RAW domain HDR low-light image reconstruction via joint optimization of neural exposure field and low-rank denoising

Abstract

We improve the RAW domain high dynamic range (HDR) low-light image reconstruction algorithm based on the joint optimization of neural exposure field and low-rank denoising. We attempt to construct a sensor-specific RAW domain degradation model that improves the characterization of noise, underexposure, and dynamic range compression in real low-light scenes, and avoids information loss caused by ISP processing. Using the neural exposure field module, we build a model with variable exposure features in image space to achieve adaptive exposure adjustment and HDR expansion. We introduce a low-rank denoising module integrating non-local similarities, which uses the low-rank structure and redundant information of image blocks to preserve edge texture details and suppress Gaussian mixture noise. We jointly optimize the loss function to realize end-to-end collaborative optimization of exposure adjustment and denoising. Experiments on public datasets (SID, RAW-LOL) show that our algorithm achieves a PSNR of 32.67 dB and an SSIM of 0.945, with optimized quantitative metrics and visual effects. It effectively alleviates overexposure and underexposure in bright and dark areas, suitable for high-quality low-light imaging in professional photography and medical imaging.