Illumination Invariants and Visual Perception
Problem being addressed
Visual perception by computers is a difficult problem in part due to changes in illumination; light sources and albedo (whiteness). What methods can be used to overcome these difficulties?
In order to assist computer vision, deep learning method that first uses light invariants such as the position of the light source, the object albedo, and surface normals is proposed. The method combines deep belief networks with lighting model assumptions from optics known as Lambertian reflectance. By explcitly taking the lighting in account, the proposed method outpeforms many of the other standard methods for one shot facial recognition under different lighting conditions.
Advantages of this solution
Visual perception is an important task for computer vision and for robotics. The ability to replicate how humans "see" will be important for any number of automated systems in the future. This research proposes a solution to make the task of seeing more robust under different lighting conditions.
Possible New Application of the Work
Electronics and Sensors Industry
Sensors need to adjust to different lighting conditions at different times. This algorithm could be embedded directly into sensors and cameras in order to adjust and compensate for changes in the lighting.
This algorithm could be used in robotics to improve robot vision and navigation systems. This would include driver less car navigation where changes in the light source can interfere with optical systems.
This research relies on a knowledge of invariant properties to enhance deep learning system. Perhaps invariant parameters from networks could be used to enhance deep learning systems that related to the internet.
Source URL: #############