Feature Consistency-Based Style Transfer for Landscape Images Using Dual-Channel Attention

Abstract

With the rapid development of artificial intelligence technology, style transfer has become an important topic in current research. However, existing models are deficient in fusing content and style features, resulting in a large gap between the generated image and the target image. To solve this problem, we propose a feature-consistent landscape image style transfer model based on two-channel attention. Unlike traditional models that rely on VGG as a content encoder, VGG has a more limited detail extraction capability when dealing with high-resolution landscape images, so we introduce a Visual Transformer (VIT) to enhance the extraction of content features. In addition, by incorporating a channel attention mechanism in the latent space, we achieve consistency between content and style features, which in turn completes the alignment and fusion of feature distributions. Finally, contrast constraints are applied to accelerate the style transfer process. Comparison experiments show that our method outperforms other existing methods on the style transfer task.

Keywords

• Convolutional neural network
• cross-attention mechanism
• contrastive constraints
• feature extraction
• style transfer