Feature Extraction and Selection for Pattern Recognition of Two-Dimensional Hydrometeor Images

View More View Less
  • 1 University of Wyoming, Laramie, WY 82071
© Get Permissions Rent on DeepDyve
Restricted access

Abstract

New feature extraction techniques are developed for two-dimensional binary images of ice particles and raindrops. These features are employed in the statistical classification of these patterns into one of seven basic hydrometeor shapes. These images have been recorded by an airborne two-dimensional probe in order to provide information leading to an understanding of important physical processes in clouds. Minimum average probability of classification error is employed as the performance criterion with informal minimization of computing time.

A synthetic image set was generated to develop feature extraction techniques. Moment normalization and rotation normalization are employed to convert all images to a common size and orientation.

Ten time domain features are explored including a new development (circular deficiency) and a new application (cross correlation). Three frequency domain features are investigated including a new Fourier descriptor (centroid distance). An original method of reducing aliasing due to binary quantization is tested. Final features are selected on the basis of classification performance when applied to the synthetic image set. Bayes classification with four selected time domain features (independent of size, position and angular orientation) was twenty times faster than frequency domain classification. The average recognition time was 2032 images per minute for time domain classification.

Abstract

New feature extraction techniques are developed for two-dimensional binary images of ice particles and raindrops. These features are employed in the statistical classification of these patterns into one of seven basic hydrometeor shapes. These images have been recorded by an airborne two-dimensional probe in order to provide information leading to an understanding of important physical processes in clouds. Minimum average probability of classification error is employed as the performance criterion with informal minimization of computing time.

A synthetic image set was generated to develop feature extraction techniques. Moment normalization and rotation normalization are employed to convert all images to a common size and orientation.

Ten time domain features are explored including a new development (circular deficiency) and a new application (cross correlation). Three frequency domain features are investigated including a new Fourier descriptor (centroid distance). An original method of reducing aliasing due to binary quantization is tested. Final features are selected on the basis of classification performance when applied to the synthetic image set. Bayes classification with four selected time domain features (independent of size, position and angular orientation) was twenty times faster than frequency domain classification. The average recognition time was 2032 images per minute for time domain classification.

Save