Fondazione Bruno Kessler - Technologies of Vision

contains material from
Template Matching Techniques in Computer Vision: Theory and Practice
Roberto Brunelli © 2009 John Wiley & Sons, Ltd

[prev] [prev-tail] [tail] [up]

4.3 L1 similarity measures

While the robustness achievable with tm.robustifiedCorrelation is remarkable, there are two related problems:

  1. it requires a substantial amount of computation;
  2. it can be, paradoxically, too robust: it looses discrimination efficiency.

Both problems can be solved using the similarity estimators based on the L1 norm as described in Section TM:4.3.

1 tm.noiseSimilarityImpact <- function(img, cRange=c(0,0.2)) { 
2 ...  c0  <- cRange[1] 
3 ...  c1  <- cRange[2] 
4 ...  cs  <- seq(c0,c1,(c1-c0)/10) 
5 ...  n   <- length(cs) 
6 ...  # 
7 ...  img <- ia.scale(img) 
8 ...  res <- array(0, dim=c(n, 5)) 
9 ...  # 
10 ...  i   <- 0 
11 ...  X   <- tm.normalizeImage(img) 
12 ...  for(c in seq(c0,c1,(c1-c0)/10)) { 
13 ...    i <- i + 1 
14 ...    res[i,1] <- c 
15 ...    # 
16 ...    Y <- tm.addNoise(img, noiseType="saltpepper", scale = 1, percent=c) 
17 ...    Y <- tm.normalizeImage(Y) 
18 ...    # 
19 ...    res[i,2] <- tm.robustifiedCorrelation(X, Y, "p") 
20 ...    res[i,3] <- tm.robustifiedCorrelation(X, Y, "O") 
21 ...    res[i,4] <- max(ia.correlation(X,Y, type="G")[[1]]@data)/100 
22 ...    res[i,5] <- max(ia.correlation(X,Y, type="L")[[1]]@data)/100 
23 ...  } 
24 ...  # 
25 ...  res 
26 ... }

PIC

Figure 4.6: The presence of noise adversely affects the correct assessment of pattern similarity. The plot compares the decay of the similarity computed by different estimators at varying degrees of contamination. The superior robustness of the tanh robustified correlation estimator has a downside: it is so robust that its discriminatory ability is reduced.

1 f   <- ia.get(img1, animask(32,87,104,104)) 
2 nsi <- tm.noiseSimilarityImpact(f) 
3 tm.dev("figures/robustCorrelation") 
4  matplot(nsi[,1], nsi[,2:5], type="b", pch=2:5, lty=2:5, 
5 ...          xlab="Salt&pepper contamination", ylab="Similarity") 
6  legend(0.1,0.9,c("Pearson", "Tanh", "G", "L"), lty=2:5, pch=2:5) 
7  grid() 
8 dev.off()

[prev] [prev-tail] [front] [up]