IST-OAM:Image Style Transfer via Optimized Self-Attention Mechanism
Ning Jia, Mengshi Li, Nan Yin · 2024
In recent years, style transfer has found increasingly broad applications in daily life, particularly in the domain of photography image style transfer. However, current methods face limitations in handling local and global information in photography style images due to their inherent structural constraints. This often leads to issues such as content loss and erroneous transformations between unrelated content. To address these critical issues, this paper proposes the IST-OAM style transfer model based on the self-attention mechanism of the Transformer model. By integrating image segmentation channels with transfer branches and defining enhanced style loss, the model strengthens structural constraints on images during the transfer process, ensuring that the image structure is preserved and not lost or disrupted. This approach aims to achieve more accurate results in style transfer. The proposed method is evaluated using various scenes from photography images. Experimental results demonstrate that the IST-OAM model outperforms other methods both visually and in quantitative metrics.