Low-Resolution Image Rendering Method Based on Reflected Radiance and Intrinsic Decomposition

Neural Radiance Fields (NeRF) utilizes neural networks to generate high-quality novel view images, but their rendering speed is slow. Reducing the resolution of input images can accelerate the rendering process, but it affects the quality of the synthesized images. Specifically, using low-resolution images as input means that each pixel contains less scene information, limiting the model's ability to reconstruct fine details in the scene. In this paper, we replace the original outgoing radiance parameterization with a view-based reflected radiance parameterization method to enhance the low-resolution novel view synthesis. Meanwhile, we decompose the surface appearance into two parts, view-independent diffuse color and view-dependent specular color, where the diffuse color is further decomposed into illumination-invariant reflectance and illumination-varying shading by intrinsic decomposition, and the specular color is calculated by combining the roughness and normal information. It is experimentally demonstrated that our model can accurately render details as well as edge regions of the scene in low-resolution scenes to achieve high-quality rendering results. In addition, our method accurately renders patterns that overlap with reflected light, enabling a more complete scene reconstruction.