Research Article
Image Cross-Domain Generation and Adaptation: A Frequency-Domain Decoupling Perspective
@INPROCEEDINGS{10.4108/eai.22-5-2026.2365354, author={Qiwang Gao}, title={ Image Cross-Domain Generation and Adaptation: A Frequency-Domain Decoupling Perspective}, proceedings={Proceedings of the 4th International Conference on Image, Algorithms, and Artificial Intelligence, ICIAAI 2026, 22-24 May 2026, Singapore, Singapore}, publisher={EAI}, proceedings_a={ICIAAI}, year={2026}, month={8}, keywords={Frequency Domain Analysis; Image Synthesis; Domain Adaptation; Feature Decoupling; Fourier Transform}, doi={10.4108/eai.22-5-2026.2365354} }- Qiwang Gao
Year: 2026
Image Cross-Domain Generation and Adaptation: A Frequency-Domain Decoupling Perspective
ICIAAI
EAI
DOI: 10.4108/eai.22-5-2026.2365354
Abstract
This paper takes a systematic approach to summarize the literature on the frequency-domain view of image generation and cross-domain transfer. Mechanisms analysis, Fourier and wavelet transforms are some of the underlying mechanisms that are analyzed to aid the construction of a systematic classification framework. The framework has a set of 5 basic operation mechanisms and three transfer levels: input space, feature space, and model/latent space. The frequency domain approaches are successful in addressing the bottlenecks of the spatial domain approaches in managing the long-range dependencies, attenuating the artifacts, and aligning cross-domain distributions. This effectiveness is attributable to inherent statistical decoupling characteristics as well as global modeling capabilities. The history of technological development makes evident a consistent progression of the external data alignment technology to feature-level calibration, and finally to latent-space intervention and reconstruction at the architecture level operator.
