Middle Level Vision Systems for Low Vision

Low vision is a significant reduction of visual function that cannot be fully corrected by ordinary lenses, medical treatment, or surgery. Aging, injuries, and diseases can cause low vision, with the most common disease being age-related macular degeneration (AMD). Our overarching goal is to develop devices to aid people with low vision. We use techniques of computer vision and computational neuroscience to build systems that enhance natural images. These techniques exploit the relatively homogeneous properties of surfaces of natural objects and of natural regions (e.g., the sky). Visual systems first make local measurements of these properties. Then, these systems combine the measurements to build idealized region and surface models. The initial measurements and the region and surface models are called low- and mid-level vision respectively. We hypothesize that because these models represent images in a simplified form, they are easier to see, aiding people with low vision. Alternatively, one may superimpose some of these models, e.g., contours, on original images, hypothesizing that this helps people parse them. We develop systems implementing such models and test these hypotheses on normal people, visually impaired aging people, and AMD patients.

To achieve the goal of building mid-level-vision systems to aid people with low vision, we developed an interdisciplinary team. Computational engineers from USC develop the systems’ algorithms, mimicking strategies found in biological processes. These engineers work in consultation with neurobiologists and cognitive scientists from USC and Harvard University. In addition, a team of psychologists and clinicians from USC, Harvard, and the University of Houston develop and administer probes to test the effectiveness of the systems in humans. This team has expertise in blindness and low vision, and in the development of visual aids. The project is led by the USC team. Semi-monthly 2-hr meetings among the USC scientists helps coordinate the technical developments. Semi-monthly conference calls and bi-annual trips from the Harvard and University of Houston partners coordinate the human-testing efforts.

Intranet

Public Data Sharing