High dynamic range (HDR) imaging is still a significant yet challenging problem due to the limited dynamic range of generic image sensors. Most existing learning-based HDR reconstruction methods take a set of bracketed-exposure sRGB images to extend the dynamic range, and thus are computational- and memory-inefficient by requiring the Image Signal Processor (ISP) to produce multiple sRGB images from the raw ones. In this paper, we propose to broaden the dynamic range from the raw inputs and perform only one ISP processing for the reconstructed HDR raw image. Our key insights are threefold: (1) we design a new computational raw HDR data formation pipeline and construct the first real-world raw HDR dataset, RealRaw-HDR; (2) we develop a lightweight-efficient HDR model, RepUNet, using the structural re-parameterization technique; (3) we propose a plug-and-play motion alignment loss to mitigate motion misalignment between short- and long-exposure images. Extensive experiments demonstrate that our approach achieves state-of-the-art performance in both visual quality and quantitative metrics.

翻译：高动态范围成像仍是因通用图像传感器动态范围有限而极具挑战性的重要问题。现有基于学习的高动态范围重建方法大多利用一组包围曝光sRGB图像来扩展动态范围，因此需要图像信号处理器从原始图像生成多张sRGB图像，导致计算和内存效率低下。本文提出从原始输入直接扩展动态范围，仅需对重建的高动态范围原始图像进行一次ISP处理。我们的核心创新体现在三个方面：（1）设计了一套新的计算式原始高动态范围数据生成流程，并构建了首个真实世界原始高动态范围数据集RealRaw-HDR；（2）采用结构重参数化技术开发了轻量高效的高动态范围模型RepUNet；（3）提出即插即用的运动对齐损失函数以缓解长短曝光图像间的运动错位。大量实验表明，本方法在视觉质量和量化指标上均达到当前最优水平。