I want to carry out a binary classification/segmentation on a grayscale image with very little texture. The only prior knowledge that's supposed to be available is that the object of interest is rather dark (i.e. low gray values) and comprises a significant portion of the image (maybe around 30%-70%). The object of interest should basically be a connected component (although this needn't be the case every single time), but can contain "blobs" of brighter clutter objects. Also, there might be patches of dark pixels in the clutter areas as well, but they should not be as homogeneous and as large as the object of interest.
Of course, I have already checked on simple thresholding (e.g. Otsu's method) or 2-means classification followed by some kind of in-painting, which both lead to similar results. However, these results are only satisfactory if the object of interest is very homogeneous and clearly separable from the clutter object's brighter gray value range. For rather heterogeneous sample images, there is often also clutter, which is put into the object class.
Therefore my question: Is there any interesting literature available about single-channel image classification without exploitation of texture features, but maybe incorporating spatial context? I have already done some research on google, but I have a hard time coming up with proper search terms...`
ADDITION: Active Contours is very close tho the solution I am looking for. However, they generally work best for one distinct, convex, object, which is enclosed by background. In my situation, however, many convex objects are located in front of the background, so either one active contour for every foreground object is needed or a special case of active contours considering the background as object, being able to deal with clutter objects WITHIN the shape of the object, is needed. Any such method or literature available?