Multi-Scale Convolutional Attention Enhanced Arbitrary Image Style Transfer

This research addresses the balance of style expression and retention of content in arbitrary image style transfer by proposing an improved CAST framework integrated with a Multi-Scale Convolutional Attention module. Existing style transfer methods often suffer from detail loss or structural distortion when processing complex style features. To solve these issues, we designed a parallel multi-scale convolutional structure that simultaneously captures local and global style features through different receptive fields, enhancing the system's perception of complex textures. We also introduced a Structural Similarity loss function to improve content structure preservation. Compared to existing methods, our framework achieves significant improvements in both style expressiveness and content preservation, generating more natural and artistic stylized images.