This problem is so interesting that I try a "solution" as well.

Let "img" be the original image. Then the trick can be done by:

imgLens = 
 Thinning@Binarize@
   DeleteBorderComponents@
    SelectComponents[Blur@EdgeDetect[img, 4], 
     "EnclosingComponentCount", # == 0 &]

which gives:

I am not really satisfied with this method, because it depends quite sensitive on the parameter of EdgeDetect.

From this one can easily extract the coordinates of the boundary of the lens:

imgDims = ImageDimensions[imgLens];
pts = {#2, imgDims[[1]] - #1 + 1} & @@@ 
   Position[ImageData@imgLens, 1, {2}];

Graphics[{Blue, PointSize[.004], Point[pts]}, Axes -> True, 
 Frame -> True]

Ciao Henrik

Hi,

this is far from being efficient, it uses too many image processing commands and could hardly be less elegant. But it does find the edge - more or less.

ImageMultiply[img, 
 ColorNegate@
  EdgeDetect[
   WatershedComponents@
     Binarize@
      ImageMultiply[img, 
       Sharpen[Erosion[
          Dilation[
           MeanFilter[
            SelectComponents[
              SelectComponents[
               SelectComponents[(MorphologicalComponents[
                  BottomHatTransform[
                   GradientFilter[
                    DeleteSmallComponents@(MorphologicalComponents[
                    img, 0.562] // Colorize), 1.64], DiskMatrix[6]], 
                  Method -> "Connected"]), Large], "Count", -3], 
              "Count", 1] // Colorize, 6], 6], 7] // ImageAdjust, 
        10]] // Colorize]]

I called the image just img. Here's the result:

Cheers,

Marco

Dear Henrik,

that looks like a really great solution, much better than mine. I tried it out and got a different result when I use your command:

This was because I imported the file by "pulling it from the post". If I get it into Mathematica using Import I get what you report. That is curious because in both cases the image has the same resolution, i.e. ImageDimensions gives the same, namely {1311, 976}. Also the different result does not seem to depend on the parameter of EdgeDetect.

Anyway, your solution is much nicer than mine.

Cheers,

Marco