Arbitrary style transfer system with split-and-transform scheme

doi:https://doi.org/10.1007/s11042-023-16582-5

機構典藏 > College of Engineering > Graduate Institute & Department of Computer Science and Information Engineering > Journal Article > Item 987654321/126784

Please use this identifier to cite or link to this item: https://tkuir.lib.tku.edu.tw/dspace/handle/987654321/126784

Title:	Arbitrary style transfer system with split-and-transform scheme
Authors:	Tu, Ching-Ting;Lin, Hwei Jen;Tsai, Yihjia;Lin, Zi-Jun
Date:	2024-01-12
Issue Date:	2025-03-20 09:24:47 (UTC+8)
Abstract:	For the subject of arbitrary image style transfer, there have been some proposed architectures that directly compute the transformation matrix of the whitening and coloring transformation (WCT) to obtain more satisfactory transformation results. However, calculating the transformation matrix of WCT is time-consuming. Li et al. trained a linear transformation module to generate a WCT transformation matrix for any pair of images, i.e., content image and style image, to avoid complex calculations and improves time efficiency. In this work, we introduce a flexible arbitrary image style transfer framework based on the LST, which uses deep neural networks to train a linear transformation matrix as the standard matrix for WCT. For the first part, inverse relationship between the Whitening matrix and the Coloring matrix w.r.t. the same image is enforced during the training of the linear transformation matrix, so that the resulting matrix will be more accurate and closer to the standard matrix of WCT. For the second part, a split-and-transform scheme is proposed. Unlike LST, which transforms the block of feature maps as a whole, the split-and-transform scheme divides the feature block into several smaller blocks and transforms them individually, so that the transformation is more localized, and the more the number of divided blocks, the more localized. In addition, the proposed split-and-transform scheme allows users to determine the number of divided blocks to flexibly control the locality of the transformations. Experimental results demonstrate the effectiveness and flexibility of the proposed framework by the high-quality stylized images and adjustable balance between globality and locality of transformations. The use of the split-and-transform scheme can reduce the computational time while preserving or even improving the stylization results.
Relation:	Multimedia Tools and Applications 83，p. 62497-62517
DOI:	https://doi.org/10.1007/s11042-023-16582-5
Appears in Collections:	[Graduate Institute & Department of Computer Science and Information Engineering] Journal Article

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	27	View/Open

Loading...