The KUIM Image Analysis System
Project Award Date: 0000-00-00
The KUIM image analysis system is a collection of general- and special-purpose image processing and computer vision applications developed at the University of Kansas. At the heart of the system is an image I/O library which supports a variety of image formats. Over fifty application programs provide basic image analysis operations, such as contrast enhancement, noise removal, image restoration, and image compression. The KUIM image analysis system is currently available on most UNIX platforms (DEC, HP, SGI, Sun) and is now being ported to Windows NT.
Our current research focuses on fast and effective tools for image segmentation and motion analysis. Specific segmentation applications include identifying cells in microscopic images, classifying ice types in remote sensing images, segmentation of multispectral images and medical images, and temporal segmentation of video sequences. Other research areas include multisensor image fusion, stereo correspondence, and parallel algorithms for image processing. Some of our 1996 projects follow:
o Algorithms for Video Segmentation (J. Bride, S. Bouix, S. Daugherty). This project's goal is to partition video sequences into short clips for use in a digital video library. To do this, we must identify the starting and ending points of video segments that were spliced together. Simple measures like mean squared pixel difference are effective for identifying sharp transitions, but more advanced computer vision methods are needed to distinguish camera fadings from camera or object motions. To support this computationally demanding approach, we are developing parallel algorithms using PVM on a network of workstations.
o Color Image Segmentation Using Gradient Watersheds (X. Hong). The intensity gradient helps identify edges in an image (high gradients) and homogeneous regions within objects (low gradients). One method for segmenting an image into visually sensible regions is to view the intensity gradient image as elevation, and calculate the watershed regions (i.e., collections of pixels that are surrounded by pixels with high gradients). Extending this notion to color images requires an appropriate metric for calculating differences in pixel color. Our goal is to evaluate different color spaces and difference metrics for this purpose.
o Framework for Parallel Image Processing. A number of image processing applications have high computational requirements and lend themselves well to parallel implementation. For example, non-linear contrast enhancement methods that process local pixel neighborhoods to generate output images can be partitioned into subimages and run in parallel on multiple machines. In a heterogeneous network of workstations, the challenge is how to "best" divide the input image into subimages to balance the work while minimizing communication among processors. The project goal was to develop an object-oriented framework on top of PVM to assist in the development of load balanced parallel implementations of image processing applications.
Primary Sponsor(s): NSF Infrastructure Grant