Image generation method based on a conditional capsule generative adversarial network

Abstract

The invention discloses an image generation method based on a conditional capsule generative adversarial network. The image generation method comprises the following steps: collecting image data and preprocessing an image; Extracting a condition vector; Acquiring random noise; Designing a neural network as a generator to map the random noise and the condition vector into a generated picture; Designing another neural network as a discriminator to receive the generated picture, the real picture and the condition vector to obtain a loss value; When the adversarial network is trained, adjusting the weight of the generator network according to the loss value minimization objective function; And after the training is completed, adjusting the weight of the generator network to be optimal, abandoning the discriminator at the moment, and reserving the generator model as the optimal neural network of the generated image. The discriminator structure is designed by using the capsule neural network, the network can effectively avoid the gradient disappearance problem in combination with the advantages of the existing WGAN and CGAN, and meanwhile, the generated sample is high in quality and hascertain advantages compared with an algorithm in the prior art.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image generation method based on a conditional capsule generation confrontation network. Background technique [0002] The generative modeling method based on the differentiable generator network is currently the most popular research field, but due to the complexity of the real sample distribution, there are many problems in the stability and generation efficiency of the GAN generative model. Therefore, the industry urgently needs generation algorithms that can provide stable and high-quality samples. [0003] GAN theory is based on a game-theoretic scenario where the generator network learns to transform from some simple input distribution (usually the standard multivariate normal or uniform distribution) to a distribution in image space by competing with an opponent—that is, more and more The more realistic samples; as an adversary, the discriminator network tries...

AI Technical Summary

Problems solved by technology

[0005] Traditional GAN ​​usually uses a convolutional neural network to build an adversarial structure. This structure not only completes the detection of important features in image pixels through techniques such as sparse weights, parameter sharing, and pooling, but also greatly reduces network parameters. However, the internal data representation of CNN does not take into account the important spatial hierarchy between simple and complex objects. In a recent study, a dynamic routing-based capsule neural network was applied to learn the spatial structure relationship between input features.

Images