A NeRF Rendering Method for Low-Light Scenes Using Occlusion Perception and Network Decomposition

Neural Radiance Fields (NeRF) is a technique where spatial coordinates or other dimensions (such as time, camera poses, etc.) are utilized as input to simulate the target function through Multilayer Perceptron (MLP) networks. This process generates scalar values (such as color, depth, etc.). However, in scenes with uneven lighting and shadows, NeRF tends to produce overly blurry or artifact-laden rendering effects. To address this issue, a NeRF low-light scene rendering method based on occlusion perception and network decomposition is proposed. Firstly, camera poses and geometric information are extracted using the Structure from Motion (SFM) and Multiple View Stereo (MVS) pipeline methods. Secondly, a combination of local occlusion fields and multilayer perceptron networks is employed to reduce light transmission during the volume rendering stage. Convolutional Neural Networks (CNN) are used to enhance spatial relationships between pixels. The convolutional layer adjusts each pixel based on the lighting conditions of surrounding pixels to generate occlusion fields with more lighting information, contributing to smoother rendering results. Finally, a global occlusion field is introduced during the volume rendering process to globally control light occlusion effects on light propagation throughout the entire scene. To expedite training time, network decomposition is utilized. This involves splitting the original NeRF network, which accepts 5-dimensional data, into two networks. This approach effectively reduces training time. Experimental results demonstrate that the proposed method achieves at least a 16.65% improvement over NeRF, SCI, and other methods in terms of rendering quality on the LOM and RELLISUR datasets, rendering richer dark scenes with superior visual effects.