| 基于卷积神经网络的镜头角度校准方法 |
| Lens calibration method based on convolutional neural network |
| 投稿时间：2025-04-14 修订日期：2025-07-18 |
| HTML 查看/发表评论 下载PDF阅读器 |
| 中文摘要: |
| 基于层析技术的三维燃烧诊断需要利用多个方向的火焰图像进行三维重建，而角度位置关系的准确性直接影响重建的质量和精度。针对角度标定问题，提出了一种基于神经网络的角度标定技术，包括一种新型十字型校准块以及训练集标定方法。所提出的十字型校准块具有更复杂的空间结构，能够减少训练集标定误差，同时为神经网络提供更高的分辨率，增强模型对角度特征的提取能力。测试结果表明，现有的三棱柱校准块方案由于角度标签标定误差较大，且各方向分辨率较低，导致神经网络的角度预测效果较差。而十字型校准块方案在一定的标定误差范围内，能够有效提升神经网络方法的校准精度，表现出更优的性能。 |
| 英文摘要: |
| Three-dimensional (3D) combustion diagnostics based on tomographic technology requires flame images from multiple viewing angles for 3D reconstruction, where the determination of angular positions directly impacts the quality and accuracy of the reconstruction. To address angle calibration challenges, this work proposes a machine learning-based angle calibration method. This method introduces a novel cross-shaped calibration block along with a training set design approach. The proposed cross-shaped calibration block features a more complex spatial struc-ture, which helps reduce calibration errors in the training set while providing higher resolution for the neural net-work, thereby enhancing its ability to extract angular features. Experimental results indicate that the existing triangu-lar prism calibration block has significant angle labeling errors and low directional resolution, leading to poor angle prediction by the neural network. In contrast, the cross-shaped calibration block effectively improves calibration accuracy within a certain error range, demonstrating superior performance in neural network-based calibration methods. |
|
| 中文关键词: 三维测量 角度标定 卷积神经网络 |
| 英文关键词:three-dimensional measurement angle calibration convolutional neural network |
| 基金项目:国家财政稳定支持项目（GJCZ-0202-04）光学图像处理相关的课题 |
| DOI： |
| 引用本文： |
| 关闭 |
