Boundary identification

My high-level strategy would be to use WatershedSegmentation on a gradient image (see this answer for a description of what this function does), starting from a marker image that looks like this:

So WatershedSegmentation can grow the highlighted "seeds" until they touch at the highest gradient border.

So first step: Get the marker image. It doesn't have to follow the grain contours perfectly, so I can use a large blurring filter to smooth out all the high-frequency details:

GaussianFilter[img, 100]

The grains are easily recognizable in this image using any segmentation technique you like, I'll just use Binarize:

binary = Binarize[gaussian];

To get the markers as above, I take the perimeter of that binary image and make it "wider" using Erosion:

markers = 
  Erosion[ColorNegate[MorphologicalPerimeter[binary]], 
   DiskMatrix[50]];

That's it, now we can pass the whole thing to WatershedSegmentation:

(seg = WatershedComponents[
    ColorConvert[ImageAdd[GradientFilter[img, 5]], "Grayscale"], 
    markers]) // Colorize

(You might want to play with the GradientFilter filter size to improve the result)

You can get the perimeter using

perimeter = Binarize[ColorNegate[Image[seg]]];
HighlightImage[img, perimeter]

To visualize the result, I'll show it filled:

HighlightImage[img, FillingTransform@perimeter]

One approach is to smooth the image (here I've used the curvature flow filter, which is a nonlinear edge-preserving smoothing filter) and then binarize before doing the edge detection.

img = Import["http://i.stack.imgur.com/TIiTZ.jpg"]  
curve = CurvatureFlowFilter[img, 40];
EdgeDetect[FillingTransform[DeleteSmallComponents[Binarize[curve]]]]

Dilating makes it look a bit smoother, but you may (or may not) want to do that, depending on what you are doing with the images.

Dilation[EdgeDetect[FillingTransform[DeleteSmallComponents[Binarize[curve]]]], 1]