Dynamic Spatial Aggregation Network for Joint Denoising and HDR Imaging

Imaging systems often produce underexposed or overexposed images due to their limited capacity to capture the full range of natural illumination. To address this limitation, deep high dynamic range (HDR) imaging methods rely on fusing multiple low dynamic range (LDR) images, demonstrating remarkable performance in recent years. However, most existing methods solely focus on object motions, disregarding real-world noise in images, which typically leads to suboptimal solutions in actual situations. In this paper, we propose a dynamic spatial aggregation network, which can simultaneously handle noise and large motion of objects in LDR images in real-world scenarios. Our method employs a dynamic convolutional operator, which is image-specific and ensures robustness to noise and misaligned image content, to adaptively extract features from input LDR images. Experimental results demonstrate that our proposed method significantly outperforms state-of-the-art deep HDR models, resulting in high-quality HDR images in real-world situations.