AFDU-NERF: Deblurring NeRF with Adaptiveaware Fusion and Iterative Data Updating

Neural Radiance Fields (NeRF) have garnered significant attention and application in 3D scene reconstruction and view synthesis. However, the high quality often rely on sharp multi-view images for NeRF training. In real-world scenarios, image data acquisition frequently depends on handheld devices such as smartphones, which can lead to image blurriness due to device instability, significantly degrading rendering quality. In this study, we propose deblurring NeRF with adaptive-aware fusion and iterative data updating (AFDU-NeRF). This method integrates a deblurring hybrid training module constructed with MIMO-UNet, which performs real-time deblurring during network training while dynamically updating the training dataset. The mechanism enables progressive optimization of 3D representations via hybrid training with updated data, achieving globally consistent deblurring. Additionally, our method employs dual parallel prediction networks to generate intermediate views. We use two prediction networks to parallel predict and generate intermediate views, dynamically merging the two intermediate views by learning fusion weights based on multi-view feature consistency. This approach alleviates the erroneous gradient propagation caused by the inclusion of blurred view images in the training process. Experiments under defocus blur, camera motion blur, and sharp input conditions demonstrate that AFDU-NeRF substantially enhances NeRF quality with blurred inputs while maintaining performance parity with sharp inputs.