Bayesian classification of ultrasound signals using wavelet coefficients

Abstract

Ultrasound is a common tool in the nondestructive evaluation of composite material integrity. Echoes from high frequency (5-10 MHz) sound waves vary with subsurface flaws and delaminations. To improve detection of internal defects in composite materials, a linear Bayes classification is applied to the wavelet transform coefficients of ultrasound scan lines. Using a Daubechies basis, wavelet transforms are taken of the ultrasound signals. A subset of the coefficients is then used as features for the classifier. A forward sequential feature selection (FSFS) algorithm was used to determine the optimal features. The training set was comprised of scanned signals from both damaged and undamaged samples. Performance statistics for classification of damaged materials were calculated using a separate set of test samples. Application of a shift-invariant wavelet transform removed some of the variability of the wavelet coefficients and improved the classification.