Ever wonder how those old photos of you would look if you hadn't been wearing glasses? Experimental software developed at Microsoft Corp. can help. The program will automatically remove eyeglasses from photos of faces. Starting with an image of a person, the program locates and isolates the face. Then it uses a statistical learning method to find the eyeglasses, remove them, and fill the blanks in with the appropriate pixel colors.
Determining these fill colors is where the statistical learning really shines. The system uses models based on prior knowledge of which colors and textures belong underneath glasses to work out how to replace the missing pixels. The group's work is the first to perform this type of editing on entire objects, such as glasses, instead of on individual pixels. What's more, the statistical approach can be used to automatically edit out a wide variety of objects--someday, perhaps even unfortunate facial piercings.
Automatic Eyeglasses Removal from Face Images , by C. Wu, C. Liu, H.-Y. Shum, Y.-Q. Xu, and Z. Zhang, IEEE Transactions on Pattern Analysis and Machine Intelligence, March 2004, pp. 32236.The Bandwidth of Sign Language
While deaf people can use text-based telecommunications to speak with one another, the rate of communication is significantly slower than with face-to-face sign languages. Video telephony would seem the solution, but is the available bandwidth from residential networks enough to make it intelligible? To figure out the limits of intelligibility, a professor at the New Jersey Institute of Technology, in Newark, has calculated the fundamental bandwidth of American Sign Language (ASL), a system of gestures used by at least 500 000 people in North America. He found that ASL movements are low-frequency affairs; a mere 6 frames per second of video will do the trick so long as those frames are smoothly interpolated back to the 30-hertz video rate.
Biomechanical and Perceptual Constraints on the Bandwidth Requirements of Sign Language , by Richard A. Foulds, IEEE Transactions on Neural Systems and Rehabilitation Engineering, March 2004, pp. 6572.Scratchy Film
Movie studios are in the midst of digitizing their archived films so that they can be repackaged for digital distribution. Part of the process involves finding and removing the defects that inevitably come with age, particularly those distracting long, vertical scratch lines. So far, automatic detection of these lines has proved elusive, requiring either lots of manual tuning or many false alarms. But workers at the Istituto per le Applicazioni del Calcolo, in Rome, now have an improved scheme for automatically detecting these lines, which are usually a mere 3 to 10 pixels wide.
The key insight is that scratches should be modeled in a more complex manner than if they were purely noise added to the image. In addition to being fast enough to run in real time, the new method keeps the number of false positives low, so there is little fear that it will accidentally remove, for example, the hanging rope in your favorite cowboy movie.
A Generalized Model for Scratch Detection , by V. Bruni and D. Vitulano, IEEE Transactions on Image Processing, January 2004, pp. 4450.