computer vision - prior based image segmentation -

i have problem @ hand, in image composed of strange objects not have closed contours. (more rivers , channels on plain ground).

i provided set of prior images of same size different rivers general orientation , structure matches river under study while position in image might deviate.

i looking image segmentation method, (theory or practice, looking clues start with) can use set of prior examples in segmenting river. in case there multiple rivers of same general orientation present in image.

i interested in ways of statistically representing these complex structures. example, if not river image (binary image), , knew had gaussian structure, use information in covariance estimated examples. in binary or trinary images, can not.

here outline image segmentation

sample small region (possible rectangle) inside river, assumption belong foreground , provide estimate color distribution. should have algorithm can find small region inside river high confidence, algorithm can trained on data have.

since know little background, ideal chose pixels lying on image frame background pixels.

the idea use these pre-selected foreground , background pixels seeds in graph cut algorithm segmentation. selecting seeds important part of graph cut algorithm segmentation, once have seeds, segmentation more or less correct. there plenty of literature/code available online on how segmentation using graph cuts.