Identification and boundary extraction of blobs in complex imagery.

Automated identification and boundary extraction of blobs in "real world" imagery is a difficult task because the boundaries are so irregular that there is often insufficient a priori information describing these boundaries and traditional methods fail. This paper has proposed a progressive segmentation approach and a boundary estimation method to identify the blobs and to yield an accurate description of its boundary. The multiresolution image processing technique is incorporated into the whole work. This work has been applied to the problem of identifying and extracting the boundaries of major vessels (e.g., the aorta) in Magnetic Resonance (MR) imagery and the results are satisfactory. A Laplacian of Gaussian (LOG) operator is utilized as a spot detector to locate the approximate position of the blob of interest. A subimage centered on this approximate position is extracted to eliminate unwanted portions of the image and facilitate further processing. A histogram pyramid is created for the subimage histogram for automated determination of the threshold in the noisy histogram. A shrink-expand operation is then employed to reduced noise and undesired structures in the subimage. The rough and irregular boundary of the blob of interest obtained by thresholding is reparameterized into polar coordinates to create a Fourier descriptor representation of the boundary. Then, the discrete Fourier transform is applied to the reparameterized 1-D discrete curve to permit appropriate smoothing, as required, in frequency space. Finally, the boundary estimation is completed by taking the inverse Fourier transform to reconstruct the boundary of interest.