| ABSTRACT: | ULTRASONIC NDE IMAGE ENHANCEMENT WITH NONLINEAR FILTERS IN
SIGNAL SUBSPACES
C.H. Chen, and Xianju Wang
University of Massachusetts Dartmouth, Electrical and Computer Engineering Dept.
N. Dartmouth, MA, USA
Abstract: The objective of this study to examine an effective method to remove speckles for
enhancing ultrasonic NDE images. The new method presented is based on Independent
Component Analysis (ICA). Firstly, in terms of the characteristic of NDE images, we believe that
most NDE images are constructed by "edges" and "textures", thus we build "Edge Model" and
"Texture Model" to describe their basis images and propose a new ICA learning algorithm to
estimate the model parameters. Then a demixing transform of the original image is employed, and
we design nonlinear filters in each signal subspace and obtain our restored image after a mixing
transform. Finally, we compare our method with median filtering and Wiener filtering. The
experimental results show that the proposed method can remove the speckle noise effectively, as
measured by signal-to-noise ratio, and efficiently with the use of fast algorithms, while at the
same time retaining important details without introducing artificial structures. Based on our
study, the ICA based ultrasonic NDE image enhancement is better than existing procedures. In
the Appendix detailed mathematical analysis and comparisons are made among the several signal
subspace models including the Ridgelet model, Gabor model, Fourier model, and Wavelet model
|