.png){kind=icon}
.png){kind=icon}
Return
Fft-guided multi-scale frequency enhancement for diffusion-based RAW-to-sRGB mapping
.png){kind=icon}
PRE
AI
Z
Y
W
DOI：10.1007/s00530-026-02582-6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The image signal processor (ISP) plays a critical role in converting raw sensor data into images with DSLR-like quality. Recent learning-based ISP methods have achieved significant progress; however, challenges remain in the RAW-to-sRGB mapping task, such as the insufficient recovery of high-frequency details and unstable color consistency. Diffusion models, despite their advantages in structural reconstruction, often introduce over-smoothing during iterative denoising, leading to texture degradation. This paper proposes the DFEM-ISP framework, which combines frequency-domain enhancement driven by the Fast Fourier Transform (FFT) with a conditional diffusion framework to improve grayscale detail reconstruction. In this framework, a multi-scale frequency enhancement module (MSFE) applies FFT analysis to features extracted from the diffusion backbone at multiple spatial scales, and explicitly compensates for diverse frequency components. Moreover, the frequency enhancement block (FEB) within MSFE is embedded into the decoder, combining adaptive frequency filtering and channel gating to progressively enhance edges and textures while suppressing noise. Additionally, a histogram-based color consistency module (HCCM) ensures global color stability in grayscale to sRGB mapping. Experimental results on the MAI dataset demonstrate that DFEM-ISP attains a 0.79 dB PSNR enhancement, coupled with markedly improved high-frequency details and color consistency, thereby confirming the efficacy of FFT-driven frequency enhancement for RAW-to-sRGB reconstruction.
Keywords:
Image signal processor
RAW-to-sRGB reconstruction
Diffusion model
Frequency enhancement
Fast fourier transform
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers