Selective kernel convolution manifold alignment network for arbitrary style transfer

doi:10.1016/j.compeleceng.2025.110811

Selective kernel convolution manifold alignment network for arbitrary style transfer

2025-11-14

0

PRE

AI

DOI：10.1016/j.compeleceng.2025.110811

Original

Original request for help

Save

Abstract

En 中文

Although existing style transfer methods have achieved good stylization results, most methods neglect the preservation of content details in stylized images. To alleviate this problem, we introduce a selective kernel convolution manifold alignment network for arbitrary style transfer (SCMANet). Specifically, we propose the Shallow Style Transfer (SST) module and the Selective Kernel Convolution Manifold Alignment (SCMA) module in our SCMANet. In the quantitative comparison, our SCMANet achieves scores of 0.40, 31.30, 0.68, and 299.54 for the evaluation metrics Structural Similarity Index (SSIM), Art Fréchet Inception Distance (ArtFID), Learning Perceived Image Patch Similarity (LPIPS), and Fréchet Inception Distance (FID) respectively. Extensive experiments demonstrate the effectiveness of SCMANet in preserving content details and achieving good stylization performance.

Selective kernel convolution manifold alignment network for arbitrary style transfer

Abstract

Journal

Organization