next up previous
Next: Metal detector image preprocessing Up: Data preprocessing - noise Previous: GPR data preprocessing

IR data preprocessing

An efficient denoising method in the wavelet domain has been proposed by the RUG [15]. This method adds spatial constraints to the criterion for selecting noisy wavelet coefficients and for each coefficient the probability of being noise-free is computed. The spatial constraints are derived from prior geometrical assumptions expressing the fact that meaningful wavelet coefficients appear in spatially connected clusters, at the location of characteristic image features like edges, corners, etc. For the criterion itself the magnitudes of the wavelet coefficients are used.
  
Figure 16: original IR images, with histogram equalization for the 2th and the 3th images (top images) - corresponding restored IR image (bottom image)
\includegraphics[width=4cm]{psfiles/ir4_orig_histequ.ps} \includegraphics[width=4cm]{psfiles/ir2_orig_bw.ps} \includegraphics[width=4cm]{psfiles/ir7_orig_histequ.ps}
\includegraphics[width=4cm]{psfiles/ir4_rest_bw.ps} \includegraphics[width=4cm]{psfiles/ir2_rest_bw.ps} \includegraphics[width=4cm]{psfiles/ir7_rest_bw.ps}

Fig.(16) summarizes the results. The top images are the IR original images, the two last ones being processed using histogram equalization techniques. The bottom images are their respective restored versions.

The analysis of an IR sequence by mean of the Karhunen-Loève transform or the Kittler-Young transform, which as a matter of fact requires a learning phase, leads to interesting results [8] as well. Fig. 17 shows the obtained results. On the left side, sample IR images of the sequence are presented. On the right side, the two most significant images (with the largest variances) after transformation are presented for the two transforms.

  
Figure 17: Karhunen-Loève and Kittler-Young transforms
\includegraphics[width=14cm]{psfiles/Meerdaal_ir_sequence.ps}


next up previous
Next: Metal detector image preprocessing Up: Data preprocessing - noise Previous: GPR data preprocessing
Marc Acheroy
2000-08-03