Scholarlink

<p>Aiming at the problem that the images acquired by industrial meters in complex environments have an angular skew and large errors in the reading results, an improved Reliable and Repeatable Detector and Descriptor (R2D2) image auto-correction method with two-channel soft thresholding is proposed by introducing a soft threshold function in the Convolutional Block Attention Module (CBAM) residual neural network and replacing the original backbone network of the R2D2 model with the improved network. Additionally, pointer-type instrument images cover multiple instrument types and require data collection under complex operating conditions. Model training and real-time registration necessitate processing large volumes of noisy instrument feature data. This demands leveraging the parallel processing capabilities of high-performance computing (HPC) to ensure efficiency, thereby preventing excessive computational time per device from impacting industrial applications. The experimental results show that the improved model can increase the number of feature points and the matching accuracy of image alignment, thus improving the accuracy and robustness of image alignment. The improved algorithm has good image alignment performance compared with the traditional R2D2 algorithm, especially in illuminated scenes.</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

Image registration of pointer gauges based on improved reliable and repeatable detector and descriptor algorithm

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Hongli Liu,
  • Jingyi Qu,
  • Ji Li,
  • Chao Li,
  • Wentao Sun,
  • Tiancheng Zhang

摘要

Aiming at the problem that the images acquired by industrial meters in complex environments have an angular skew and large errors in the reading results, an improved Reliable and Repeatable Detector and Descriptor (R2D2) image auto-correction method with two-channel soft thresholding is proposed by introducing a soft threshold function in the Convolutional Block Attention Module (CBAM) residual neural network and replacing the original backbone network of the R2D2 model with the improved network. Additionally, pointer-type instrument images cover multiple instrument types and require data collection under complex operating conditions. Model training and real-time registration necessitate processing large volumes of noisy instrument feature data. This demands leveraging the parallel processing capabilities of high-performance computing (HPC) to ensure efficiency, thereby preventing excessive computational time per device from impacting industrial applications. The experimental results show that the improved model can increase the number of feature points and the matching accuracy of image alignment, thus improving the accuracy and robustness of image alignment. The improved algorithm has good image alignment performance compared with the traditional R2D2 algorithm, especially in illuminated scenes.