Generative adversarial network method fusing self-attention mechanism

Abstract

The invention relates to a generative adversarial network method fusing a self-attention mechanism, belongs to the field of computer vision, and particularly relates to a generative adversarial network for image generation. The generation of the image is an important challenge in the field of computer vision, and if a large number of high-quality image samples can be generated, the artificial intelligence field can be more rapidly developed in the era depending on the big data background. Therefore, the invention provides the generative adversarial network fusing the self-attention mechanism,the network can generate the high-quality images, and meanwhile, the images have higher diversity. Specifically, the generative adversarial network uses a Waserstein distance to measure an evaluationcriterion of the generator and discriminator distribution, and a loss function is correspondingly improved; meanwhile, the self-attention mechanism is introduced into the neural network architecture corresponding to the generator and the discriminator, so that the relevance between the local pixel regions of the generated image is improved, and the quality of the generated image is improved.

Description

technical field [0001] The invention belongs to the field of computer vision, and relates to a generation confrontation network method integrating a self-attention mechanism. Background technique [0002] In recent years, deep learning technology with neural network as the core has been in full swing in the field of computer vision. The discriminative model in neural network has been applied to solve basic problems such as image recognition, image classification and image text description; however, for The generative model that generates image data faces problems such as high difficulty in the modeling process and poor definition of the generated effect. These reasons make it difficult to apply the generative model to the field of image generation. In order to solve this problem, the generative model and the discriminative model are combined to conduct confrontational training on the image data, which solves the difficult problem of the modeling process and improves the qual...

AI Technical Summary

Problems solved by technology

[0004] The earlier generation confrontation network technology has the problem of mode collapse, and the sample mode is single and lacks diversity.

In detail, the generation confrontation network draws the data distribution of the image generated by the generator to the data distribution of the real image through confrontational training, but the data distribution of the real image is difficult to obtain, and the discriminator is trained by using real image data to obtain a close to the real Image data distribution, the original generative confrontation network technology describes the distance between two image data distributions as cross entropy or JS divergence. When the two distributions do not intersect at the beginning of training, the training results using JS divergence are poor , cannot provide effective gradients for the training of the generator and lead to mode collapse

In addition, it was difficult for the previously uncontrollable generative adversarial network to generate pixel-clear images. The reason is that its network architecture only includes convolutional layers and cannot obtain the relationship between local pixel regions in the entire image. Therefore, the generated Image with full image outline but poor clarity and missing details

Images