图像复原
计算机科学
计算机视觉
人工智能
图像(数学)
图像处理
图像分割
迭代重建
图像压缩
可视化
特征提取
数据压缩
计算机图形学(图像)
模式识别(心理学)
作者
Yuan Meng,Yubin Gu,Wenting Luo,Jiayi Ji,Xiaoshuai Sun,Weijian Ruan,Rongrong Ji
标识
DOI:10.1109/tcsvt.2026.3728473
摘要
While “All-in-One” (AIO) image restoration (IR) models demonstrate remarkable versatility by unifying multiple degradations via prompts or degradation-aware modules, their generalization to specific tasks remains limited. A fundamental bottleneck lies in their neglect of distinct frequency-domain energy distributions associated with diverse degradations, which severely restricts feature discriminability and the explicit modeling of cross-frequency interactions. We revisit unified image restoration from a frequency perspective and observe that existing dynamic adaptation mechanisms frequently suffer fromfrequency response collapse, wherein different attention heads or dynamic kernels degenerate to exhibit nearly identical spectral behaviors. This redundancy weakens spectral diversity and restricts the model’s capacity to adapt to varied degradations. To tackle this issue, we propose UFSIR, a unified image restoration framework driven by structured frequency modeling. Specifically, we introduce the Spectral Sub-band Decoupling Learning (SSDL) module, which partitions the frequency spectrum into radial sub-bands and learns sub-band-specific dynamic filters, enforcing diverse and adaptive spectral responses. Building on these representations, the Spectral Graph Interaction (SGI) module explicitly models complementary cross-band dependencies through a lightweight graph structure, enabling coherent and explicitly quantifiable spectral fusion. Extensive experiments demonstrate that UFSIR achieves promising results across allin-one, composite-in-one, and single-specific scenarios with low computational cost and fast inference, eliminating the need for explicit degradation prompts. The source code will be made publicly available.
科研通智能强力驱动
Strongly Powered by AbleSci AI