ISPFormer: Learning RAW-to-sRGB mappings with wavelet-based self-attention

doi:10.1016/j.neucom.2025.131084

ISPFormer: Learning RAW-to-sRGB mappings with wavelet-based self-attention

2025-07-30

0

PRE

AI

DOI：10.1016/j.neucom.2025.131084

Original

Original request for help

Save

Abstract

En 中文

RAW-to-sRGB mapping, or the simulation of the traditional camera image signal processor (ISP), aims to generate DSLR-quality sRGB images from RAW data captured by smartphone sensors. Despite achieving comparable results to sophisticated handcrafted camera ISP solutions, existing CNN-based methods still suffer from detail disparity and color distortion due to their inherent locality restrictions. In this paper, we present ISPFormer, a novel Transformer-based framework utilizing self-attention to tackle the learnable ISP problem. Specifically, we propose a Wavelet-based Transformer Block (WTB) with two loss functions to correct color and enhance high-frequency details, where WTB integrates wavelet transformation into window-based self-attention to perform self-attention in sub-bands across frequency domains, enabling larger sliding window modeling without additional computational overhead. Based on the proposed WTB, we construct the ISPFormer as a multi-stage network, where a multi-scale window adjustment strategy is further proposed to flexibly assign varying window sizes for each stage, reconstructing visually satisfactory results in a coarse-to-fine manner. Extensive experiments demonstrate that our ISPFormer achieves competitive quantitative and qualitative results. Code is available at https://github.com/RenYangSCU/ISPFormer .

Keywords:

RAW-to-sRGB mapping

Transformer-based framework

Wavelet-based Transformer Block

color correction

detail enhancement

ISPFormer: Learning RAW-to-sRGB mappings with wavelet-based self-attention

Abstract

Journal

Organization