A DNeRF Image Denoising Method Based on MSAF-DT

Abstract

Rendering novel and realistic images is crucial in applications such as augmented reality, virtual reality, 3D content creation, gaming, and the film industry. However, dynamic image rendering often suffers from significant noise, which compromises clarity and realism. Dynamic-Neural Radiance Fields (D-NeRF), an extension of the original NeRF model, addresses this challenge by enabling the rendering of dynamic images. Despite its advantages, D-NeRF often generates significant noise in the rendered images. Addressing this limitation, this paper proposes a Transformer-based model, Multi-Scale Attention Fusion Denoise Transformer (MSAF-DT), designed to enhance the clarity of rendered images. MSAF-DT constructs a deep neural network by stacking multiple Transformer blocks, with each block adaptively extracting complex features and dependencies from the data. The multi-head self-attention (MHSA) mechanism effectively captures long-range dependencies, which is crucial for processing sequences in dynamic radiance fields. Additionally, the model supports parallel processing of the entire sequence, significantly enhancing training efficiency. This design enables MSAF-DT to handle the noise present in D-NeRF outputs while preserving essential features. Experimental results on the Nerf_Synthetic dataset demonstrate that the proposed method outperforms D-NeRF in both image clarity and processing efficiency, achieving higher PSNR scores and faster convergence during training.