Computer Science Colloquium - Bruce Maxwell

Removing Illumination as a Confounding Signal Improves Machine Learning

An image is the reflection of light from surfaces. Most computer vision tasks are based on trying to identify the properties or identity of the surfaces in an image. For example, what are the objects in an image? While illumination is necessary in order to see, variation in an image can occur because of changes in either illumination or reflectance (surface color). In order to correctly understand the surface properties, we have to either discount or remove the illumination signal. In this talk IÕll present two separate methods of identifying and removing illumination as a confounding signal for two important tasks: optical character recognition and vision for autonomous vehicles. These are tasks where deep networks are the top performing systems. For both tasks IÕll demonstrate that removing the illumination signal first enables a machine learning system to achieve better performance than using standard images. Using a physics-based pre-processor in tandem with machine learning boosts performance by providing a better signal for computer vision.

Dr. Maxwell started exploring computer science as an undergraduate at Swarthmore College, where he earned a B.A. in Political Science, a B.S. in Engineering, and a Concentration in Computer Science. He went on to obtain an M.Phil. in Speech Recognition at Cambridge University and a Ph.D. in Robotics at Carnegie Mellon University. He taught for two years at the University of North Dakota and nine years at Swarthmore College before coming to Colby in the fall of 2007 as chair of the CS department. His interests include robotics, computer vision, computer graphics, scientific data analysis and visualization.