| Image segmentation is a process of dividing an image into functionally distinct regions. In this thesis we develop a scheme for supervisory verification and modification of the computer generated segmentation. Our supervised image segmentation scheme combines the reliability of a human expert with the precision of the automatic image segmentation algorithms, thereby yielding precise and reliable segmentation.; To enable active and natural interaction between a supervisor and the relevant image processing algorithms, we facilitate two common modes of natural interactive gestures, namely pointing and circling. To account for imprecision and extents of manual suggestions, we use the human visual attention distribution model obtained from the published psychological and psychophysical research. The supervisor's interactions by pointing activate our region growing algorithm, and interactions by circling activate our constrained contouring algorithm.; The important contributions of this research are: (1) Development of a threshold propagation algorithm for image segmentation that facilitates supervisory specification of width of an elongated pattern in an image. (2) Development of a graph search based constrained contouring algorithm, that refines a rough input contour to yield the boundaries of the intended object. (3) The overall supervised image segmentation scheme is significantly more sophisticated than the published interactive segmentation schemes, because the interactions are approximate natural gestures, and are used in processing stage of the algorithms. The algorithms account for both the image data constraints and the approximations in the human gestures, by incorporating the human visual attention map. |
