Artificial Intelligence and generative adversarial networks is a new deep learning technology that has emerged in recent years.The successful applications of generative adversarial networks in image generation,speech enhancement,video prediction,etc.have attracted much academic and industrial attentions.At present,the main challenges faced by generative adversarial networks are stable training and finding Nash equilibrium.In image generation tasks,these challenges are solved in order to generate higher quality images.At present,the commonly used method for stabilizing training introduces Lipschitz constraint to the discriminator,but when used with spectral normalization,it suffers from unstable quality of generated images.At the same time,some heuristic methods are also used to improve the convergence of generative adversarial networks.These methods aim to generate high quality images by finding Nash equilibrium,but fail to achieve expected results in practice.In order to enhance the quality and stability of the images generated by generative adversarial networks,this paper conducts an in-depth study of generative adversarial networks in terms of constraining the generator and the discriminator based on L2 norm.This paper chooses the L2 norm constraint instead of the other norm constraints because: compared with other norms,the L2 norm is more conducive to the weight update;in addition,the consideration of the generation of the adversarial network for each sample should be considered from the same perspective,L2 norm is more suitable to deal with this situation than other norms.The main work of this paper is as follows:1)The generator constraint method based on L2 norm is studied.In order to solve the unstable quality of images generated by the spectral normalization method,this paper proposes an L2 norm constraint method for the generator.This method makes the image quality generated by the spectral normalization method more stable by subjecting the difference between the generated data and the real data to the L2 norm.In the actual training process,this method can bias the generator to generate a batch of images with a small distance between the Euclidean space and the real image.In this paper,the method is compared and verified on the public datasets.The experiments show that the L2 norm constraint on the generator can stably improve generation quality,and to some extent,generate images with higher quality.2)The discriminator constraint method based on L2 norm is studied.Inspired by Nash equilibrium,this paper proposes a method for constraining the discriminator with L2 norm.This method makes the network converge better by constraining the relative output value of the discriminator.Although the constraint and the previous constraint are both L2 norms,the mathematical form and starting point of the constraint are different.Constraining the discriminator helps it converge better to generate high-quality images.Experiments on public datasets show that this method can effectively improve the quality of generated images.Even if the generative adversarial network cannot find the Nash equilibrium,the method still has advantages under certain conditions.3)The verification of the proposed method is carried out.In order to further verify the effectiveness of the proposed method,in addition to verifying the effectiveness of the proposed method on public datasets,the paper also established a design patent dataset and used different methods to perform image generation experiments on this dataset.The results show that this method is also validated on the specific dataset.