# PoseNeRF: In Situ 3D Reconstruction Method Based on Joint Optimization of Pose and Neural Radiation Field for Smooth and Weakly Textured Aeroengine Blade

## TL;DR

This paper introduces PoseNeRF, a new 3D reconstruction method for aeroengine blades that improves accuracy by combining pose estimation and neural radiance fields.

## Contribution

The novel contribution is a joint optimization approach combining pose and neural radiance field for high-fidelity 3D reconstruction of smooth, weakly textured blades.

## Key findings

- The ComBFNet network achieves 95.5% mean intersection over union for background filtering.
- PoseNeRF outperforms other models with PSNR of 25.59, SSIM of 0.719, and LPIPS of 0.239.
- The method reduces blurring and aliasing artifacts in 3D reconstructions of aeroengine blades.

## Abstract

Digital twins are essential for the real-time health management and monitoring of aeroengines, and the in situ three-dimensional (3D) reconstruction technology of key components of aeroengines is an important support for the construction of a digital twin model. In this paper, an in situ high-fidelity 3D reconstruction method, named PoseNeRF, for aeroengine blades based on the joint optimization of pose and neural radiance field (NeRF), is proposed. An aeroengine blades background filtering network based on complex network theory (ComBFNet) is designed to filter out the useless background information contained in the two-dimensional (2D) images and improve the fidelity of the 3D reconstruction of blades, and the mean intersection over union (mIoU) of the network reaches 95.5%. The joint optimization loss function, including photometric loss, depth loss, and point cloud loss is proposed. The method solves the problems of excessive blurring and aliasing artifacts, caused by factors such as smooth blade surface and weak texture information in 3D reconstruction, as well as the cumulative error problem caused by camera pose pre-estimation. The PSNR, SSIM, and LPIPS of the 3D reconstruction model proposed in this paper reach 25.59, 0.719, and 0.239, respectively, which are superior to other general models.