• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文期刊>Journal of visualization >Deep learning-assisted segmentation of bubble image shadowgraph
【24h】

Deep learning-assisted segmentation of bubble image shadowgraph

机译：Deep learning-assisted segmentation of bubble image shadowgraph

获取原文
获取原文并翻译 | 示例
           
页面导航
  • 摘要
  • 著录项
  • 相关主题

摘要

Abstract Global thresholding parameters for the semantic segmentation of bubbles from experimental bubble image shadowgraph were implemented. Traditional image processing algorithms for experimental visualization of multiphase flows require very rigorous and time-consuming trial by error of applying thresholding to be able to obtain the bubble statistics. More so, due to the varying flow conditions and lighting system during experimentation, it is impossible to apply a global threshold for in the post-processing the results of visualized flows. BIMSNet (modified U-Net architecture) was trained with bubble shadowgraph images obtained from experiments with varying flows and lightning conditions and developed global threshold parameters (binarization threshold) to semantically segment clustered bubbles with irregular shapes. The variation of pixel intensity of the sequence of images was taken into consideration in training the network. The average dice coefficient score (accuracy) of the network on the validation dataset was 99.3% with a 1.2% loss. Evaluation of the trained network on the test dataset gave an average precision and dice coefficient score of 99.73%, respectively. The detection of bubbles with the trained model when compared with the local average adaptive threshold image extraction process yields a higher bubble detection rate with less amount of misdetection and eliminates the trial-by-error method of obtaining the threshold limits for the binarization of images when post-processing images. Graphical abstract

著录项

获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号