Image segmentation based on the integration of pixel affinity and deformable models

This paper describes a general-purpose method we have developed for automatically segmenting objects of an unknown number and unknown locations in images. Our method integrates deformable models and statistics of image cues including intensity, gradient, color and texture. By using a combination of image features rather than a single feature such as gradient our method is more robust to noise and sparse data. To allow for the automated segmentation of an unknown number and locations of objects, we simultaneously segment objects initialized at uniformly distributed points in the image. A method is developed to automatically merge models corresponding to the same object. Results of the method are presented for several examples, including greyscale, color and noisy images.