Scholarlink

<p>Scanned point clouds offer a flexible and effective 3D representation for aerospace thin-walled parts, which are increasingly applied in downstream tasks such as size measurement, defect detection, and object tracking. However, due to the small production batches and varying working conditions of these parts, only a limited number of scanned point clouds are available, often containing diverse outliers and complex distributed noise, which makes it challenging to obtain clean data required for downstream tasks. The geometry-aware multi-task learning (GAMTL) network is therefore proposed to remove outliers and denoise scanned point clouds. This architecture consists of a shared geometry-aware encoder based on the dynamic graph convolutional network and two branch networks that perform outlier removal and denoising tasks. The outlier removal branch network incorporates a multi-scale neighborhood attention mechanism to enhance feature learning. The denoising branch network introduces a score estimation network to learn the noisy distribution, followed by an adaptive step-size algorithm that dynamically denoises the input points based on the estimated score. Additionally, an uncertainty-weighted loss function is employed for joint training. Experiments confirm that the GAMTL network demonstrates superior performance in both quantitative metrics and visual quality.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Geometry-aware multi-task learning network for outlier removal and denoising in scanned point clouds of aerospace thin-walled parts

  • 期刊论文

  • 发表日期：

    2025年6月17日

查看全文
  • Liling Zuo,
  • Jie Zhang,
  • Hui Cheng,
  • Youlong Lyu

摘要

Scanned point clouds offer a flexible and effective 3D representation for aerospace thin-walled parts, which are increasingly applied in downstream tasks such as size measurement, defect detection, and object tracking. However, due to the small production batches and varying working conditions of these parts, only a limited number of scanned point clouds are available, often containing diverse outliers and complex distributed noise, which makes it challenging to obtain clean data required for downstream tasks. The geometry-aware multi-task learning (GAMTL) network is therefore proposed to remove outliers and denoise scanned point clouds. This architecture consists of a shared geometry-aware encoder based on the dynamic graph convolutional network and two branch networks that perform outlier removal and denoising tasks. The outlier removal branch network incorporates a multi-scale neighborhood attention mechanism to enhance feature learning. The denoising branch network introduces a score estimation network to learn the noisy distribution, followed by an adaptive step-size algorithm that dynamically denoises the input points based on the estimated score. Additionally, an uncertainty-weighted loss function is employed for joint training. Experiments confirm that the GAMTL network demonstrates superior performance in both quantitative metrics and visual quality.