A Generative Adversarial Network-Based Approach to Solving Model Collapse Using Perceptual Loss

Abstract

The invention belongs to the technical field of image reconstruction, and discloses a method for solving model collapse based on generative confrontation network using perceptual loss. A random vector z is used to generate an image similar to the distribution of actual data. In the training process, perceptual loss is used to compare z with real The data is mapped into the feature space to extract higher-level features, and combined with the adversarial loss to encourage the generation network to generate image samples similar to the actual image; finally, the discriminator cannot judge that this is a fake image. The present invention solves the problem of model collapse by adopting a smaller data set for the existing network, and VGG-GAN is evaluated on two small scene data sets; the experimental results show that the image quality generated by the VGG-GAN method is better than the existing method .

Description

Technical field [0001] The present invention belongs to the technical field of image reconstruction, particularly to a network based on the generated against loss method using a perceptual model to solve collapse. VGG-GAN particularly to a method of using a perceptual model collapse loss generated against network-based solution. Background technique [0002] At present, commonly used in the prior art is this: [0003] GAN has generated against traditional network builder network and discriminator networks. Training GAN network and the generator discriminator competing networks to achieve optimal process, namely to achieve Nash equilibrium. However, there is no clear way to determine whether GAN network has entered a Nash equilibrium. The problem is a high-dimensional non-convex optimization objectives. In the next step the network tries to minimize non-convex optimization goals, which may eventually lead to oscillations, rather than converge to the real underlying goal. As long a...

AI Technical Summary

Problems solved by technology

[0007] (1) At present, there is no clear method to determine whether the GAN network has entered Nash equilibrium;

[0008] (2) GAN is widely used, but there are problems such as unstable training, gradient disappearance, and model collapse. The results of the experiment will be poor, and the final result will not be improved even if the training time is increased;

However, when the training data is a small scene dataset, the experimental results produce mode collapse; resulting in a loss of diversity in the generated samples

Images