基于十字形窗口的生成对抗网络模型

Generative Adversarial Network Model Based on Cross-Shaped Window

由于传统的生成对抗网络(generative adversarial network,GAN)都是以卷积神经网络(convolutional neural networks,CNN)作为基本框架,CNN无法处理远程依赖关系,因此会导致图片特征分辨率低和精细细节损失的问题。CSWin Transformer中的十字形窗口自注意力机制可以有效捕获图像组件之间的远程依赖关系,本文提出一种基于CSWin Transformer的生成对抗网络模型CTGAN(CSWin Transformer GAN),模型在CIFAR-10数据集和更高分辨率的CelebA数据集上进行测试,模型表现出了较好的生成效果,可以生成保真度高且细节丰富的图片。

Since traditional generative adversarial networks(GAN)are based on convolutional neural networks(CNN)as the basic framework,CNN cannot process remote dependency relationships.As a result,image feature resolution and fine detail loss will be caused.The cross-shaped window attention mechanism in CSWin Transformer can effectively capture remote dependencies between image components.Therefore,in this article we propose a generative adjoint network model CTGAN(CSWin Transformer GAN)based on CSWin Transformer.The model was tested on the CIFAR-10 datasets and the CelebA datasets with higher resolution,and it showed a good generation effect.Moreover,high fidelity and detailed images can be generated.