1.College of Mathematics and Computer Science, Zhejiang A & F University;2.Key Laboratory of Forestry Perception Technology and Intelligent Equipment of the State Forestry and Grassland Administration;3.Lin'an Agricultural Information Center Practice Base
Abstract: 3D reconstruction plays a critical role in various fields, including computer vision and artificial intelligence, medical imaging, architecture, and urban planning. To address the inefficiency of traditional manual modeling methods, this paper proposes a method based on Neural Radiance Fields(NeRF) that incorporates multi-scale fusion and attention mechanisms. The approach introduces a multi-scale feature module combined with graph convolutional networks to enhance the network's understanding of spatial structures, allowing for more accurate capture of both local and global geometric relationships. The multi-scale feature module extracts information at different levels, improving the accuracy and comprehensiveness of detail reconstruction, which in turn enhances overall reconstruction quality.Additionally, to further improve the model's robustness and precision, a feature pyramid network is introduced to ensure the network can effectively capture important information across different scales, particularly in complex scenes where details might otherwise be lost. The integration of the Squeeze-and-Excitation attention mechanism allows the model to adaptively focus on key regions in the image, enhancing the representation of important features and improving reconstruction performance in challenging environments.Experimental results demonstrate that the proposed method outperforms the NeRF model on a self-built building dataset, achieving SSIM, PSNR and LPIPS of 0.784, 25.42 and 0.183, respectively. These metrics show improvements of 4.39%, 3.29% and 15.84% over the NeRF model, indicating better handling of complex reconstruction tasks. This method provides a new approach for 3D reconstruction in various application domains.
Key words : 3D reconstruction;graph convolutional network;feature pyramid network;NeRF;SE attention