Robust 3D Gaussian Splatting Using a Hybrid Image Restoration Network

Recent advancements in 3D Gaussian Splatting (3DGS) have shown great promise for novel view synthesis in applications such as 3D scene reconstruction, digital human modeling, and virtual reality. 3DGS represents objects and scenes through explicit radiance fields, significantly accelerating training and rendering speeds while maintaining high-quality modeling. However, most existing 3DGS methods rely on clean images as input, overlooking the effects of image degradation. To address this issue, we propose a hybrid image restoration network for 3DGS that mitigates rendering quality degradation caused by uneven exposure, motion blur, and noise. Our approach employs a multi-stage image restoration network that integrates exposure correction, deblurring, and denoising, substantially enhancing the input image quality. Experimental results demonstrate that our method yields more robust modeling and superior rendering quality on both public datasets and photos captured by mobile devices.