关键词: 图像处理, 图像风格迁移, 生成对抗网络, CycleGAN, Ghost卷积, 反残差模块, 特征提取, 颜色重建损失

Abstract:

Image style transfer has been a research hotspot in the field of image processing, but the current style transfer models have problems such as fuzzy details of generated image, the poor color effect of style texture and excessive model parameters. An image style transfer method based on improved cycle-consistent generative adversarial network is proposed in this study. The generator network architecture is improved by adding the Ghost convolution module and the inverted residual improved module to reduce the number of model parameters and computation cost, as well as enhance the feature extraction capability of the network. And the content style loss item, the color reconstruction loss item and the map identity loss item are added to the loss function for enhancing generative capability of the model and improving the quality of generated images. The experimental results show that the proposed method has a stronger ability for style transfer, which enhances the content details and color effect of style texture from generated images effectively, improves the image quality significantly, and the model performance has also been improved well.

Key words: image processing, image style transfer, generative adversarial network, CycleGAN, Ghost convolution, inverted residual module, feature extraction, color reconstruction loss