Just another site

Thought this was cool: image processing – Help find a bright object on Mars! – Mathematica

leave a comment »

Comments: “image processing – Help find a bright object on Mars! – Mathematica”


Here’s another, slightly more scientific method. One that works for many kinds of anomalies (darker, brighter, different hue, different saturation).

First, I use a part of the image that only contains sand as my training set (I use the high-res image from the NASA site instead of the one linked in the question. The results are similar, but I get much saner probabilities without the JPEG artifacts):

img = Import[""];
sandSample = ImageTake[img, {0, 200}, {1000, 1200}]

We can visualize the distribution of the R/G channels in this sample:

SmoothHistogram3D[sandPixels[[All, {1, 2}]], Automatic, "PDF", AxesLabel -> {"R", "G", "PDF"}]

The histogram looks a bit skewed, but it’s close enough to treat it as gaussian. So I’ll assume for simplicity that the “sand” texture is a gaussian random variable where each pixel is independent. Then I can estimate it’s distribution like this:

sandPixels = Flatten[ImageData[sandSample], 1];
dist = MultinormalDistribution[{mR, mG, mB}, {{sRR, sRG, sRB}, {sRG, sGG, SGB}, {sRB, sGB, sBB}}];
edist = EstimatedDistribution[sandPixels, dist];
logPdf = PowerExpand@Log@PDF[edist, {r, g, b}]

Now I can just apply the PDF of this distribution to the complete image (I use the Log PDF to prevent overflows/underflows):

rgb = ImageData /@ ColorSeparate[GaussianFilter[img, 3]];
p = logPdf /. {r -> rgb[[1]], g -> rgb[[2]], b -> rgb[[3]]};

We can visualize the negative log PDF with an appropriate scaling factor:


Here we can see:

  • The sand areas are dark – these pixels fit the estimated distribution from the sand sample
  • Most of the Curiosity area in the image is very bright – it’s very unlikely that these pixels are from the same distribution
  • The shadows of the Curiosity probe are gray – they’re not from the same distribution as the sand sample, but still closer than the anomaly
  • The anomaly we’re looking is very bright – It can be detected easily

To find the sand/non-sand areas, I use MorphologicalBinarize. For the sand pixels, the PDF is > 0 everywhere, for the anomaly pixels, it’s < 0, so finding a threshold isn’t very hard.

bin = MorphologicalBinarize[Image[-p], {0, 10}]

Here, areas where the Log[PDF] < -10 are selected. PDF < e^-10 is very unlikely, so you won’t have to check too many false positives.

Final step: find connected components, ignoring components above 10000 Pixels (that’s the rover) and mark them in the image:

components = 
 ComponentMeasurements[bin, {"Area", "Centroid", "CaliperLength"}, 
 10 < #1 < 10000 &][[All, 2]]
 Graphics[{Red, AbsoluteThickness[5], Circle[#[[2]], 2 #[[3]]] & /@ components}]]

Obviously, the assumption that “sand pixels” are independent gaussian random variables is a gross oversimplification, but the general method would work for other distributions as well. Also, r/g/b values alone are probably not the best features to find alien objects. Normally you’d use more features (e.g. a set of Gabor filters)

from Hacker News 200:


Written by cwyalpha

十月 15, 2012 在 5:07 上午

发表在 Uncategorized


Fill in your details below or click an icon to log in: 徽标

You are commenting using your account. Log Out /  更改 )

Google+ photo

You are commenting using your Google+ account. Log Out /  更改 )

Twitter picture

You are commenting using your Twitter account. Log Out /  更改 )

Facebook photo

You are commenting using your Facebook account. Log Out /  更改 )


Connecting to %s

%d 博主赞过: