3D reconstruction in industrial environments based on an improved neural radiation field method research

Abstract

Neural Radiation Field (NeRF) is driving the development of 3D reconstruction technology. Several NeRF variants have been proposed to improve rendering accuracy and reconstruction speed. One of the most significant variants, TensoRF, uses a 4D tensor to model the radiation field, resulting in improved accuracy and speed. However, reconstruction quality remains limited. This study presents an improved TensoRF that addresses the aforementioned issues by reconstructing its multilayer perceptron network. Increasing the number of neurons in the input and network layers improves the render accuracy. To accelerate the reconstruction speed, we utilized the Nadam optimization algorithm and the RELU6 activation function. Our experiments on various classical datasets demonstrate that the PSNR value of the improved TensoRF is higher than that of the original TensoRF. Additionally, the improved TensoRF has a faster reconstruction speed (≤30min). Finally, we applied the improved TensoRF to a self-made industrial dataset. The results showed better global accuracy and local texture in the reconstructed image.