Gray-statistics-based Twin Feature Extraction for Hyperbola Classification in Ground Penetrating Radar images

Abstract

The row vector and column vector of a ground penetrating radar (GPR) B-scan image have different physical meanings, and the features of heterogeneous medium properties based on these vectors can provide new possibilities for hyperbola classification. This study uses the features of both row and column vectors (i.e., twin vectors), gray statistics, and united coding to produce a twin gray statistics sequence (TGSS), a representation of the GPR image, based on information entropy. An actual dataset and multiple classification methods are used to compare and evaluate the robustness and dimension reduction performance of TGSS. The results show that the proposed method has relatively favorable robustness and steady dimension reduction performance in the test environment with a small number of samples and class imbalance.